Merge branch 'master' into master
diff --git a/docs-site/README.md b/docs-site/README.md
index 49cd920..0a116e3 100644
--- a/docs-site/README.md
+++ b/docs-site/README.md
@@ -35,6 +35,12 @@
The generated html files are under `website/build/singa-doc/`.
+To host the website locally for testing, under `website/build/singa-doc/` type:
+
+```sh
+$ python -m http.server
+```
+
## Update and Add Pages
### Navigation Bar
diff --git a/docs-site/docs/autograd.md b/docs-site/docs/autograd.md
index cc6bd1e..292a861 100644
--- a/docs-site/docs/autograd.md
+++ b/docs-site/docs/autograd.md
@@ -192,8 +192,7 @@
### Using the Model API
-The following <<<<<<< HEAD
-[example](https://github.com/apache/singa/blob/master/examples/cnn/model/cnn.py)
+The following [example](https://github.com/apache/singa/blob/master/examples/cnn/model/cnn.py)
implements a CNN model using the [Model API](./graph).
#### Define the subclass of Model
diff --git a/docs-site/docs/download.md b/docs-site/docs/download.md
index 667a09a..d58524a 100644
--- a/docs-site/docs/download.md
+++ b/docs-site/docs/download.md
@@ -1,5 +1,5 @@
---
-id: download-singa
+id: downloads
title: Download SINGA
---
diff --git a/docs-site/docs/half-precision.md b/docs-site/docs/half-precision.md
new file mode 100644
index 0000000..126ae70
--- /dev/null
+++ b/docs-site/docs/half-precision.md
@@ -0,0 +1,104 @@
+---
+id: half-precision
+title: Half Precision
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+Half precision training could bring benefits:
+- using less GPU memory, supporting larger network.
+- training faster.
+
+## Half data type
+
+### Half data type definition
+The IEEE 754 standard specifies a binary16 as having the following
+ [format](https://en.wikipedia.org/wiki/Half-precision_floating-point_format):
+Sign bit: 1 bit
+Exponent width: 5 bits
+Significand precision: 11 bits (10 explicitly stored)
+
+### Half data type operation
+Load data in fp32 and easily convert to fp16 by casting.
+```python
+>>> from singa import tensor, device
+>>> dev = device.create_cuda_gpu()
+>>> x = tensor.random((2,3),dev)
+>>> x
+[[0.7703407 0.42764223 0.5872884 ]
+ [0.78362167 0.70469785 0.64975065]], float32
+>>> y = x.as_type(tensor.float16)
+>>> y
+[[0.7705 0.4277 0.5874]
+ [0.7837 0.7046 0.65 ]], float16
+```
+
+Primary operations are supported in fp16.
+```python
+>>> y+y
+[[1.541 0.8555 1.175 ]
+ [1.567 1.409 1.3 ]], float16
+```
+
+## Training in Half
+
+### Training in Half three step
+Training in half precision could be done easily in three steps:
+1. Load data and convert to half
+2. Set data type of optimizer
+3. Train model as usual
+``` python
+# cast input data to fp16
+x = load_data()
+x = x.astype(np.float16)
+tx = tensor.from_numpy(x)
+
+# load model
+model = build_model()
+# set optimizer dtype to fp16
+sgd = opt.SGD(lr=0.1, dtype=tensor.float16)
+
+# train as usual
+out, loss = model(tx, ty)
+```
+
+### Example
+An example script is `train_cnn.py`, run below command to train in half.
+```python
+python examples/cnn/train_cnn.py cnn mnist -pfloat16
+```
+
+## Implementation
+
+### Half Type Dependency
+This half implementation is integrated in C++ backend as general half type
+support.
+
+To run on GPU, `__half` is available in Cuda math API. To support `__half`
+math operation, it is required to compile against Nvidia compute arch > 6.0
+ (Pascal).
+
+### Nvidia Hardware Acceleration: Tensor Core
+Tensor Core released by Nvidia further accelerates half precision and multiples
+throughput for operations like GEMM(CuBlas) and convolution(CuDNN). To enable
+Tensor core operation, there are a few restriction on GEMM dimensions,
+convolution channel size, Cuda version, and GPU version(Turing or later) and etc.
+
+### Implement Operations
+Half operations are primarily implemented in `tensor_math_cuda.h`, by specializing
+operation template with half type and implementation the low level computation.
+
+For example, GEMM operation is implemented as:
+```c++
+template <>
+void GEMM<half_float::half, lang::Cuda>(const half_float::half alpha,
+ const Tensor& A, const Tensor& B,
+ const half_float::half beta, Tensor* C,
+ Context* ctx) {
+ // ...
+ CUBLAS_CHECK(cublasGemmEx(handle, transb, transa, ncolB, nrowA, ncolA,
+ alphaPtr, BPtr, Btype, ldb, APtr, Atype, lda,
+ betaPtr, CPtr, Ctype, ldc, computeType, algo));
+ // ...
+}
+```
\ No newline at end of file
diff --git a/docs-site/docs/releases/RELEASE_NOTES_3.1.0.md b/docs-site/docs/releases/RELEASE_NOTES_3.1.0.md
index b74f7bf..f8e556f 100644
--- a/docs-site/docs/releases/RELEASE_NOTES_3.1.0.md
+++ b/docs-site/docs/releases/RELEASE_NOTES_3.1.0.md
@@ -25,7 +25,7 @@
- Add SGD operators to opt module, including RMSProp, Adam, and AdaGrad.
- Extend the sonnx module to support DenseNet121, ShuffleNetv1, ShuffleNetv2,
- SqueezeNet, VGG19, GPT2, and RoBERTa,
+ SqueezeNet, VGG19, GPT2, and RoBERTa.
- Reconstruct sonnx to
diff --git a/docs-site/docs/wheel-cpu-dev.md b/docs-site/docs/wheel-cpu-dev.md
index 9f8f7af..c1b0101 100644
--- a/docs-site/docs/wheel-cpu-dev.md
+++ b/docs-site/docs/wheel-cpu-dev.md
@@ -1,6 +1,6 @@
---
id: wheel-cpu-dev
-title: CPU only Wheel Packages (develop version)
+title: CPU only (dev version)
---
<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
diff --git a/docs-site/docs/wheel-cpu.md b/docs-site/docs/wheel-cpu.md
index 430d9d1..7952a31 100644
--- a/docs-site/docs/wheel-cpu.md
+++ b/docs-site/docs/wheel-cpu.md
@@ -1,6 +1,6 @@
---
id: wheel-cpu
-title: CPU only Wheel Packages
+title: CPU only
---
<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
diff --git a/docs-site/docs/wheel-gpu-dev.md b/docs-site/docs/wheel-gpu-dev.md
index 3c061e9..d2cad4b 100644
--- a/docs-site/docs/wheel-gpu-dev.md
+++ b/docs-site/docs/wheel-gpu-dev.md
@@ -1,6 +1,6 @@
---
id: wheel-gpu-dev
-title: Wheel Packages with CUDA enabled (develop version)
+title: CUDA enabled (dev version)
---
<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
diff --git a/docs-site/docs/wheel-gpu.md b/docs-site/docs/wheel-gpu.md
index 2c0a1a9..bd91c4f 100644
--- a/docs-site/docs/wheel-gpu.md
+++ b/docs-site/docs/wheel-gpu.md
@@ -1,6 +1,6 @@
---
id: wheel-gpu
-title: Wheel Packages with CUDA Enabled
+title: CUDA enabled
---
<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
diff --git a/docs-site/docs_chs/assets/AllReduce.png b/docs-site/docs_chs/assets/AllReduce.png
new file mode 100644
index 0000000..ba9cd43
--- /dev/null
+++ b/docs-site/docs_chs/assets/AllReduce.png
Binary files differ
diff --git a/docs-site/docs_chs/assets/GraphOfMLP.png b/docs-site/docs_chs/assets/GraphOfMLP.png
new file mode 100644
index 0000000..19bfbad
--- /dev/null
+++ b/docs-site/docs_chs/assets/GraphOfMLP.png
Binary files differ
diff --git a/docs-site/docs_chs/assets/GraphPipeline.png b/docs-site/docs_chs/assets/GraphPipeline.png
new file mode 100644
index 0000000..c2c5ada
--- /dev/null
+++ b/docs-site/docs_chs/assets/GraphPipeline.png
Binary files differ
diff --git a/docs-site/docs_chs/assets/MPI.png b/docs-site/docs_chs/assets/MPI.png
new file mode 100644
index 0000000..536fb03
--- /dev/null
+++ b/docs-site/docs_chs/assets/MPI.png
Binary files differ
diff --git a/docs-site/docs_chs/assets/benchmark.png b/docs-site/docs_chs/assets/benchmark.png
new file mode 100644
index 0000000..711e6b5
--- /dev/null
+++ b/docs-site/docs_chs/assets/benchmark.png
Binary files differ
diff --git a/docs-site/docs_chs/assets/singav1-sw.png b/docs-site/docs_chs/assets/singav1-sw.png
new file mode 100644
index 0000000..e443c6e
--- /dev/null
+++ b/docs-site/docs_chs/assets/singav1-sw.png
Binary files differ
diff --git a/docs-site/docs_chs/assets/singav3-sw.png b/docs-site/docs_chs/assets/singav3-sw.png
new file mode 100644
index 0000000..cf6b044
--- /dev/null
+++ b/docs-site/docs_chs/assets/singav3-sw.png
Binary files differ
diff --git a/docs-site/docs_chs/assets/singav3.1-sw.png b/docs-site/docs_chs/assets/singav3.1-sw.png
new file mode 100644
index 0000000..2bfd2c0
--- /dev/null
+++ b/docs-site/docs_chs/assets/singav3.1-sw.png
Binary files differ
diff --git a/docs-site/docs_chs/autograd.md b/docs-site/docs_chs/autograd.md
new file mode 100644
index 0000000..7dbc525
--- /dev/null
+++ b/docs-site/docs_chs/autograd.md
@@ -0,0 +1,240 @@
+---
+id: autograd
+title: Autograd
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+实现autograd有两种典型的方式，一种是通过如[Theano](http://deeplearning.net/software/theano/index.html)的符号微分（symbolic differentiation）或通过如[Pytorch](https://pytorch.org/docs/stable/notes/autograd.html)的反向微分（reverse differentialtion）。SINGA遵循Pytorch方式，即通过记录计算图，并在正向传播后自动应用反向传播。自动传播算法的详细解释请参阅[这里](https://pytorch.org/docs/stable/notes/autograd.html)。我们接下来对SINGA中的相关模块进行解释，并举例说明其使用方法。
+
+## 相关模块
+
+在autograd中涉及三个类，分别是`singa.tensor.Tensor`，`singa.autograd.Operation`和`singa.autograd.Layer`。在本篇的后续部分中，我们使用Tensor、Operation和Layer来指代这三个类。
+
+### Tensor
+
+Tensor的三个属性被autograd使用：
+
+- `.creator`是一个`Operation`实例。它记录了产生Tensor实例的这个操作。
+- `.request_grad`是一个布尔变量。它用于指示autograd算法是否需要计算张量的梯度。例如，在反向传播的过程中，线性层的权重矩阵和卷积层（非底层）的特征图的张量梯度应该被计算。
+- `.store_grad`是一个布尔变量。它用于指示张量的梯度是否应该被存储并由后向函数输出。例如，特征图的梯度是在反向传播过程中计算出来的，但不包括在反向函数的输出中。
+
+开发者可以改变Tensor实例的`requires_grad`和`stores_grad`。例如，如果将后者设置为True，那么相应的梯度就会被包含在后向函数的输出。需要注意的是，如果`stores_grad`是True，那么 `requires_grad`一定是真，反之亦然。
+
+
+### Operation
+
+它将一个或多个`Tensor`实例作为输入，然后输出一个或多个`Tensor`实例。例如，ReLU可以作为一个具体的Operation子类来实现。当一个`Operation`实例被调用时（实例化后），会执行以下两个步骤。
+
+1.记录源操作，即输入张量的`创建者`。
+2.通过调用成员函数`.forward()`进行计算。
+
+有两个成员函数用于前向和反向传播，即`.forward()`和`.backward()`。它们以`Tensor.data`作为输入（类型为`CTensor`），并输出`Ctensor`s。要添加一个特定的操作，子类`Operation`应该实现自己的`.forward()`和`.backward()`函数。在后向传播过程中，autograd的`backward()`函数会自动调用`backward()`函数来计算输入的梯度（根据`require_grad`字段的参数和约束）。
+
+### Layer
+
+对于那些需要参数的Operation，我们把它们封装成一个新的类，`Layer`。例如，卷积操作被封装到卷积层(Convolution layer)中。`层`管理（存储）参数，并调用相应的`Operation`来实现变换。
+
+## 样例
+
+在[example folder](https://github.com/apache/singa/tree/master/examples/autograd)中提供了很多样例。在这里我我们分析两个最具代表性的例子。
+
+### 只使用Operation
+
+下一段代码展示了一个只使用`Operation`的多层感知机（MLP）模型：
+
+#### 调用依赖包
+
+```python
+from singa.tensor import Tensor
+from singa import autograd
+from singa import opt
+```
+
+#### 创建权重矩阵和偏置向量
+
+在将`requires_grad`和`stores_grad`都设置为`True`的情况下，创建参数张量。
+
+```python
+w0 = Tensor(shape=(2, 3), requires_grad=True, stores_grad=True)
+w0.gaussian(0.0, 0.1)
+b0 = Tensor(shape=(1, 3), requires_grad=True, stores_grad=True)
+b0.set_value(0.0)
+
+w1 = Tensor(shape=(3, 2), requires_grad=True, stores_grad=True)
+w1.gaussian(0.0, 0.1)
+b1 = Tensor(shape=(1, 2), requires_grad=True, stores_grad=True)
+b1.set_value(0.0)
+```
+
+#### 训练
+
+```python
+inputs = Tensor(data=data) # data matrix
+target = Tensor(data=label) # label vector
+autograd.training = True # for training
+sgd = opt.SGD(0.05) # optimizer
+
+for i in range(10):
+ x = autograd.matmul(inputs, w0) # matrix multiplication
+ x = autograd.add_bias(x, b0) # add the bias vector
+ x = autograd.relu(x) # ReLU activation operation
+
+ x = autograd.matmul(x, w1)
+ x = autograd.add_bias(x, b1)
+
+ loss = autograd.softmax_cross_entropy(x, target)
+
+ for p, g in autograd.backward(loss):
+ sgd.update(p, g)
+```
+
+### 使用Operation和Layer
+
+下面的[例子](https://github.com/apache/singa/blob/master/examples/autograd/mnist_cnn.py)使用autograd模块提供的层实现了一个CNN模型。
+
+#### 创建层
+
+```python
+conv1 = autograd.Conv2d(1, 32, 3, padding=1, bias=False)
+bn1 = autograd.BatchNorm2d(32)
+pooling1 = autograd.MaxPool2d(3, 1, padding=1)
+conv21 = autograd.Conv2d(32, 16, 3, padding=1)
+conv22 = autograd.Conv2d(32, 16, 3, padding=1)
+bn2 = autograd.BatchNorm2d(32)
+linear = autograd.Linear(32 * 28 * 28, 10)
+pooling2 = autograd.AvgPool2d(3, 1, padding=1)
+```
+
+#### 定义正向传播函数
+
+在正向传播中的operations会被自动记录，用于反向传播。
+
+```python
+def forward(x, t):
+ # x is the input data (a batch of images)
+ # t is the label vector (a batch of integers)
+ y = conv1(x) # Conv layer
+ y = autograd.relu(y) # ReLU operation
+ y = bn1(y) # BN layer
+ y = pooling1(y) # Pooling Layer
+
+ # two parallel convolution layers
+ y1 = conv21(y)
+ y2 = conv22(y)
+ y = autograd.cat((y1, y2), 1) # cat operation
+ y = autograd.relu(y) # ReLU operation
+ y = bn2(y)
+ y = pooling2(y)
+
+ y = autograd.flatten(y) # flatten operation
+ y = linear(y) # Linear layer
+ loss = autograd.softmax_cross_entropy(y, t) # operation
+ return loss, y
+```
+
+#### 训练
+
+```python
+autograd.training = True
+for epoch in range(epochs):
+ for i in range(batch_number):
+ inputs = tensor.Tensor(device=dev, data=x_train[
+ i * batch_sz:(1 + i) * batch_sz], stores_grad=False)
+ targets = tensor.Tensor(device=dev, data=y_train[
+ i * batch_sz:(1 + i) * batch_sz], requires_grad=False, stores_grad=False)
+
+ loss, y = forward(inputs, targets) # forward the net
+
+ for p, gp in autograd.backward(loss): # auto backward
+ sgd.update(p, gp)
+```
+
+### Using the Model API
+
+下面的[样例](https://github.com/apache/singa/blob/master/examples/cnn/model/cnn.py)使用[Model API](./graph)实现了一个CNN模型。.
+
+#### 定义Model的子类
+
+定义模型类，它应该是Model的子类。只有这样，在训练阶段使用的所有操作才会形成一个计算图以便进行分析。图中的操作将被按时序规划并有效执行，模型类中也可以包含层。
+
+```python
+class MLP(model.Model): # the model is a subclass of Model
+
+ def __init__(self, data_size=10, perceptron_size=100, num_classes=10):
+ super(MLP, self).__init__()
+
+ # init the operators, layers and other objects
+ self.relu = layer.ReLU()
+ self.linear1 = layer.Linear(perceptron_size)
+ self.linear2 = layer.Linear(num_classes)
+ self.softmax_cross_entropy = layer.SoftMaxCrossEntropy()
+
+ def forward(self, inputs): # define the forward function
+ y = self.linear1(inputs)
+ y = self.relu(y)
+ y = self.linear2(y)
+ return y
+
+ def train_one_batch(self, x, y):
+ out = self.forward(x)
+ loss = self.softmax_cross_entropy(out, y)
+ self.optimizer(loss)
+ return out, loss
+
+ def set_optimizer(self, optimizer): # attach an optimizer
+ self.optimizer = optimizer
+```
+
+#### 训练
+
+```python
+# create a model instance
+model = MLP()
+# initialize optimizer and attach it to the model
+sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
+model.set_optimizer(sgd)
+# input and target placeholders for the model
+tx = tensor.Tensor((batch_size, 1, IMG_SIZE, IMG_SIZE), dev, tensor.float32)
+ty = tensor.Tensor((batch_size, num_classes), dev, tensor.int32)
+# compile the model before training
+model.compile([tx], is_train=True, use_graph=True, sequential=False)
+
+# train the model iteratively
+for b in range(num_train_batch):
+ # generate the next mini-batch
+ x, y = ...
+
+ # Copy the data into input tensors
+ tx.copy_from_numpy(x)
+ ty.copy_from_numpy(y)
+
+ # Training with one batch
+ out, loss = model(tx, ty)
+```
+
+#### 保存模型checkpoint
+
+```python
+# define the path to save the checkpoint
+checkpointpath="checkpoint.zip"
+
+# save a checkpoint
+model.save_states(fpath=checkpointpath)
+```
+
+#### 加载模型checkpoint
+
+```python
+# define the path to load the checkpoint
+checkpointpath="checkpoint.zip"
+
+# load a checkpoint
+import os
+if os.path.exists(checkpointpath):
+ model.load_states(fpath=checkpointpath)
+```
+
+### Python API
+
+关于Python API的更多细节，请参考[这里](https://singa.readthedocs.io/en/latest/autograd.html#module-singa.autograd)。
diff --git a/docs-site/docs_chs/benchmark-train.md b/docs-site/docs_chs/benchmark-train.md
new file mode 100644
index 0000000..1a72da4
--- /dev/null
+++ b/docs-site/docs_chs/benchmark-train.md
@@ -0,0 +1,16 @@
+---
+id: benchmark-train
+title: Benchmark for Distributed Training
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+
+项目：我们使用深度卷积神经网络[ResNet-50](https://github.com/apache/singa/blob/master/examples/cnn/model/resnet.py)。它有50个卷积层，用于图像分类。它需要3.8个GFLOPs来通过网络处理一张图像（尺寸为224x224）。输入的图像大小为224x224。
+
+
+硬件方面：我们使用的是AWS的p2.8xlarge实例，每个实例有8个Nvidia Tesla K80 GPU，共96GB GPU内存，32个vCPU，488GB主内存，10Gbps网络带宽。
+
+衡量标准：我们衡量不同数量worker的每次迭代时间，以评估SINGA的可扩展性。Batch-size固定为每个GPU32个。采用同步训练方案。因此，有效的batch-size是`32N`，其中N是GPU的数量。我们与一个流行的开源系统进行比较，该系统采用参数服务器拓扑结构。选择第一个GPU作为服务器。
+
+ <br/> **可扩展性测试。条形为吞吐量，折线形为通信成本。**
diff --git a/docs-site/docs_chs/build.md b/docs-site/docs_chs/build.md
new file mode 100644
index 0000000..e1f7f30
--- /dev/null
+++ b/docs-site/docs_chs/build.md
@@ -0,0 +1,430 @@
+---
+id: build
+title: Build SINGA from Source
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+源文件可以通过[tar.gz文件](https://dist.apache.org/repos/dist/dev/singa/)或git repo的形式下载：
+
+```shell
+$ git clone https://github.com/apache/singa.git
+$ cd singa/
+```
+
+如果您想为SINGA贡献代码，请参考[贡献代码](contribute-code.md)页面的步骤和要求。
+
+## 使用Conda构筑SINGA
+
+Conda-build 是一个构建工具，它从anaconda cloud安装依赖的库并执行构建脚本。
+
+安装conda-build(需要先安装conda)：
+
+```shell
+conda install conda-build
+```
+
+### 构建CPU版本
+
+构建SINGA的CPU版本：
+
+```shell
+conda build tool/conda/singa/
+```
+
+以上命令已经在Ubuntu（14.04，16.04和18.04）和macOS 10.11上测试过。更多信息请参考[Travis-CI page](https://travis-ci.org/apache/singa)页面。
+
+### 构建GPU版本
+
+要构建GPU版的SINGA，计算机必须装备有Nvida GPU，而且需要安装CUDA driver(>=384.81)、CUDA toolkit(>=9)和cuDNN(>=7)。以下两个Docker镜像提供了构建环境：
+
+1. apache/singa:conda-cuda9.0
+2. apache/singa:conda-cuda10.0
+
+构建环境准备好后，需要先导出CUDA版本，然后运行conda命令构建SINGA：
+
+```shell
+export CUDA=x.y (e.g. 9.0)
+conda build tool/conda/singa/
+```
+
+### 后处理
+
+生成的包文件的位置(`.tar.gz`)将打印在终端上，生成的包可以直接安装：
+
+```shell
+conda install -c conda-forge --use-local <path to the package file>
+```
+
+若要上传到anaconda云端供他人下载安装，需要在 anaconda 上注册一个账号，才能[上传包](https://docs.anaconda.com/anaconda-cloud/user-guide/getting-started/)：
+
+```shell
+conda install anaconda-client
+anaconda login
+anaconda upload -l main <path to the package file>
+```
+
+将包上传到云端后，您可以在[Anaconda Cloud](https://anaconda.org/)上看到，也可以通过以下命令查看：
+
+```shell
+conda search -c <anaconda username> singa
+```
+
+每个特定的SINGA软件包都由版本和构建字符串来标识。要安装一个特定的SINGA包，需要提供所有信息，例如：
+
+```shell
+conda install -c <anaconda username> -c conda-forge singa=2.1.0.dev=cpu_py36
+```
+
+为了使安装命令简单化，您可以创建以下附加包，这些包依赖于最新的CPU和GPU SINGA包：
+
+```console
+# for singa-cpu
+conda build tool/conda/cpu/ --python=3.6
+conda build tool/conda/cpu/ --python=3.7
+# for singa-gpu
+conda build tool/conda/gpu/ --python=3.6
+conda build tool/conda/gpu/ --python=3.7
+```
+
+因此，当您运行：
+
+```shell
+conda install -c <anaconda username> -c conda-forge singa-xpu
+```
+
+时(`xpu`表示'cpu' or 'gpu'), 相应的真正的SINGA包将作为依赖库被安装。
+
+## 使用本地工具在Ubuntu上构建SINGA
+
+请参阅 SINGA [Dockerfiles](https://github.com/apache/singa/blob/master/tool/docker/devel/ubuntu/cuda9/Dockerfile#L30)，了解在 Ubuntu 16.04 上安装依赖库的说明。您也可以使用 devel 映像创建一个 Docker 容器，并在容器中构建 SINGA。要使用GPU、DNNL、Python和单元测试来构建SINGA，请运行以下命令：
+
+```shell
+mkdir build # at the root of singa folder
+cd build
+cmake -DENABLE_TEST=ON -DUSE_CUDA=ON -DUSE_DNNL=ON -DUSE_PYTHON3=ON ..
+make
+cd python
+pip install .
+```
+CMake选项的详细内容在本页最后一节解释，上面最后一条命令是安装Python包。你也可以运行`pip install -e .`，它可以创建符号链接，而不是将 Python 文件复制到 site-package 文件夹中。
+
+如果SINGA在ENABLE_TEST=ON的情况下编译，您可以通过以下方式运行单元测试:
+
+```shell
+$ ./bin/test_singa
+```
+
+您可以看到所有的测试案例与测试结果。如果SINGA通过了所有测试，那么您就成功安装了SINGA。
+
+## 使用本地工具在Centos7上构建SINGA
+
+由于Centos7的软件包名称不同，因此从源码开始构建会有所不同。
+
+### 安装依赖项
+
+基础包和库文件：
+
+```shell
+sudo yum install freetype-devel libXft-devel ncurses-devel openblas-devel blas-devel lapack devel atlas-devel kernel-headers unzip wget pkgconfig zip zlib-devel libcurl-devel cmake curl unzip dh-autoreconf git python-devel glog-devel protobuf-devel
+```
+
+构建必需的包：
+
+```shell
+sudo yum group install "Development Tools"
+```
+
+若要安装swig：
+
+```shell
+sudo yum install pcre-devel
+wget http://prdownloads.sourceforge.net/swig/swig-3.0.10.tar.gz
+tar xvzf swig-3.0.10.tar.gz
+cd swig-3.0.10.tar.gz
+./configure --prefix=${RUN}
+make
+make install
+```
+
+安装gfortran：
+
+```shell
+sudo yum install centos-release-scl-rh
+sudo yum --enablerepo=centos-sclo-rh-testing install devtoolset-7-gcc-gfortran
+```
+
+安装pip和其他包：
+
+```shell
+sudo yum install epel-release
+sudo yum install python-pip
+pip install matplotlib numpy pandas scikit-learn pydot
+```
+
+### 安装SINGA
+
+按照《使用本地工具在Ubuntu上构建SINGA》的步骤1-5进行操作
+
+### 测试
+
+您可以通过如下方式进行
+
+```shell
+$ ./bin/test_singa
+```
+
+您可以看到所有的测试案例与测试结果。如果SINGA通过了所有测试，即表示安装成功。
+
+## 在Windows中编译SINGA
+在Windows上使用Python支持构建SINGA的说明可以在[install-win页面](install-win.md)找到。
+
+## 关于编译选项的更多细节
+
+### USE_MODULES (已过期废弃)
+
+如果没有安装protobuf和openblas，你可以用它们一起编译SINGA
+
+```shell
+$ In SINGA ROOT folder
+$ mkdir build
+$ cd build
+$ cmake -DUSE_MODULES=ON ..
+$ make
+```
+cmake 会下载 OpenBlas 和 Protobuf (2.6.1) 并与 SINGA 一起编译。
+
+您可以使用`ccmake ..`来配置编译选项。如果一些依赖的库不在系统默认路径中，则您需要手动导出以下环境变量：
+
+```shell
+export CMAKE_INCLUDE_PATH=<path to the header file folder>
+export CMAKE_LIBRARY_PATH=<path to the lib file folder>
+```
+
+### USE_PYTHON
+
+编译SINGA的Python封装器选项：
+
+```shell
+$ cmake -DUSE_PYTHON=ON ..
+$ make
+$ cd python
+$ pip install .
+```
+
+### USE_CUDA
+我们推荐用户安装CUDA和[cuDNN](https://developer.nvidia.com/cudnn)，以便在GPU上运行SINGA，以获得更好的性能。
+
+SINGA已经在CUDA 9/10和cuDNN 7上进行了测试。如果cuDNN安装在非系统文件夹中，例如 /home/bob/local/cudnn/，则需要执行以下命令来让cmake在编译时能够找到它们：
+
+```shell
+$ export CMAKE_INCLUDE_PATH=/home/bob/local/cudnn/include:$CMAKE_INCLUDE_PATH
+$ export CMAKE_LIBRARY_PATH=/home/bob/local/cudnn/lib64:$CMAKE_LIBRARY_PATH
+$ export LD_LIBRARY_PATH=/home/bob/local/cudnn/lib64:$LD_LIBRARY_PATH
+```
+
+CUDA和cuDNN的cmake选项应该设置成“ON”：
+
+```shell
+# Dependent libs are install already
+$ cmake -DUSE_CUDA=ON ..
+$ make
+```
+
+### USE_DNNL
+
+用户可以启用DNNL来提高CPU的计算性能，DNNL的安装指南可以在[这里](https://github.com/intel/mkl-dnn#installation)找到：
+
+
+SINGA在DNNL v1.1环境下已经进行过测试并通过，
+
+若要启用DNNL支持来编译SINGA:
+
+```shell
+# Dependent libs are installed already
+$ cmake -DUSE_DNNL=ON ..
+$ make
+```
+
+### USE_OPENCL
+
+SINGA使用opencl-headers和viennacl（版本1.7.1及以上）来支持OpenCL，它们可以通过如下方式安装：
+
+```shell
+# On Ubuntu 16.04
+$ sudo apt-get install opencl-headers, libviennacl-dev
+# On Fedora
+$ sudo yum install opencl-headers, viennacl
+```
+
+此外，你需要在你想运行OpenCL的平台安装OpenCL Installable Client Driver（ICD）。
+
+- 对于AMD和Nvidia的GPU，驱动包也应该安装与之匹配的OpenCL ICD。
+- 对于Intel的CPU和/或GPU，请从[Intel官方网站](https://software.intel.com/en-us/articles/opencl-drivers)上获取驱动程序。请注意，该网站上提供的驱动程序只支持最新的CPU和Iris GPU。
+- 对于旧的Intel CPU，你可以使用beignet-opencl-icd包。
+
+请注意，目前不建议在CPU上运行OpenCL，因为运行速度会很慢。内存传输是以整数秒为单位的（直观来说，CPU上是1000毫秒，而GPU上是1毫秒）。
+
+更多关于建立OpenCL工作环境的信息可以在[这里](https://wiki.tiker.net/OpenCLHowTo)找到。
+
+如果ViennaCL的软件包版本不是至少1.7.1，则需要从源码构建它：
+
+从[这个git repo](https://github.com/viennacl/viennacl-dev)clone版本库，切换（checkout）到release-1.7.1分支，然后构建它，并把项目路径添加到PATH，再把构建的库文件添加到LD_LIBRARY_PATH。
+
+构建支持OpenCL的SINGA（在SINGA 1.1上测试）：
+
+```shell
+$ cmake -DUSE_OPENCL=ON ..
+$ make
+```
+
+### PACKAGE
+
+这个设置是用来构建 Debian 软件包的。设置PACKAGE=ON，然后用make命令来编译软件包，如下所示：
+
+```shell
+$ cmake -DPACKAGE=ON
+$ make package
+```
+
+## FAQ
+
+- Q: 'import singa'阶段报错
+
+ A: 请检查`python -c "from singa import _singa_wrap`中的详细错误。有时是由依赖库引起的，比如protobuf有多个版本，cudnn缺失，numpy版本不匹配等问题。下面展示了不同情况下的解决方案
+
+ 1. 检查cudnn和cuda。如果cudnn缺失或与wheel包的版本不一致，你可以下载正确的cudnn版本到~/local/cudnn/，然后：
+
+ ```shell
+ $ echo "export LD_LIBRARY_PATH=/home/<yourname>/local/cudnn/lib64:$LD_LIBRARY_PATH" >> ~/.bashrc
+ ```
+
+ 2. 如果是protobuf的问题。你可以将protobuf (3.6.1)从源码安装到本地文件夹，比如 ~/local/，解压tar文件，然后：
+
+ ```shell
+ $ ./configure --prefix=/home/<yourname>local
+ $ make && make install
+ $ echo "export LD_LIBRARY_PATH=/home/<yourname>/local/lib:$LD_LIBRARY_PATH" >> ~/.bashrc
+ $ source ~/.bashrc
+ ```
+
+ 3. 如果找不到包括python在内的其他类库，则使用`pip`或`conda`创建虚拟环境.
+
+ 4. 如果不是上述原因造成的，请到`_singa_wrap.so`文件夹中查看：
+ ```shell
+ $ python
+ >> import importlib
+ >> importlib.import_module('_singa_wrap')
+ ```
+
+ 来检查错误信息。例如，如果是numpy的版本不匹配，错误信息会是：
+
+ ```shell
+ RuntimeError: module compiled against API version 0xb but this version of numpy is 0xa
+ ```
+
+ 那么你就需要更新numpy到更高版本。
+
+* Q: 运行`cmake ...`时出错，找不到依赖库。
+
+ A: 如果你还没有安装这些依赖库，请先安装它们。如果你在系统文件夹之外的文件夹中安装了库，例如/usr/local，那您需要手动导出以下变量:
+
+ ```shell
+ $ export CMAKE_INCLUDE_PATH=<path to your header file folder>
+ $ export CMAKE_LIBRARY_PATH=<path to your lib file folder>
+ ```
+
+- Q: 来自`make`的错误，例如linking阶段的错误.
+
+ A: 如果您的库文件在系统默认路径以外的其他文件夹中，则需要手动导出以下变量。
+
+ ```shell
+ $ export LIBRARY_PATH=<path to your lib file folder>
+ $ export LD_LIBRARY_PATH=<path to your lib file folder>
+ ```
+
+* Q: 来自头文件的错误，例如'cblas.h文件不存在'
+
+ A: 您需要手动将cblas.h的文件夹包含在CPLUS_INCLUDE_PATH中，例如：
+
+ ```shell
+ $ export CPLUS_INCLUDE_PATH=/opt/OpenBLAS/include:$CPLUS_INCLUDE_PATH
+ ```
+
+* Q: 在编译SINGA时，我收到错误信息`SSE2 instruction set not enabled`
+
+ A:您可以尝试如下指令:
+
+ ```shell
+ $ make CFLAGS='-msse2' CXXFLAGS='-msse2'
+ ```
+
+* Q:当试图导入.py文件时，从google.protobuf.internal收到`ImportError: cannot import name enum_type_wrapper`。
+
+ A: 您需要安装protobuf的Python绑定包，它可以通过如下方式安装：
+
+ ```shell
+ $ sudo apt-get install protobuf
+ ```
+
+ 或从源文件编译：
+
+ ```shell
+ $ cd /PROTOBUF/SOURCE/FOLDER
+ $ cd python
+ $ python setup.py build
+ $ python setup.py install
+ ```
+
+* Q: 当我从源码构建OpenBLAS时，被告知需要一个Fortran编译器。
+
+ A: 您可以通过如下方式编译OpenBLAS：
+
+ ```shell
+ $ make ONLY_CBLAS=1
+ ```
+ 或者通过如下方式安装：
+
+ ```shell
+ $ sudo apt-get install libopenblas-dev
+ ```
+
+* Q: 当我构建协议缓冲区时，它报告说在`/usr/lib64/libstdc++.so.6`中没有找到`GLIBC++_3.4.20`？
+
+ A: 这意味着链接器找到了libstdc++.so.6，但该库属于旧的GCC（用于编译和链接程序的GCC）版本。此程序依赖于新的libstdc++中定义的代码，而该代码属于较新版本的GCC，所以必须告诉链接器如何找到较新的libstdc++共享库。最简单的解决方法是找到正确的libstdc++，并将其导出到LD_LIBRARY_PATH。例如，如果下面命令的输出中列出了GLIBC++_3.4.20：
+
+ $ strings /usr/local/lib64/libstdc++.so.6|grep GLIBC++
+
+ 那么接下来需要设置环境变量为：
+
+ $ export LD_LIBRARY_PATH=/usr/local/lib64:$LD_LIBRARY_PATH
+
+* Q: 当构建glog时，报告说 "src/logging_unittest.cc:83:20: error: 'gflags' is not a namespace-name"。
+
+ A: 可能是由于安装的gflags用了不同的命名空间，比如 "google"，所以glog找不到'gflags'的命名空间。实际上建立glog并不需要gflags，所以你可以修改configure.ac文件来忽略gflags。
+
+ 1. cd to glog src directory
+ 2. change line 125 of configure.ac to "AC_CHECK_LIB(gflags, main, ac_cv_have_libgflags=0, ac_cv_have_libgflags=0)"
+ 3. autoreconf
+
+ 执行上述命令后，就可以重新构建glog了。
+
+* Q: 在使用虚拟环境时，每次运行pip install，都会重新安装numpy。但是，当运行`import numpy`时，numpy并没有被调用。
+
+ A: 可能是由于`PYTHONPATH`造成的，在使用虚拟环境时，应将`PYTHONPATH`设置为空，以避免与虚拟环境本身的路径冲突。
+
+* Q: 当从源代码编译PySINGA时，由于缺少<numpy/objectarray.h>，出现的编译错误。
+
+ A: 请安装numpy并导出numpy头文件的路径为
+
+ $ export CPLUS_INCLUDE_PATH=`python -c "import numpy; print numpy.get_include()"`:$CPLUS_INCLUDE_PATH
+
+* Q: 当我在Mac OS X中运行SINGA时，报错 "Fatal Python error:
+ PyThreadState_Get: no current thread Abort trap: 6"
+
+ A: 这个错误通常发生在系统上有多个版本的Python，并且是通过pip安装的SINGA (通过conda安装时不会出现这个问题)，例如，操作系统自带的版本和Homebrew安装的版本。PySINGA所链接的Python必须与Python解释器（interpreter）相同。 您可以通过 `which python` 检查您的解释器路径并通过`otool -L <path to _singa_wrap.so>`检查PySINGA链接的Python路径。要解决这个问题，请用正确的 Python 版本编译 SINGA。需要注意的是，如果您从源码编译 PySINGA，您需要在调用[cmake](http://stackoverflow.com/questions/15291500/i-have-2-versions-of-python-installed-but-cmake-is-using-older-version-how-do)时指定路径：
+
+
+ $ cmake -DPYTHON_LIBRARY=`python-config --prefix`/lib/libpython2.7.dylib -DPYTHON_INCLUDE_DIR=`python-config --prefix`/include/python2.7/ ..
+
+ 如果从二进制包中安装PySINGA，例如debian或wheel，那么你需要改变python解释器的路径，例如，重新设置$PATH，并把Python的正确路径放在前面的位置。
diff --git a/docs-site/docs_chs/contribute-code.md b/docs-site/docs_chs/contribute-code.md
new file mode 100644
index 0000000..dee1df8
--- /dev/null
+++ b/docs-site/docs_chs/contribute-code.md
@@ -0,0 +1,108 @@
+---
+id: contribute-code
+title: How to Contribute Code
+---
+
+<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 代码风格
+
+SINGA代码库在[CPP](http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml)和[Python](http://google.github.io/styleguide/pyguide.html)代码中都遵循Google风格。
+
+强制执行Google编码风格的一个简单方法是使用Visual Studio Code编辑器中的linting和格式化工具:
+
+- [C/C++扩展](https://marketplace.visualstudio.com/items?itemName=ms-vscode.cpptools)
+- [Python扩展](https://marketplace.visualstudio.com/items?itemName=ms-python.python)
+- [cpplint扩展](https://marketplace.visualstudio.com/items?itemName=mine.cpplint)
+- [Clang-Format](https://marketplace.visualstudio.com/items?itemName=xaver.clang-format)
+
+安装扩展后，编辑`settings.json`文件：
+
+```json
+{
+ "[cpp]": {
+ "editor.defaultFormatter": "xaver.clang-format"
+ },
+ "cpplint.cpplintPath": "path/to/cpplint",
+
+ "editor.formatOnSave": true,
+ "python.formatting.provider": "yapf",
+ "python.linting.enabled": true,
+ "python.linting.lintOnSave": true,
+ "clang-format.language.cpp.style": "google",
+ "python.formatting.yapfArgs": ["--style", "{based_on_style: google}"]
+}
+```
+
+根据您的操作系统，用户设置文件位于以下位置：
+
+1. Windows %APPDATA%\Code\User\settings.json
+2. macOS "\$HOME/Library/Application Support/Code/User/settings.json"
+3. Linux "\$HOME/.config/Code/User/settings.json"
+
+配置是在相应的配置文件中指定的。而这些工具会自动查找项目根目录下的配置文件，比如`.pylintrc`。
+
+#### 安装必要工具
+
+最理想的情况是所有贡献者都使用相同版本的代码格式化工具（clang-format 9.0.0和yapf 0.29.0），这样在不同PR中的代码格式化就会完全相同，从而摆脱github pull request冲突。
+
+首先，安装LLVM 9.0，它提供了clang-format 9.0.0版本，LLVM的下载页面如下:
+
+- [LLVM](http://releases.llvm.org/download.html#9.0.0)
+
+ - Ubuntu系统：
+
+ ```sh
+ sudo apt-get install clang-format-9
+ ```
+
+ - Windows系统，下载预编译包并安装。
+
+然后，安装cpplint, pylint和yapf
+
+- Ubuntu或OSX:
+
+ ```
+ $ sudo pip install cpplint
+ $ which cpplint
+ /path/to/cpplint
+
+ $ pip install yapf==0.29.0
+ $ pip install pylint
+ ```
+
+- Windows: 安装Anaconda进行包管理
+
+ ```
+ $ pip install cpplint
+ $ where cpplint
+ C:/path/to/cpplint.exe
+
+ $ pip install yapf==0.29.0
+ $ pip install pylint
+ ```
+
+#### 使用
+
+- 配置后，在编辑源代码文件时，linting会自动启用。错误和警告会在Visual Studio Code `PROBLEMS`面板中列出。
+- 代码格式化可以通过调出Command Palette(Windows中为`Shift+Ctrl+P`，OS X中为`Shift+Command+P`)并输入`Format Document`来完成。
+
+#### 提交
+
+修正格式错误以后就可以提交pull request了。
+
+## 开发环境
+
+推荐使用Visual Studio Code作为编辑器。可以安装Python、C/C++、代码拼写检查器、autoDocstring、vim、远程开发等扩展。这些扩展的参考配置（即`settings.json`）可以在[这里](https://gist.github.com/nudles/3d23cfb6ffb30ca7636c45fe60278c55)查看。
+
+如果更新CPP代码，需要从[源文件](./build.md)重新编译SINGA。建议使用`*-devel Docker`镜像中的原生构建工具或使用`conda build`。
+
+如果要只更新Python代码，您可以安装一次SINGA，然后复制更新后的Python文件来替换Python安装文件夹中的文件。
+
+```shell
+cp python/singa/xx.py <path to conda>/lib/python3.7/site-packages/singa/
+```
+
+## 工作流程
+
+请参阅[git工作流程页面](./git-workflow.md).
diff --git a/docs-site/docs_chs/contribute-docs.md b/docs-site/docs_chs/contribute-docs.md
new file mode 100644
index 0000000..6c9ef19
--- /dev/null
+++ b/docs-site/docs_chs/contribute-docs.md
@@ -0,0 +1,82 @@
+---
+id: contribute-docs
+title: How to Contribute to Documentation
+---
+
+<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+文档有两种类型，即markdown文件和API使用参考。本指南介绍了一些工具，并指导如何准备md文件和API注释。
+
+md文件将通过[Docusaurus](https://docusaurus.io/)构建成HTML页面；API注释（来自源代码）将用于使用Sphinx（对应Python）和Doxygen（对应CPP）生成API参考页面。
+
+## Markdown文件
+
+请尽量遵循[Google Documentation style](https://developers.google.com/style)。例如：
+
+1. 删除指令中的"please"。如：'Please click...' VS 'Click...'。
+2. 遵循标准的[大小写规则](https://owl.purdue.edu/owl/general_writing/mechanics/help_with_capitals.html)。
+3. 在说明中用"you"代替"we"。
+4. 使用现在时态，避免使用`will`。
+5. 尽量使用主动语态而不是被动。
+
+此外，为了使文件一致：
+
+1. 句子不宜过长, e.g., 长度<=80
+2. 尽量使用相对路径，假设我们在repo的根目录下，例如，`doc-site/docs`指的是`singa-doc/docs-site/docs`。
+3. 使用背标将命令、路径、类函数和变量亮出来，例如，`Tensor`, `singa-doc/docs-site/docs`。
+4. 为了突出其他术语/概念，使用 _斜体_ or **加粗**
+
+
+本项目使用的[prettier tool](https://prettier.io/)会在我们进行git提交时，根据[配置](https://github.com/apache/singa-doc/blob/master/docs-site/.prettierrc)自动格式化代码。例如，它会将markdown文件中的文本最多折叠成80个字符（注释行除外）。
+
+在介绍一个概念（如`Tensor`类）时，要提供概述（目的和与其他概念的关系）、API和例子，还可以用Google colab来演示其用法。
+
+详细的编辑md文件和建立网站的方法请参考[本页面](https://github.com/apache/singa-doc/tree/master/docs-site)。
+
+## API References
+
+### CPP API
+
+请遵循[Google CPP注释风格](https://google.github.io/styleguide/cppguide.html#Comments).
+
+要生成文档，请从doc文件夹中运行 "doxygen"（推荐使用Doxygen >= 1.8）。
+
+### Python API
+
+请遵循[Google Python DocString风格](http://google.github.io/styleguide/pyguide.html#38-comments-and-docstrings).
+
+## Visual Studio Code (vscode)
+
+如果你使用vscode作为编辑器，我们推荐使用以下插件。
+
+### Docstring Snippet
+
+[autoDocstring](https://marketplace.visualstudio.com/items?itemName=njpwerner.autodocstring)生成函数、类等的docstring，要注意选择使用`google`的docString格式。
+
+### Spell Check
+
+[Code Spell Checker](https://marketplace.visualstudio.com/items?itemName=streetsidesoftware.code-spell-checker)可以用来检查代码的注释，或`.md`和`.rst`文件。
+
+要只对Python代码的注释进行拼写检查，可以在`File - Preferences - User Snippets - python.json`中添加以下代码段：
+
+ "cspell check" : {
+ "prefix": "cspell",
+ "body": [
+ "# Directives for doing spell check only for python and c/cpp comments",
+ "# cSpell:includeRegExp #.* ",
+ "# cSpell:includeRegExp (\"\"\"|''')[^\1]*\1",
+ "# cSpell: CStyleComment",
+ ],
+ "description": "# spell check only for python comments"
+ }
+
+如果要只对Cpp代码的注释进行拼写检查，可以在`File - Preferences - User Snippets - cpp.json`中添加以下代码段：
+
+ "cspell check" : {
+ "prefix": "cspell",
+ "body": [
+ "// Directive for doing spell check only for cpp comments",
+ "// cSpell:includeRegExp CStyleComment",
+ ],
+ "description": "# spell check only for cpp comments"
+ }
diff --git a/docs-site/docs_chs/device.md b/docs-site/docs_chs/device.md
new file mode 100644
index 0000000..944d5a0
--- /dev/null
+++ b/docs-site/docs_chs/device.md
@@ -0,0 +1,29 @@
+---
+id: device
+title: Device
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+Device代表一个有内存和计算单元的硬件设备。所有的[Tensor操作](./tensor)都由常驻设备安排执行的，Tensor内存也由设备的内存管理器管理。因此，需要在Device类中实现内存和执行的优化。
+
+
+## 特定设备
+
+目前，SINGA支持三种设备：
+
+1. CudaGPU：用于运行Cuda代码的Nvidia GPU。
+2. CppCPU：用于运行Cpp代码的CPU。
+3. OpenclGPU：用于运行OpenCL代码的GPU卡。
+
+## 用法示例
+
+下面的代码提供了创建设备的例子：
+
+```python
+from singa import device
+cuda = device.create_cuda_gpu_on(0) # use GPU card of ID 0
+host = device.get_default_device() # get the default host device (a CppCPU)
+ary1 = device.create_cuda_gpus(2) # create 2 devices, starting from ID 0
+ary2 = device.create_cuda_gpus([0,2]) # create 2 devices on ID 0 and 2
+```
diff --git a/docs-site/docs_chs/dist-train.md b/docs-site/docs_chs/dist-train.md
new file mode 100644
index 0000000..50aa6ad
--- /dev/null
+++ b/docs-site/docs_chs/dist-train.md
@@ -0,0 +1,378 @@
+---
+id: dist-train
+title: Distributed Training
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA支持跨多个GPU的数据并行训练（在单个节点上或跨不同节点）。下图说明了数据并行训练的情况：
+
+
+
+在分布式训练中，每个进程(称为worker)在单个GPU上运行一个训练脚本，每个进程都有一个单独的通信等级，训练数据被分发给各个worker，模型在每个worker上被复制。在每次迭代中，worker从其分区中读取数据的一个mini-batch（例如，256张图像），并运行BackPropagation算法来计算权重的梯度，通过all-reduce（由[NCCL](https://developer.nvidia.com/nccl)提供）进行平均，按照随机梯度下降算法（SGD）进行权重更新。
+
+NCCL的all-reduce操作可以用来减少和同步不同GPU的梯度。假设我们使用4个GPU进行训练，如下图所示。一旦计算出4个GPU的梯度，all-reduce将返回GPU的梯度之和，并使其在每个GPU上可用，然后就可以轻松计算出平均梯度。
+
+
+
+## 使用
+
+SINGA提供了一个名为`DistOpt`（`Opt`的一个子类）的模块，用于分布式训练。它封装了一个普通的SGD优化器，并调用`Communicator`进行梯度同步。下面的例子说明了如何使用`DistOpt`在MNIST数据集上训练一个CNN模型。源代码在[这里](https://github.com/apache/singa/blob/master/examples/cnn/)，或者可以使用[Colab notebook]()。
+
+### 代码示例
+
+1. 定义神经网络模型：
+
+```python
+class CNN(model.Model):
+
+ def __init__(self, num_classes=10, num_channels=1):
+ super(CNN, self).__init__()
+ self.conv1 = layer.Conv2d(num_channels, 20, 5, padding=0, activation="RELU")
+ self.conv2 = layer.Conv2d(20, 50, 5, padding=0, activation="RELU")
+ self.linear1 = layer.Linear(500)
+ self.linear2 = layer.Linear(num_classes)
+ self.pooling1 = layer.MaxPool2d(2, 2, padding=0)
+ self.pooling2 = layer.MaxPool2d(2, 2, padding=0)
+ self.relu = layer.ReLU()
+ self.flatten = layer.Flatten()
+ self.softmax_cross_entropy = layer.SoftMaxCrossEntropy()
+
+ def forward(self, x):
+ y = self.conv1(x)
+ y = self.pooling1(y)
+ y = self.conv2(y)
+ y = self.pooling2(y)
+ y = self.flatten(y)
+ y = self.linear1(y)
+ y = self.relu(y)
+ y = self.linear2(y)
+ return y
+
+ def train_one_batch(self, x, y, dist_option='fp32', spars=0):
+ out = self.forward(x)
+ loss = self.softmax_cross_entropy(out, y)
+
+ # Allow different options for distributed training
+ # See the section "Optimizations for Distributed Training"
+ if dist_option == 'fp32':
+ self.optimizer(loss)
+ elif dist_option == 'fp16':
+ self.optimizer.backward_and_update_half(loss)
+ elif dist_option == 'partialUpdate':
+ self.optimizer.backward_and_partial_update(loss)
+ elif dist_option == 'sparseTopK':
+ self.optimizer.backward_and_sparse_update(loss,
+ topK=True,
+ spars=spars)
+ elif dist_option == 'sparseThreshold':
+ self.optimizer.backward_and_sparse_update(loss,
+ topK=False,
+ spars=spars)
+ return out, loss
+
+# create model
+model = CNN()
+```
+
+2. 创建`DistOpt`实例并将其应用到创建的模型上：
+
+```python
+sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
+sgd = opt.DistOpt(sgd)
+model.set_optimizer(sgd)
+dev = device.create_cuda_gpu_on(sgd.local_rank)
+```
+
+下面是关于代码中一些变量的解释：
+
+(i) `dev`
+
+dev代表`Device`实例，在设备中加载数据并运行CNN模型。
+
+(ii)`local_rank`
+
+Local rank表示当前进程在同一节点中使用的GPU数量。例如，如果你使用的节点有2个GPU，`local_rank=0`表示这个进程使用的是第一个GPU，而`local_rank=1`表示使用的是第二个GPU。使用MPI或多进程，你能够运行相同的训练脚本，唯一的区别`local_rank`的值不同。
+
+(iii)`global_rank`
+
+global中的rank代表了你使用的所有节点中所有进程的全局排名。让我们考虑这样的情况：你有3个节点，每个节点有两个GPU， `global_rank=0`表示使用第1个节点的第1个GPU的进程， `global_rank=2`表示使用第2个节点的第1个GPU的进程， `global_rank=4`表示使用第3个节点的第1个GPU的进程。
+
+3. 加载和分割训练/验证数据：
+
+```python
+def data_partition(dataset_x, dataset_y, global_rank, world_size):
+ data_per_rank = dataset_x.shape[0] // world_size
+ idx_start = global_rank * data_per_rank
+ idx_end = (global_rank + 1) * data_per_rank
+ return dataset_x[idx_start:idx_end], dataset_y[idx_start:idx_end]
+
+train_x, train_y, test_x, test_y = load_dataset()
+train_x, train_y = data_partition(train_x, train_y,
+ sgd.global_rank, sgd.world_size)
+test_x, test_y = data_partition(test_x, test_y,
+ sgd.global_rank, sgd.world_size)
+```
+
+这个`dev`的数据集的一个分区被返回。
+
+
+这里，`world_size`代表你用于分布式训练的所有节点中的进程总数。
+
+4. 初始化并同步所有worker的模型参数:
+
+```python
+#Synchronize the initial parameter
+tx = tensor.Tensor((batch_size, 1, IMG_SIZE, IMG_SIZE), dev, tensor.float32)
+ty = tensor.Tensor((batch_size, num_classes), dev, tensor.int32)
+model.compile([tx], is_train=True, use_graph=graph, sequential=True)
+...
+#Use the same random seed for different ranks
+seed = 0
+dev.SetRandSeed(seed)
+np.random.seed(seed)
+```
+
+5. 运行BackPropagation和分布式SGD
+
+```python
+for epoch in range(max_epoch):
+ for b in range(num_train_batch):
+ x = train_x[idx[b * batch_size: (b + 1) * batch_size]]
+ y = train_y[idx[b * batch_size: (b + 1) * batch_size]]
+ tx.copy_from_numpy(x)
+ ty.copy_from_numpy(y)
+ # Train the model
+ out, loss = model(tx, ty)
+```
+
+### 执行命令
+
+有两种方式可以启动训练，MPI或Python multiprocessing。
+
+#### Python multiprocessing
+
+它可以在一个节点上使用多个GPU，其中，每个GPU是一个worker。
+
+1. 将上述训练用的代码打包进一个函数：
+
+```python
+def train_mnist_cnn(nccl_id=None, local_rank=None, world_size=None):
+ ...
+```
+
+2. 创建`mnist_multiprocess.py`。
+
+```python
+if __name__ == '__main__':
+ # Generate a NCCL ID to be used for collective communication
+ nccl_id = singa.NcclIdHolder()
+
+ # Define the number of GPUs to be used in the training process
+ world_size = int(sys.argv[1])
+
+ # Define and launch the multi-processing
+ import multiprocessing
+ process = []
+ for local_rank in range(0, world_size):
+ process.append(multiprocessing.Process(target=train_mnist_cnn,
+ args=(nccl_id, local_rank, world_size)))
+
+ for p in process:
+ p.start()
+```
+
+下面是关于上面所创建的变量的一些说明：
+
+(i) `nccl_id`
+
+需要注意的是，我们在这里需要生成一个NCCL ID，用于集体通信，然后将其传递给所有进程。NCCL ID就像一个门票，只有拥有这个ID的进程才能加入到all-reduce操作中。(如果我们接下来使用MPI，NCCL ID的传递就没有必要了，因为在我们的代码中，这个ID会由MPI自动广播。)
+
+(ii) `world_size`
+
+world_size是您想用于训练的GPU数量。
+
+(iii) `local_rank`
+
+local_rank决定分布式训练的本地顺序，以及在训练过程中使用哪个gpu。在上面的代码中，我们使用for循环来运行训练函数，其中参数local_rank从0迭代到world_size。在这种情况下，不同的进程可以使用不同的GPU进行训练。
+
+创建`DistOpt`实例的参数应按照如下方式更新：
+
+```python
+sgd = opt.DistOpt(sgd, nccl_id=nccl_id, local_rank=local_rank, world_size=world_size)
+```
+
+3. 运行`mnist_multiprocess.py`：
+
+```sh
+python mnist_multiprocess.py 2
+```
+
+与单GPU训练相比，它最主要的意义是速度提升：
+
+```
+Starting Epoch 0:
+Training loss = 408.909790, training accuracy = 0.880475
+Evaluation accuracy = 0.956430
+Starting Epoch 1:
+Training loss = 102.396790, training accuracy = 0.967415
+Evaluation accuracy = 0.977564
+Starting Epoch 2:
+Training loss = 69.217010, training accuracy = 0.977915
+Evaluation accuracy = 0.981370
+Starting Epoch 3:
+Training loss = 54.248390, training accuracy = 0.982823
+Evaluation accuracy = 0.984075
+Starting Epoch 4:
+Training loss = 45.213406, training accuracy = 0.985560
+Evaluation accuracy = 0.985276
+Starting Epoch 5:
+Training loss = 38.868435, training accuracy = 0.987764
+Evaluation accuracy = 0.986278
+Starting Epoch 6:
+Training loss = 34.078186, training accuracy = 0.989149
+Evaluation accuracy = 0.987881
+Starting Epoch 7:
+Training loss = 30.138697, training accuracy = 0.990451
+Evaluation accuracy = 0.988181
+Starting Epoch 8:
+Training loss = 26.854443, training accuracy = 0.991520
+Evaluation accuracy = 0.988682
+Starting Epoch 9:
+Training loss = 24.039650, training accuracy = 0.992405
+Evaluation accuracy = 0.989083
+```
+
+#### MPI
+
+只要有多个GPU，MPI既适用于单节点，也适用于多节点。
+
+1. 创建`mnist_dist.py`。
+
+```python
+if __name__ == '__main__':
+ train_mnist_cnn()
+```
+
+2. 为MPI生成一个hostfile，例如下面的hostfile在一个节点上使用了2个进程（即2个GPU）：
+
+```txt
+localhost:2
+```
+
+3. 通过`mpiexec`启动训练：
+
+```sh
+mpiexec --hostfile host_file python mnist_dist.py
+```
+
+与单GPU训练相比，它同样可以带来速度的提升：
+```
+Starting Epoch 0:
+Training loss = 383.969543, training accuracy = 0.886402
+Evaluation accuracy = 0.954327
+Starting Epoch 1:
+Training loss = 97.531479, training accuracy = 0.969451
+Evaluation accuracy = 0.977163
+Starting Epoch 2:
+Training loss = 67.166870, training accuracy = 0.978516
+Evaluation accuracy = 0.980769
+Starting Epoch 3:
+Training loss = 53.369656, training accuracy = 0.983040
+Evaluation accuracy = 0.983974
+Starting Epoch 4:
+Training loss = 45.100403, training accuracy = 0.985777
+Evaluation accuracy = 0.986078
+Starting Epoch 5:
+Training loss = 39.330826, training accuracy = 0.987447
+Evaluation accuracy = 0.987179
+Starting Epoch 6:
+Training loss = 34.655270, training accuracy = 0.988799
+Evaluation accuracy = 0.987780
+Starting Epoch 7:
+Training loss = 30.749735, training accuracy = 0.989984
+Evaluation accuracy = 0.988281
+Starting Epoch 8:
+Training loss = 27.422146, training accuracy = 0.991319
+Evaluation accuracy = 0.988582
+Starting Epoch 9:
+Training loss = 24.548153, training accuracy = 0.992171
+Evaluation accuracy = 0.988682
+```
+
+## 针对分布式训练的优化
+
+SINGA为分布式训练提供了多种优化策略，以降低模块间的通信成本。参考`DistOpt`的API，了解每种策略的配置。
+
+
+当我们使用`model.Model`建立模型时，我们需要在`training_one_batch`方法中启用分布式训练的选项，请参考本页顶部的示例代码。我们也可以直接复制这些选项的代码，然后在其他模型中使用。
+
+有了定义的选项，我们可以在使用`model(tx, ty, dist_option, spars)`开始训练时，设置对应的参数`dist_option`和`spars`。
+
+### 不采取优化手段
+
+```python
+out, loss = model(tx, ty)
+```
+
+`loss`是损失函数的输出张量，例如分类任务中的交叉熵。
+
+### 半精度梯度（Half-precision Gradients）
+
+```python
+out, loss = model(tx, ty, dist_option = 'fp16')
+```
+
+在调用all-reduce之前，它将每个梯度值转换为16-bit表示（即半精度）。
+
+### 部分同步（Partial Synchronization）
+
+```python
+out, loss = model(tx, ty, dist_option = 'partialUpdate')
+```
+
+在每一次迭代中，每个rank都做本地SGD更新。然后，只对一个部分的参数进行平均同步，从而节省了通信成本。分块大小是在创建`DistOpt`实例时配置的。
+
+### 梯度稀疏化（Gradient Sparsification）
+
+该策略应用稀疏化方案来选择梯度的子集进行all-reduce，有两种方式：
+
+- 选择前k大的元素，spars是被选择的元素的一部分（比例在0 - 1之间）。
+
+```python
+out, loss = model(tx, ty, dist_option = 'sparseTopK', spars = spars)
+```
+
+- 所有绝对值大于预定义阈值的梯度都会被选中。
+
+```python
+out, loss = model(tx, ty, dist_option = 'sparseThreshold', spars = spars)
+```
+超参数在创建`DistOpt`实例时配置。
+
+## 实现
+
+本节主要是让开发者了解分布训练模块的代码是如何实现的。
+
+### NCCL communicator的C接口
+
+
+首先，通信层是用C语言[communicator.cc](https://github.com/apache/singa/blob/master/src/io/communicator.cc)编写的，它调用用NCCL库进行集体通信。
+
+communicator有两个构造器，一个是MPI的，另一个是Multiprocess的。
+
+(i) MPI构造器
+
+构造器首先先获取全局rank和world_size，计算出本地rank，然后由rank 0生成NCCL ID并广播给每个rank。之后，它调用setup函数来初始化NCCL communicator、cuda流和缓冲区。
+
+(ii) Python multiprocess构造器
+
+构造器首先从输入参数中获取rank、world_size和NCCL ID。之后，调用setup函数来初始化NCCL communicator、cuda流和缓冲区。
+
+在初始化之后，它提供了all-reduce功能来同步模型参数或梯度。例如，synch接收一个输入张量，通过NCCL例程进行all-reduce，在我们调用synch之后，需要调用wait函数来等待all-reduce操作的完成。
+
+### DistOpt的Python接口
+
+然后，python接口提供了一个[DistOpt](https://github.com/apache/singa/blob/master/python/singa/opt.py)类来封装一个[optimizer](https://github.com/apache/singa/blob/master/python/singa/opt.py)对象，以执行基于MPI或Multiprocess的分布式训练。在初始化过程中，它创建了一个NCCL communicator对象（来自于上面小节提到的C接口），然后，`DistOpt`中的每一次all-reduce操作都会用到这个communicator对象。
+
+在MPI或Multiprocess中，每个进程都有一个独立的rank，它给出了各个进程使用的GPU的信息。训练数据是被拆分的，因此每个进程可以根据一部分训练数据来评估子梯度。一旦每个进程的子梯度被计算出来，就可以将所有进程计算出的子梯度做all-reduce，得到总体随机梯度。
\ No newline at end of file
diff --git a/docs-site/docs_chs/download.md b/docs-site/docs_chs/download.md
new file mode 100644
index 0000000..4ab2825
--- /dev/null
+++ b/docs-site/docs_chs/download.md
@@ -0,0 +1,158 @@
+---
+id: download-singa
+title: Download SINGA
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## Verify
+
+要验证下载的tar.gz文件，下载[KEYS](https://www.apache.org/dist/singa/KEYS)和ASC文件，然后执行以下命令:
+
+```shell
+% gpg --import KEYS
+% gpg --verify downloaded_file.asc downloaded_file
+```
+
+你也可以检查SHA512或MD5值判断下载是否完成。
+
+## V3.0.0 (18 April 2020):
+
+- [Apache SINGA 3.0.0](http://www.apache.org/dyn/closer.cgi/singa/3.0.0/apache-singa-3.0.0.tar.gz)
+ [\[SHA512\]](https://www.apache.org/dist/singa/3.0.0/apache-singa-3.0.0.tar.gz.sha512)
+ [\[ASC\]](https://www.apache.org/dist/singa/3.0.0/apache-singa-3.0.0.tar.gz.asc)
+- [Release Notes 3.0.0](releases/RELEASE_NOTES_3.0.0)
+- 新特性及重要更新：
+ - 增强了ONNX。在SINGA中测试了多种ONNX模型。
+ - 使用 MPI 和 NCCL Communication进行分布式训练，通过梯度稀疏化和压缩以及分块传输进行了优化。
+ - 计算图的构建，利用图优化了速度和内存。
+ - 新的文档网站（singa.apache.org）和API参考网站（apache-singa.rtfd.io）。
+ - 使用CI实现代码质量检查。
+ - 将MKLDNN替换为DNNL。
+ - 更新Tensor API以支持广播操作。
+ - 实现了支持ONNX模型的新的autograd操作符。
+
+## 孵化版本（incubating） v2.0.0 (20 April 2019):
+
+- [Apache SINGA 2.0.0 (incubating)](http://www.apache.org/dyn/closer.cgi/incubator/singa/2.0.0/apache-singa-incubating-2.0.0.tar.gz)
+ [\[SHA512\]](https://www.apache.org/dist/incubator/singa/2.0.0/apache-singa-incubating-2.0.0.tar.gz.sha512)
+ [\[ASC\]](https://www.apache.org/dist/incubator/singa/2.0.0/apache-singa-incubating-2.0.0.tar.gz.asc)
+- [Release Notes 2.0.0 (incubating)](releases/RELEASE_NOTES_2.0.0.html)
+- 新特性及重要更新：
+ - 增强了autograd功能(适用于卷积网络和循环网络)。
+ - 支持ONNX。
+ - 通过英特尔MKL DNN lib改进CPP操作。
+ - 实现tensor广播。
+ - 在Apache用户名下移动Docker镜像。
+ - 更新conda-build配置中依赖的lib版本。
+
+## 孵化版本（incubating） v1.2.0 (6 June 2018):
+
+- [Apache SINGA 1.2.0 (incubating)](https://archive.apache.org/dist/incubator/singa/1.2.0/apache-singa-incubating-1.2.0.tar.gz)
+ [\[SHA512\]](https://archive.apache.org/dist/incubator/singa/1.2.0/apache-singa-incubating-1.2.0.tar.gz.sha512)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/1.2.0/apache-singa-incubating-1.2.0.tar.gz.asc)
+- [Release Notes 1.2.0 (incubating)](releases/RELEASE_NOTES_1.2.0.html)
+- 新特性及重要更新：
+ - 实现了autograd（目前支持MLP模式）。
+ - 升级PySinga以支持Python 3
+ - 改进Tensor类的stride范围。
+ - 将cuDNN从V5升级到V7。
+ - 增加VGG、Inception V4、ResNet和DenseNet进行ImageNet分类。
+ - 为conda包创建别名
+ - 完整的中文文档
+ - 添加在Windows上运行Singa的说明
+ - 更新编译，CI
+ - 修复一些错误
+
+## 孵化版本（incubating） v1.1.0 (12 February 2017):
+
+- [Apache SINGA 1.1.0 (incubating)](https://archive.apache.org/dist/incubator/singa/1.1.0/apache-singa-incubating-1.1.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/1.1.0/apache-singa-incubating-1.1.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/1.1.0/apache-singa-incubating-1.1.0.tar.gz.asc)
+- [Release Notes 1.1.0 (incubating)](releases/RELEASE_NOTES_1.1.0.html)
+- 新特性和重要更新：
+ - 创建Docker镜像(CPU和GPU版本)
+ - 为SINGA创建Amazon AMI（CPU版）。
+ - 集成Jenkins，自动生成Wheel和Debian包（用于安装），并更新网站。
+ - 增强FeedFowardNet，例如，多输入和用于调试的verbose模式。
+ - 添加Concat和Slice层。
+ - 优化CrossEntropyLoss以接受带有多个标签的实例。
+ - 添加包含图像增强方法的image_tool.py。
+ - 支持通过快照API加载和保存模型。
+ - 在Windows上编译SINGA源代码。
+ - 将必要依赖库与SINGA代码一起编译。
+ - 为SINGA启用Java binding（基本）。
+ - 在checkpoint文件中添加版本ID。
+ - 增加Rafiki工具包以提供RESTFul APIs。
+ - 添加了从Caffe预训练的例子，包括GoogleNet。
+
+## 孵化版本（incubating） v1.0.0 (8 September 2016):
+
+- [Apache SINGA 1.0.0 (incubating)](https://archive.apache.org/dist/incubator/singa/1.0.0/apache-singa-incubating-1.0.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/1.0.0/apache-singa-incubating-1.0.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/1.0.0/apache-singa-incubating-1.0.0.tar.gz.asc)
+- [Release Notes 1.0.0 (incubating)](releases/RELEASE_NOTES_1.0.0.html)
+- 新特性和重要更新：
+ - 创建Tensor概念，用于支持更多的机器学习模型。
+ - 创建Device概念，用于运行在不同的硬件设备上，包括CPU，(Nvidia/AMD) GPU 和 FPGA (将在以后的版本中测试)。
+ - 用 cmake 取代 GNU autotool 进行编译。
+ - 支持 Mac OS。
+ - 改进Python binding，包括安装和编程。
+ - 更多的深度学习模型，包括VGG和ResNet。
+ - 更多的IO类用于读取/写入文件和编码/解码数据。
+ - 直接基于Socket的新网络通信组件。
+ - Cudnn V5，包含Dropout和RNN层。
+ - 将网页构建工具从maven更换为Sphinx。
+ - 整合Travis-CI。
+
+## 孵化版本（incubating） v0.3.0 (20 April 2016):
+
+- [Apache SINGA 0.3.0 (incubating)](https://archive.apache.org/dist/incubator/singa/0.3.0/apache-singa-incubating-0.3.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/0.3.0/apache-singa-incubating-0.3.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/0.3.0/apache-singa-incubating-0.3.0.tar.gz.asc)
+- [Release Notes 0.3.0 (incubating)](releases/RELEASE_NOTES_0.3.0.html)
+- 新特性和重要更新：
+ - 在GPU集群上进行训练，可以在GPU集群上进行深度学习模型的训练。
+ - 改进Python wrapper简化配置工作，包括神经网络和SGD算法。
+ - 新增SGD更新器，包括Adam、AdaDelta和AdaMax。
+ - 安装时减少了单节点训练的依赖库。
+ - 使用CPU和GPU进行异构训练。
+ - 支持 cuDNN V4。
+ - 数据预取。
+ - 修复一些bug。
+
+## 孵化版本（incubating） v0.2.0 (14 January 2016):
+
+- [Apache SINGA 0.2.0 (incubating)](https://archive.apache.org/dist/incubator/singa/0.2.0/apache-singa-incubating-0.2.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/0.2.0/apache-singa-incubating-0.2.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/0.2.0/apache-singa-incubating-0.2.0.tar.gz.asc)
+- [Release Notes 0.2.0 (incubating)](releases/RELEASE_NOTES_0.2.0.html)
+- 新特性和重要更新：
+ - 在GPU上进行训练，可以在一个节点上用多个GPU卡训练复杂的模型。
+ - 混合神经网分区支持数据和模型同时并行。
+ - Python wrapper简化配置，包括神经网络和SGD算法。
+ - 实现了RNN模型和BPTT算法，支持基于RNN模型的应用，如GRU。
+ - 云软件集成，包括Mesos、Docker和HDFS。
+ - 可视化神经网结构和层信息，以便优化调试。
+ - 针对Blobs和原始数据指针的线性代数函数和随机函数。
+ - 添加新的层，包括SoftmaxLayer、ArgSortLayer、DummyLayer、RNN层和cuDNN层。
+ - 更新Layer类以携带多个数据/梯度Blobs。
+ - 通过加载之前训练的模型参数，提取特征并测试新数据的性能。
+ - 为IO操作添加Store类。
+
+## Incubating v0.1.0 (8 October 2015):
+
+- [Apache SINGA 0.1.0 (incubating)](https://archive.apache.org/dist/incubator/singa/apache-singa-incubating-0.1.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/apache-singa-incubating-0.1.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/apache-singa-incubating-0.1.0.tar.gz.asc)
+- [Amazon EC2 image](https://console.aws.amazon.com/ec2/v2/home?region=ap-southeast-1#LaunchInstanceWizard:ami=ami-b41001e6)
+- [Release Notes 0.1.0 (incubating)](releases/RELEASE_NOTES_0.1.0.html)
+- 新特性和重要更新：
+ - 允许使用GNU构建工具进行安装。
+ - 实现用zookeeper进行工作管理的脚本。
+ - 实现基于NeuralNet和Layer的编程模型。
+ - 实现基于Worker、Server和Stub的系统架构。
+ - 训练模型来自三种不同的模型类别，即前馈模型、能量模型和RNN模型。
+ - 使用CPU的同步和异步分布式训练框架。
+ - checkpoint文件生成和恢复。
+ - 使用gtest进行单元测试。
diff --git a/docs-site/docs_chs/examples.md b/docs-site/docs_chs/examples.md
new file mode 100644
index 0000000..1d93ea7
--- /dev/null
+++ b/docs-site/docs_chs/examples.md
@@ -0,0 +1,62 @@
+---
+id: examples
+title: Examples
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+本页列出了一些使用SINGA的深度学习任务示例。源代码维护在 [Github](https://github.com/apache/singa/tree/master/examples) 上的 SINGA repo 内。对于使用SINGA Python API在CPU或单GPU上运行的例子，它们也可以在[Google Colab](https://colab.research.google.com/)上获得。你可以直接在谷歌云上运行它们，而无需在本地设置环境。下面给出了每个例子的链接。
+
+## 图像分类
+
+| 网络模型 | 数据集 | 链接 |
+| ----------- | --------------------------------- | ------------------------------------------------------------------------------------------------------- |
+| Simple CNN | MNIST, CIFAR10, CIFAR100 | [Colab](https://colab.research.google.com/drive/1fbGUs1AsoX6bU5F745RwQpohP4bHTktq) |
+| AlexNet | ImageNet | [Cpp]() |
+| VGG | ImageNet | [Cpp](), [Python](), [Colab](https://colab.research.google.com/drive/14kxgRKtbjPCKKsDJVNi3AvTev81Gp_Ds) |
+| XceptionNet | MNIST, CIFAR10, CIFAR100 | [Python]() |
+| ResNet | MNIST, CIFAR10, CIFAR100, CIFAR10 | [Python](), [Colab](https://colab.research.google.com/drive/1u1RYefSsVbiP4I-5wiBKHjsT9L0FxLm9) |
+| MobileNet | ImageNet | [Colab](https://colab.research.google.com/drive/1HsixqJMIpKyEPhkbB8jy7NwNEFEAUWAf) |
+
+## 目标检测
+
+| 网络模型 | 数据集 | 链接 |
+| ----------- | ---------- | ---------------------------------------------------------------------------------- |
+| Tiny YOLOv2 | Pascal VOC | [Colab](https://colab.research.google.com/drive/11V4I6cRjIJNUv5ZGsEGwqHuoQEie6b1T) |
+
+## 面部及表情识别
+
+| 模型 | 数据集 | 链接 |
+| --------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------- |
+| ArcFace | Refined MS-Celeb-1M | [Colab](https://colab.research.google.com/drive/1qanaqUKGIDtifdzEzJOHjEj4kYzA9uJC) |
+| Emotion FerPlus | [Facial Expression Recognition Challenge](https://www.kaggle.com/c/challenges-in-representation-learning-facial-expression-recognition-challenge/data) | [Colab](https://colab.research.google.com/drive/1XHtBQGRhe58PDi4LGYJzYueWBeWbO23r) |
+
+## 图像生成
+
+| 模型 | 数据集 | 链接 |
+| ----- | ------- | ---------------------------------------------------------------------------------- |
+| GAN | MNIST | [Colab](https://colab.research.google.com/drive/1f86MNDW47DJqHoIqWD1tOxcyx2MWys8L) |
+| LSGAN | MNIST | [Colab](https://colab.research.google.com/drive/1C6jNRf28vnFOI9JVM4lpkJPqxsnhxdol) |
+
+## 机器理解
+
+| 模型 | 数据集 | 链接 |
+| ---------- | ------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- |
+| Bert-Squad | [SQuAD v1.1](https://rajpurkar.github.io/SQuAD-explorer/explore/1.1/dev/) | [Colab](https://colab.research.google.com/drive/1kud-lUPjS_u-TkDAzihBTw0Vqr0FjCE-) |
+
+## 文字识别
+
+| 模型 | 数据集 | 链接 |
+| ----------- | ------- | ---------- |
+| Simple LSTM | IMDB | [python]() |
+
+## 文本排序
+
+| 模型 | 数据集 | 链接 |
+| ------ | ----------- | ---------- |
+| BiLSTM | InsuranceQA | [python]() |
+
+## 其他
+
+- MNIST数据集的有限玻尔兹曼机, [source](),
+ [Colab](https://colab.research.google.com/drive/19996noGu9JyHHkVmp4edBGu7PJSRQKsd).
diff --git a/docs-site/docs_chs/git-workflow.md b/docs-site/docs_chs/git-workflow.md
new file mode 100644
index 0000000..d6847bc
--- /dev/null
+++ b/docs-site/docs_chs/git-workflow.md
@@ -0,0 +1,85 @@
+---
+id: git-workflow
+title: Git Workflow
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 对于开发者
+
+1. 将[SINGA Github repository](https://github.com/apache/singa) fork到你自己的Github账户。
+
+2. 从你自己的git仓库中clone **repo** (short for repository):
+
+ ```shell
+ git clone https://github.com/<Github account>/singa.git
+ git remote add upstream https://github.com/apache/singa.git
+ ```
+
+3. 创建一个新的分支（例如 `feature-foo` 或 `fixbug-foo`），在这个分支上工作并提交你的代码:
+
+ ```shell
+ git checkout dev
+ git checkout -b feature-foo
+ # write your code
+ git add <created/updated files>
+ git commit
+ ```
+
+ commit信息应包括：
+
+ - 一个概括性的标题。
+ - 详细的描述。如果提交是为了修复bug，描述中最好包括问题的简短复现；如果是新功能，可以描述新功能的动机/目的。
+
+ 如果您的分支有很多小的commit，您需要通过:
+
+ ```shell
+ git rebase -i <commit id>
+ ```
+ 你可以[压制和重写](https://help.github.com/en/articles/about-git-rebase)提交的内容。
+
+4. 当你在写代码的时候，SINGA的`dev`分支可能已经被别人更新了；在这种情况下，你需要拉取最新的`dev`分支：
+
+ ```shell
+ git checkout dev
+ git pull upstream dev:dev
+ ```
+
+5. 将`feature-foo` [rebase](https://git-scm.com/book/en/v2/Git-Branching-Rebasing)到`dev`分支上，并将提交的内容推送到自己的Github账户（你刚刚创建的新分支），rebase操作是为了清理提交历史。提交当前工作后，应执行以下 git 指令：
+
+ ```shell
+ git checkout feature-foo
+ git rebase dev
+ git push origin feature-foo:feature-foo
+ ```
+
+ Rebase命令的[操作步骤](https://git-scm.com/book/en/v2/Git-Branching-Rebasing)如下: "这个操作的工作原理是进入到两个分支（你所在的分支和你要rebase的分支）的共同来源 -> 获取你所在分支的每次commit所引入的差异 -> 将这些差异保存到临时文件中 -> 将当前分支重置为与你要rebase的分支相同的commit -> 最后依次修改每个差异。"
+
+ 因此，执行后，你还是在特性分支上，但你自己的提交ID/hash被改变了，因为diffs是在rebase时提交的；而且你的分支现在有来自`dev`分支和你自己分支的最新代码。
+
+6. 在 Github 网站上创建一个针对 apache/singa `dev`分支的pull request（PR）。如果你想通知其他在相同文件上工作过的贡献者，你可以在Github上找到文件，然后点击 "Blame"，就可以看到最后修改代码的人的逐行注释。然后，你可以在PR描述中加上@username，就可以立即ping到他们。请说明该贡献是你的原创作品，并且你在项目的开源许可下将该作品授权给项目。你的新分支的进一步提交（例如，bug修复）将由Github自动添加到这个pull request中。
+
+7. 接下来等待committer审核PR。在这段时间里，SINGA的`dev`可能已经被其他人更新了，这时你需要[合并](https://docs.fast.ai/dev/git.html#how-to-keep-your-feature-branch-up-to-date)最新的`dev`来解决冲突。有些人将PR重新[rebase到最新的dev](https://github.com/edx/edx-platform/wiki/How-to-Rebase-a-Pull-Request)上，而不是合并。但是，如果其他开发者获取这个PR来添加新的功能，然后再发送PR，那么rebase操作会在未来的PR中引入**重复的提交**（不同的哈希）。关于何时避免使用rebase的细节，请参见[The Golden Rule of Rebasing](https://www.atlassian.com/git/tutorials/merging-vs-rebasing)。另一种简单的更新PR的方法（修复冲突或提交错误）是，从Apache SINGAS repo的最新开发分支中checkout出一个新的分支，复制并粘贴更新/添加的代码，然后commit并发送一个新的PR。
+
+## 对于Committers
+
+Committer可以将PR合并到上游 repo 的 dev 分支。在合并每一个PR之前，提交者应该做到：
+
+- 检查commit信息(内容和格式)
+- 检查对现有代码的修改，API的变化应该被记录下来
+- 检查Travis测试结果，检查代码/文档格式和单元测试。
+
+合并PR的方式有两种:
+
+- 在Github上，按照[说明](https://gitbox.apache.org/setup/)将你的Apache账户与Github账户链接，之后你就可以直接在GitHub上合并PR了。
+- 通过命令行合并pull request到https://github.com/apache/singa/pull/xxx，应执行以下指令：
+
+ ```shell
+ git clone https://github.com/apache/singa.git
+ git remote add asf https://gitbox.apache.org/repos/asf/singa.git
+ git fetch origin pull/xxx/head:prxxx
+ git checkout dev
+ git merge --no-ff prxxx
+ git push asf dev:dev
+ ```
+ 不要使用rebase来合并PR，并禁用fast forward。
diff --git a/docs-site/docs_chs/graph.md b/docs-site/docs_chs/graph.md
new file mode 100644
index 0000000..73fdda7
--- /dev/null
+++ b/docs-site/docs_chs/graph.md
@@ -0,0 +1,464 @@
+---
+id: graph
+title: Model
+---
+
+<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+神经网络中的前向和反向传播可以用一组操作来表示，比如卷积和池化。每个操作都需要一些输入的[tensors](./tensor)，并应用一个[operator](./autograd)来生成输出的张量。通过将每个运算符表示为一个节点，将每个张量表示为一条边，所有的运算就形成了一个计算图。有了计算图，可以通过调度运算的执行和内存的智能分配/释放来进行速度和内存优化。在SINGA中，用户只需要使用[Model](https://github.com/apache/singa/blob/master/python/singa/model.py) API定义神经网络模型，计算图则会在C++后台自动构建和优化。
+
+
+这样，一方面，用户使用[Model](./graph) API按照PyTorch那样的命令式编程风格实现网络。而与PyTorch在每次迭代中重新创建操作不同的是，SINGA在第一次迭代后就会缓冲操作，隐式地创建计算图（当该功能被启用时）。因此，另一方面，SINGA的计算图与使用声明式编程的库（如TensorFlow）创建的计算图类似，因而它可以享受在图上进行的优化。
+
+## 样例
+
+下面的代码说明了`Model`API的用法：
+
+1. 将新模型实现为Model类的子类：
+
+```Python
+class CNN(model.Model):
+
+ def __init__(self, num_classes=10, num_channels=1):
+ super(CNN, self).__init__()
+ self.conv1 = layer.Conv2d(num_channels, 20, 5, padding=0, activation="RELU")
+ self.conv2 = layer.Conv2d(20, 50, 5, padding=0, activation="RELU")
+ self.linear1 = layer.Linear(500)
+ self.linear2 = layer.Linear(num_classes)
+ self.pooling1 = layer.MaxPool2d(2, 2, padding=0)
+ self.pooling2 = layer.MaxPool2d(2, 2, padding=0)
+ self.relu = layer.ReLU()
+ self.flatten = layer.Flatten()
+ self.softmax_cross_entropy = layer.SoftMaxCrossEntropy()
+
+ def forward(self, x):
+ y = self.conv1(x)
+ y = self.pooling1(y)
+ y = self.conv2(y)
+ y = self.pooling2(y)
+ y = self.flatten(y)
+ y = self.linear1(y)
+ y = self.relu(y)
+ y = self.linear2(y)
+ return y
+
+ def train_one_batch(self, x, y):
+ out = self.forward(x)
+ loss = self.softmax_cross_entropy(out, y)
+ self.optimizer(loss)
+ return out, loss
+```
+
+2. 创建model、optimizer、device等的实例。编译模型：
+
+```python
+model = CNN()
+
+# initialize optimizer and attach it to the model
+sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
+model.set_optimizer(sgd)
+
+# initialize device
+dev = device.create_cuda_gpu()
+
+# input and target placeholders for the model
+tx = tensor.Tensor((batch_size, 1, IMG_SIZE, IMG_SIZE), dev, tensor.float32)
+ty = tensor.Tensor((batch_size, num_classes), dev, tensor.int32)
+
+# compile the model before training
+model.compile([tx], is_train=True, use_graph=True, sequential=False)
+```
+
+3. 迭代训练：
+
+```python
+for b in range(num_train_batch):
+ # generate the next mini-batch
+ x, y = ...
+
+ # Copy the data into input tensors
+ tx.copy_from_numpy(x)
+ ty.copy_from_numpy(y)
+
+ # Training with one batch
+ out, loss = model(tx, ty)
+```
+
+这个例子的Google Colab notebook可以在[这里](https://colab.research.google.com/drive/1fbGUs1AsoX6bU5F745RwQpohP4bHTktq)找到。
+
+
+更多例子：
+
+- [MLP](https://github.com/apache/singa/blob/master/examples/mlp/model.py)
+- [CNN](https://github.com/apache/singa/blob/master/examples/cnn/model/cnn.py)
+- [ResNet](https://github.com/apache/singa/blob/master/examples/cnn/model/resnet.py)
+
+## 实现
+
+### 图的构建
+
+SINGA分三步构建计算图：
+
+1. 将操作保存在缓冲区。
+2. 分析操作的依赖性。
+3. 根据依赖关系创建节点和边。
+
+以MLP模型的dense层的矩阵乘法运算为例，该操作会在[MLP model](https://github.com/apache/singa/blob/master/examples/mlp/model.py)的前向函数中被调用：
+
+```python
+class MLP(model.Model):
+
+ def __init__(self, data_size=10, perceptron_size=100, num_classes=10):
+ super(MLP, self).__init__()
+ self.linear1 = layer.Linear(perceptron_size)
+ ...
+
+ def forward(self, inputs):
+ y = self.linear1(inputs)
+ ...
+```
+
+`线性`层由`mutmul`运算符组成，`autograd`通过SWIG调用CPP中提供的`Mult`函数来实现`matmul`运算符。
+
+```python
+# implementation of matmul()
+singa.Mult(inputs, w)
+```
+
+At the backend, the `Mult` function is implemented by calling `GEMV` a CBLAS
+function. Instead of calling `GEMV` directly, `Mult` submits `GEMV` and the
+arguments to the device as follows,
+在后端，`Mult`函数是通过调用`GEMV`一个CBLAS函数来实现的。但`Mult`没有直接调用`GEMV`，而是将`GEMV`和参数提交给设备，具体如下。
+
+```c++
+// implementation of Mult()
+C->device()->Exec(
+ [a, A, b, B, CRef](Context *ctx) mutable {
+ GEMV<DType, Lang>(a, A, B, b, &CRef, ctx);
+ },
+ read_blocks, {C->block()});
+```
+
+`Device`的`Exec`函数对函数及其参数进行缓冲。此外，它还拥有这个函数要读写的块的信息（块是指张量的内存块）。
+
+一旦`Model.forward()`被执行一次，所有的操作就会被`Device`缓冲。接下来，对所有操作的读写信息进行分析，用来建立计算图。例如，如果一个块`b`被一个操作O1写入，之后又被另一个操作O2读出，我们就会知道O2依赖于O1并且有一条从A到B的有向边，它代表了块`b`（或其张量）。之后我们就构建了一个有向无环图，如下图所示。该图会构建一次。
+
+
+
+<br/>**Figure 1 - MLP例子的计算图**
+
+### 优化
+
+目前，基于计算图进行了以下优化：
+
+**惰性分配** 当创建张量/块时，设备不会立即为它们分配内存。相反，是在第一次访问块时，才会分配内存。
+
+**自动回收** 每个张量/块的参考计数是根据图计算出来的。在执行操作之前，参考计数是读取这个块的操作次数。在执行过程中，一旦执行了一个操作，每一个输入块的参考数就会减少1，如果一个块的参考数达到了0，就意味着这个块在剩下的操作中不会再被读取。因此，它的内存可以被安全释放。此外，SINGA还会跟踪图外的块的使用情况。如果一个块被Python代码使用（而不是被autograd操作符使用），它将不会被回收。
+
+**内存共享** SINGA使用内存池，如[CnMem](https://github.com/NVIDIA/cnmem)来管理CUDA内存。有了自动回收和内存池，SINGA就可以在张量之间共享内存。考虑两个操作`c=a+b`和`d=2xc`。在执行第二个操作之前，根据惰性分配原则，应该分配d的内存。假设`a`在其余操作中没有使用。根据自动回收，`a`的块将在第一次操作后被释放。因此，SINGA会向CUDA流提交四个操作：加法、释放`a`、分配`b`和乘法。这样，内存池就可以将`a`释放的内存与`b`共享，而不是要求GPU为`b`做真正的malloc。
+
+其他的优化技术，如来自编译器的优化技术，如常见的子表达式消除和不同CUDA流上的并行化操作也可以应用。
+
+## 新的操作符
+
+`autograd`模块中定义的每个运算符都实现了两个功能：前向和反向，通过在后台调用运算符来实现。如果要在`autograd`中添加一个新的运算符，需要在后台添加多个运算符。
+
+以[Conv2d](https://github.com/apache/singa/blob/master/python/singa/autograd.py)运算符为例，在Python端，根据设备类型，从后台调用运算符来实现前向和反向功能：
+
+```python
+class _Conv2d(Operation):
+
+ def forward(self, x, W, b=None):
+ ......
+ if training:
+ if self.handle.bias_term:
+ self.inputs = (x, W, b) # record x, W, b
+ else:
+ self.inputs = (x, W)
+
+ if (type(self.handle) != singa.ConvHandle):
+ return singa.GpuConvForward(x, W, b, self.handle)
+ else:
+ return singa.CpuConvForward(x, W, b, self.handle)
+
+ def backward(self, dy):
+ if (type(self.handle) != singa.ConvHandle):
+ dx = singa.GpuConvBackwardx(dy, self.inputs[1], self.inputs[0],
+ self.handle)
+ dW = singa.GpuConvBackwardW(dy, self.inputs[0], self.inputs[1],
+ self.handle)
+ db = singa.GpuConvBackwardb(
+ dy, self.inputs[2],
+ self.handle) if self.handle.bias_term else None
+ else:
+ dx = singa.CpuConvBackwardx(dy, self.inputs[1], self.inputs[0],
+ self.handle)
+ dW = singa.CpuConvBackwardW(dy, self.inputs[0], self.inputs[1],
+ self.handle)
+ db = singa.CpuConvBackwardb(
+ dy, self.inputs[2],
+ self.handle) if self.handle.bias_term else None
+ if db:
+ return dx, dW, db
+ else:
+ return dx, dW
+```
+
+对于后台的每一个操作符，应按以下方式实现：
+
+- 假设操作符是`foo()`，它的真正实现应该包装在另一个函数中，例如`_foo()`。`foo()`将`_foo`和参数一起作为lambda函数传递给`Device`的`Exec`函数进行缓冲，要读和写的块也同时被传递给`Exec`。
+
+- lambda表达式中使用的所有参数都需要根据以下规则获取：
+
+ - `值捕获`: 如果参数变量是一个局部变量，或者将被立刻释放（例如，中间时序）。否则，一旦`foo()`存在，这些变量将被销毁。
+ - `引用捕获`：如果变量是记录在python端或者是一个持久变量（例如Conv2d类中的参数W和ConvHand）。
+
+ - `可变捕获`: 如果在`_foo()`中修改了由值捕获的变量，则lambda表达式应带有mutable（可变）标签。
+
+下面是一个在后台实现的操作的[例子](https://github.com/apache/singa/blob/master/src/model/operation/convolution.cc)：
+
+```c++
+Tensor GpuConvBackwardx(const Tensor &dy, const Tensor &W, const Tensor &x,
+ const CudnnConvHandle &cch) {
+ CHECK_EQ(dy.device()->lang(), kCuda);
+
+ Tensor dx;
+ dx.ResetLike(x);
+
+ dy.device()->Exec(
+ /*
+ * dx is a local variable so it's captured by value
+ * dy is an intermediate tensor and isn't recorded on the python side
+ * W is an intermediate tensor but it's recorded on the python side
+ * chh is a variable and it's recorded on the python side
+ */
+ [dx, dy, &W, &cch](Context *ctx) mutable {
+ Block *wblock = W.block(), *dyblock = dy.block(), *dxblock = dx.block();
+ float alpha = 1.f, beta = 0.f;
+ cudnnConvolutionBackwardData(
+ ctx->cudnn_handle, &alpha, cch.filter_desc, wblock->data(),
+ cch.y_desc, dyblock->data(), cch.conv_desc, cch.bp_data_alg,
+ cch.workspace.block()->mutable_data(),
+ cch.workspace_count * sizeof(float), &beta, cch.x_desc,
+ dxblock->mutable_data());
+ },
+ {dy.block(), W.block()}, {dx.block(), cch.workspace.block()});
+ /* the lambda expression reads the blocks of tensor dy and w
+ * and writes the blocks of tensor dx and chh.workspace
+ */
+
+ return dx;
+}
+```
+
+## Benchmark
+
+### 单节点
+
+- 实验设定
+ - 模型：
+ - 使用层: ResNet50 in
+ [resnet.py](https://github.com/apache/singa/blob/master/examples/cnn/autograd/resnet_cifar10.py)
+ - 使用模型: ResNet50 in
+ [resnet.py](https://github.com/apache/singa/blob/master/examples/cnn/model/resnet.py)
+ - GPU: NVIDIA RTX 2080Ti
+- 注释：
+ - `s` ：second，秒
+ - `it` ： iteration，迭代次数
+ - `Mem`：peak memory usage of single GPU，单GPU显存峰值
+ - `Throughout`：number of images processed per second，每秒处理的图像数
+ - `Time`：total time，总时间
+ - `Speed`：iterations per second。每秒迭代次数
+ - `Reduction`：the memory usage reduction rate compared with that using layer，与使用层的内存使用率相比，内存使用率降低了多少
+ - `Speedup`: speedup ratio compared with dev branch，与dev分支相比的加速率
+- 结果：
+ <table style="text-align: center">
+ <tr>
+ <th style="text-align: center">Batchsize</th>
+ <th style="text-align: center">Cases</th>
+ <th style="text-align: center">Mem(MB)</th>
+ <th style="text-align: center">Time(s)</th>
+ <th style="text-align: center">Speed(it/s)</th>
+ <th style="text-align: center">Throughput</th>
+ <th style="text-align: center">Reduction</th>
+ <th style="text-align: center">Speedup</th>
+ </tr>
+ <tr>
+ <td rowspan="4">16</td>
+ <td nowrap>layer</td>
+ <td>4975</td>
+ <td>14.1952</td>
+ <td>14.0893</td>
+ <td>225.4285</td>
+ <td>0.00%</td>
+ <td>1.0000</td>
+ </tr>
+ <tr>
+ <td nowrap>model:disable graph</td>
+ <td>4995</td>
+ <td>14.1264</td>
+ <td>14.1579</td>
+ <td>226.5261</td>
+ <td>-0.40%</td>
+ <td>1.0049</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, bfs</td>
+ <td>3283</td>
+ <td>13.7438</td>
+ <td>14.5520</td>
+ <td>232.8318</td>
+ <td>34.01%</td>
+ <td>1.0328</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, serial</td>
+ <td>3265</td>
+ <td>13.7420</td>
+ <td>14.5540</td>
+ <td>232.8635</td>
+ <td>34.37%</td>
+ <td>1.0330</td>
+ </tr>
+ <tr>
+ <td rowspan="4">32</td>
+ <td nowrap>layer</td>
+ <td>10119</td>
+ <td>13.4587</td>
+ <td>7.4302</td>
+ <td>237.7649</td>
+ <td>0.00%</td>
+ <td>1.0000</td>
+ </tr>
+ <tr>
+ <td nowrap>model:disable graph</td>
+ <td>10109</td>
+ <td>13.2952</td>
+ <td>7.5315</td>
+ <td>240.6875</td>
+ <td>0.10%</td>
+ <td>1.0123</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, bfs</td>
+ <td>6839</td>
+ <td>13.1059</td>
+ <td>7.6302</td>
+ <td>244.1648</td>
+ <td>32.41%</td>
+ <td>1.0269</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, serial</td>
+ <td>6845</td>
+ <td>13.0489</td>
+ <td>7.6635</td>
+ <td>245.2312</td>
+ <td>32.35%</td>
+ <td>1.0314</td>
+ </tr>
+ </table>
+
+### 多线程
+
+- 实验设置：
+ - API：
+ - 使用层: ResNet50 in
+ [resnet_dist.py](https://github.com/apache/singa/blob/master/examples/cnn/autograd/resnet_dist.py)
+ - 使用模型: ResNet50 in
+ [resnet.py](https://github.com/apache/singa/blob/master/examples/cnn/model/resnet.py)
+ - GPU: NVIDIA RTX 2080Ti \* 2
+ - MPI: 在同一节点上的两个MPI processes
+- 注释: 与上面相同
+- 结果：
+ <table style="text-align: center">
+ <tr>
+ <th style="text-align: center">Batchsize</th>
+ <th style="text-align: center">Cases</th>
+ <th style="text-align: center">Mem(MB)</th>
+ <th style="text-align: center">Time(s)</th>
+ <th style="text-align: center">Speed(it/s)</th>
+ <th style="text-align: center">Throughput</th>
+ <th style="text-align: center">Reduction</th>
+ <th style="text-align: center">Speedup</th>
+ </tr>
+ <tr>
+ <td rowspan="4">16</td>
+ <td nowrap>layer</td>
+ <td>5439</td>
+ <td>17.3323</td>
+ <td>11.5391</td>
+ <td>369.2522</td>
+ <td>0.00%</td>
+ <td>1.0000</td>
+ </tr>
+ <tr>
+ <td nowrap>model:disable graph</td>
+ <td>5427</td>
+ <td>17.8232</td>
+ <td>11.2213</td>
+ <td>359.0831</td>
+ <td>0.22%</td>
+ <td>0.9725</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, bfs</td>
+ <td>3389</td>
+ <td>18.2310</td>
+ <td>10.9703</td>
+ <td>351.0504</td>
+ <td>37.69%</td>
+ <td>0.9507</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, serial</td>
+ <td>3437</td>
+ <td>17.0389</td>
+ <td>11.7378</td>
+ <td>375.6103</td>
+ <td>36.81%</td>
+ <td>1.0172</td>
+ </tr>
+ <tr>
+ <td rowspan="4">32</td>
+ <td nowrap>layer</td>
+ <td>10547</td>
+ <td>14.8635</td>
+ <td>6.7279</td>
+ <td>430.5858</td>
+ <td>0.00%</td>
+ <td>1.0000</td>
+ </tr>
+ <tr>
+ <td nowrap>model:disable graph</td>
+ <td>10503</td>
+ <td>14.7746</td>
+ <td>6.7684</td>
+ <td>433.1748</td>
+ <td>0.42%</td>
+ <td>1.0060</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, bfs</td>
+ <td>6935</td>
+ <td>14.8553</td>
+ <td>6.7316</td>
+ <td>430.8231</td>
+ <td>34.25%</td>
+ <td>1.0006</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, serial</td>
+ <td>7027</td>
+ <td>14.3271</td>
+ <td>6.9798</td>
+ <td>446.7074</td>
+ <td>33.37%</td>
+ <td>1.0374</td>
+ </tr>
+ </table>
+
+### 结论
+
+- 在启用计算图的情况下进行训练，可以显著减少内存占用。
+- 目前，在速度上有一点改进。在效率方面还可以做更多的优化。
diff --git a/docs-site/docs_chs/history-singa.md b/docs-site/docs_chs/history-singa.md
new file mode 100644
index 0000000..0a24307
--- /dev/null
+++ b/docs-site/docs_chs/history-singa.md
@@ -0,0 +1,37 @@
+---
+id: history-singa
+title: History of SINGA
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 历史
+
+SINGA由新加坡国立大学DB System Group于2014年发起，与浙江大学数据库组合作。如果您在研究中使用SINGA，请引用以下两篇论文。
+
+- B.C. Ooi, K.-L. Tan, S. Wang, W. Wang, Q. Cai, G. Chen, J. Gao, Z. Luo, A. K.
+ H. Tung, Y. Wang, Z. Xie, M. Zhang, and K. Zheng.
+ [SINGA: A distributed deep learning platform](http://www.comp.nus.edu.sg/~ooibc/singaopen-mm15.pdf).
+ ACM Multimedia (Open Source Software Competition) 2015
+
+- W. Wang, G. Chen, T. T. A. Dinh, B. C. Ooi, K.-L.Tan, J. Gao, and S. Wang.
+ [SINGA: putting deep learning in the hands of multimedia users](http://www.comp.nus.edu.sg/~ooibc/singa-mm15.pdf).
+ ACM Multimedia 2015.
+
+Rafiki是SINGA的一个子模块，如果您在研究中使用Rafiki，请引用以下论文。
+
+- Wei Wang, Jinyang Gao, Meihui Zhang, Sheng Wang, Gang Chen, Teck Khim Ng, Beng
+ Chin Ooi, Jie Shao, Moaz Reyad.
+ [Rafiki: Machine Learning as an Analytics Service System](http://www.vldb.org/pvldb/vol12/p128-wang.pdf).
+ [VLDB 2019](http://vldb.org/2019/)
+ ([BibTex](https://dblp.org/rec/bib2/journals/pvldb/WangWGZCNOS18.bib)).
+
+[NetEase](http://tech.163.com/17/0602/17/CLUL016I00098GJ5.html),
+[yzBigData](http://www.yzbigdata.com/en/index.html),
+[Shentilium](https://shentilium.com/), [Foodlg](http://www.foodlg.com/) 以及
+[Medilot](https://medilot.com/technologies) 等公司在他们的工作和应用中使用了SINGA。
+
+## License
+
+SINGA以[Apache License Version 2.0](http://www.apache.org/licenses/LICENSE-2.0)
+版本发布。
diff --git a/docs-site/docs_chs/how-to-release.md b/docs-site/docs_chs/how-to-release.md
new file mode 100644
index 0000000..887b140
--- /dev/null
+++ b/docs-site/docs_chs/how-to-release.md
@@ -0,0 +1,172 @@
+---
+id: how-to-release
+title: How to Prepare a Release
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+这是SINGA的[发布准备流程](http://www.apache.org/dev/release-publishing.html)指南。
+
+1. 选择一个发布管理者。发布管理者（RM）是发布过程的协调者，他的签名文件（.asc）将会与发布一起上传。RM 生成 KEY (RSA 4096 位)并将其上传到公钥服务器，首先需要得到其他Apache用户对他的密钥的认可（签名），才能连接到信任网，RM需要先求助其他项目管理者帮忙认证他的密钥。[如何生成密钥？](http://www.apache.org/dev/release-signing.html)
+
+2. 检查license。 [FAQ](https://www.apache.org/legal/src-headers.html#faq-docs);
+ [SINGA Issue](https://issues.apache.org/jira/projects/SINGA/issues/SINGA-447)
+
+ - 代码库不能包含与APL不兼容的第三方代码。
+ - 依赖项与APL兼容，GNU类license不兼容。
+ - 我们编写的所有源文件都必须包含Apache license头：http://www.apache.org/legal/src-headers.html.
+ 链接中有一个脚本可以帮助将这个头同步到所有文件。
+ - 更新LICENSE文件。如果我们在发行包中包含了任何非APL的第三方代码，必须要在NOTICE文件的最后注明。
+
+3. 复查版本。检查代码和文档。
+
+ - 编译过程无错误。
+ - (尽可能地)包含进单元测试。
+ - Conda包运行无误。
+ - Apache网站上的在线文档是最新的。
+
+4. 准备好RELEASE_NOTES文件。包括以下项目，介绍，特性，错误（链接到JIRA或Github PR），变更，依赖列表，不兼容问题，可以按照这个[例子]((http://commons.apache.org/proper/commons-digester/commons-digester-3.0/RELEASE-NOTES.txt))来写。
+
+5. 打包候选版本。该版本应该打包成：apache-singa-VERSION.tar.gz。这个版本不应该包含任何二进制文件，包括git文件。但是CMake的编译依赖于git标签来获取版本号；要删除这个依赖，你需要手动更新CMakeLists.txt文件来设置版本号：
+
+ ```
+ # remove the following lines
+ include(GetGitRevisionDescription)
+ git_describe(VERSION --tags --dirty=-d)
+ string(REGEX REPLACE "^([0-9]+)\\..*" "\\1" VERSION_MAJOR "${VERSION}")
+ string(REGEX REPLACE "^[0-9]+\\.([0-9]+).*" "\\1" VERSION_MINOR "${VERSION}")
+ string(REGEX REPLACE "^[0-9]+\\.[0-9]+\\.([0-9]+).*" "\\1" VERSION_PATCH "${VERSION}")
+
+ # set the numbers manually
+ SET(PACKAGE_VERSION 3.0.0)
+ SET(VERSION 3.0.0)
+ SET(SINGA_MAJOR_VERSION 3) # 0 -
+ SET(SINGA_MINOR_VERSION 0) # 0 - 9
+ SET(SINGA_PATCH_VERSION 0) # 0 - 99
+ ```
+
+ 将软件包上传到[stage repo](https://dist.apache.org/repos/dist/dev/singa/)。应包括tar文件、签名、KEY和SHA256校验和文件。MD5不再使用，详细规则在[这里](https://dist.apache.org/repos/dist/dev/singa/)。阶段文件夹应该包括：
+
+ - apache-singa-VERSION.tar.gz
+ - apache-singa-VERSION.acs
+ - apache-singa-VERSION.SHA256
+
+ 创建这些文件并上传到stage svn repo的命令如下：
+
+ ```sh
+ # in singa repo
+ rm -rf .git
+ rm -rf rafiki/*
+ cd ..
+ tar -czvf apache-singa-VERSION.tar.gz singa/
+
+ mkdir stage
+ cd stage
+ svn co https://dist.apache.org/repos/dist/dev/singa/
+ cd singa
+ # copy the KEYS file from singa repo to this folder if it is not here
+ cp ../../singa/KEYS .
+ mkdir VERSION
+ # copy the tar.gz file
+ mv ../../apache-singa-VERSION.tar.gz VERSION/
+ cd VERSION
+ sha512sum apache-singa-VERSION.tar.gz > apache-singa-VERSION.tar.gz.sha512
+ gpg --armor --output apache-singa-VERSION.tar.gz.asc --detach-sig apache-singa-VERSION.tar.gz
+ cd ..
+ svn add VERSION
+ svn commit
+ ```
+
+6) 通过发送电子邮件的方式进行投票。现举例如下：
+
+```
+ To: dev@singa.apache.org
+ Subject: [VOTE] Release apache-singa-X.Y.Z (release candidate N)
+
+ Hi all,
+
+ I have created a build for Apache SINGA 3.1.0, release candidate 2.
+
+ The release note is at
+ https://github.com/apache/singa/blob/master/RELEASE_NOTES.
+
+ The artifacts to be voted on are located here:
+ https://dist.apache.org/repos/dist/dev/singa/3.1.0.rc2/apache-singa-3.1.0.rc2.tar.gz
+
+ The hashes of the artifacts are as follows:
+ SHA512: 84545499ad36da108c6a599edd1d853f82d331bc03273b5278515554866f0c698e881f956b2eabcb6b29c07fa9fa4ff1add5a777b58db8a6a2362cf383b5c04d
+
+ Release artifacts are signed with the followingkey:
+ https://dist.apache.org/repos/dist/dev/singa/KEYS
+
+ The signature file is:
+ https://dist.apache.org/repos/dist/dev/singa/3.1.0.rc2/apache-singa-3.1.0.rc2.tar.gz.asc
+
+ The Github tag is at:
+ https://github.com/apache/singa/releases/tag/3.1.0.rc2
+
+ The documentation website is at
+ http://singa.apache.org/docs/next/installation/
+
+ Some examples are available for testing:
+ https://github.com/apache/singa/tree/master/examples
+
+
+
+ Please vote on releasing this package. The vote is open for at least 72 hours and passes if a majority of at least three +1 votes are cast.
+
+ [ ] +1 Release this package as Apache SINGA X.Y.Z
+
+ [ ] 0 I don't feel strongly about it, but I'm okay with the release
+
+ [ ] -1 Do not release this package because...
+
+ Here is my vote: +1
+
+```
+
+7) 等待至少48小时的测试回复。任何PMC、提交者或贡献者都可以测试发布的功能，以及反馈。大家在投票+1之前应该检查这些。如果投票通过，则发送如下的结果邮件，否则，从头开始重复刚刚的步骤。
+
+```
+
+To: dev@singa.apache.org Subject: [RESULT][vote] Release apache-singa-X.Y.Z
+(release candidate N)
+
+Thanks to everyone who has voted and given their comments. The tally is as
+follows.
+
+N binding +1s: <names>
+
+N non-binding +1s: <names>
+
+No 0s or -1s.
+
+I am delighted to announce that the proposal to release Apache SINGA X.Y.Z has
+passed.
+
+````
+
+8) 将软件包上传至 https://dist.apache.org/repos/dist/release/singa/，以便[distribution](http://www.apache.org/dev/release-publishing.html#distribution)。
+
+9) 更新SINGA网站的下载页面。tar.gz 文件必须从镜像下载，使用 closer.cgi 脚本；其他工件必须从 Apache 主站点下载。更多细节请看[这里](http://www.apache.org/dev/release-download-pages.html)。我们在之前的版本中得到的一些反馈。“下载页面必须只链接到正式发布的版本，所以不能包含到GitHub的链接”，“链接到KEYS, sig和Hash的链接不能使用dist.apache.org而应该使用 https://www.apache.org/dist/singa/...”“而且你只需要一个KEYS链接，而且应该描述如何使用KEYS+sig或Hash来验证下载。”
+
+10) 删除RC标签并编译conda包。
+
+11) 发布release信息：
+
+ ```
+ To: announce@apache.org, dev@singa.apache.org
+ Subject: [ANNOUNCE] Apache SINGA X.Y.Z released
+
+ We are pleased to announce that SINGA X.Y.Z is released.
+
+ SINGA is a general distributed deep learning platform
+ for training big deep learning models over large datasets.
+ The release is available at: http://singa.apache.org/downloads.html
+ The main features of this release include XXX
+ We look forward to hearing your feedback, suggestions,
+ and contributions to the project.
+
+ On behalf of the SINGA team, {SINGA Team Member Name}
+ ```
+
diff --git a/docs-site/docs_chs/install-win.md b/docs-site/docs_chs/install-win.md
new file mode 100644
index 0000000..5bb9c8c
--- /dev/null
+++ b/docs-site/docs_chs/install-win.md
@@ -0,0 +1,359 @@
+---
+id: install-win
+title: Build SINGA on Windows
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+在Microsoft Windows上从源码构建SINGA的过程有四个部分：安装依赖关系、构建SINGA源码、（可选择）安装python模块和（可选择）运行单元测试。
+
+## 安装依赖项
+
+你可以创建一个文件夹来构建依赖关系。
+
+使用到的依赖项有：
+
+- 编译器和IDE：
+ - Visual Studio，社区版是免费的，可以用来构建SINGA。
+ https://www.visualstudio.com/
+- CMake
+ - 可以从 http://cmake.org/ 下载
+ - 确保 cmake 可执行文件的路径在系统路径中，或者在调用 cmake 时使用完整路径。
+- SWIG
+ - 可以从 http://swig.org/ 下载
+ - 确保swig可执行文件的路径在系统路径中，或者在调用swig时使用完整路径。请使用最新的版本，如3.0.12。
+
+- Protocol Buffers
+ - 下载一个合适的版本，如2.6.1:
+ https://github.com/google/protobuf/releases/tag/v2.6.1 。
+ - 下载 protobuf-2.6.1.zip 和 protoc-2.6.1-win32.zip。
+ - 将这两个文件解压到dependencies文件夹中，将protoc可执行文件的路径添加到系统路径中，或者在调用它时使用完整路径。
+ - 打开Visual Studio solution，它可以在vsproject文件夹中找到。
+ - 将build settings改为Release和x64。
+ - 构建libprotobuf项目。
+
+- Openblas
+ - 从 http://www.openblas.net 下载合适的源码，如0.2.20。
+ - 将源码解压到dependencies文件夹中。
+ - 如果你没有安装Perl，请下载一个perl环境，如Strawberry Perl (http://strawberryperl.com/)。
+ - 在源文件夹中运行此命令来构建Visual Studio解决方案：
+
+ ```bash
+ cmake -G "Visual Studio 15 2017 Win64"
+ ```
+
+ - 打开Visual Studio解决方案并将build settings更改为Release和x64。
+ - 构建libopenblas项目。
+
+- Google glog
+ - 下载一个合适的版本，如0.3.5:
+ https://github.com/google/glog/releases
+ - 将源码解压到dependencies文件夹中。
+ - 打开Visual Studio solution.
+ - 将build settings改为Release and x64.
+ - 构建libglog项目。
+
+## 构建SINGA源代码
+
+- 下载SINGA源代码
+- 编译protobuf文件:
+
+ - 在src/proto目录中：
+
+ ```shell
+ mkdir python_out
+ protoc.exe *.proto --python_out python_out
+ ```
+
+- 为C++和Python生成swig接口：在src/api目录中：
+
+ ```shell
+ swig -python -c++ singa.i
+ ```
+
+- 生成SINGA的Visual Studio解决方案：在SINGA源码根目录中：
+
+ ```shell
+ mkdir build
+ cd build
+ ```
+
+- 调用 cmake 并添加系统路径，类似于下面的例子:
+
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64" ^
+ -DGLOG_INCLUDE_DIR="D:/WinSinga/dependencies/glog-0.3.5/src/windows" ^
+ -DGLOG_LIBRARIES="D:/WinSinga/dependencies/glog-0.3.5/x64/Release" ^
+ -DCBLAS_INCLUDE_DIR="D:/WinSinga/dependencies/openblas-0.2.20/lapack-netlib/CBLAS/include" ^
+ -DCBLAS_LIBRARIES="D:/WinSinga/dependencies/openblas-0.2.20/lib/RELEASE" ^
+ -DProtobuf_INCLUDE_DIR="D:/WinSinga/dependencies/protobuf-2.6.1/src" ^
+ -DProtobuf_LIBRARIES="D:/WinSinga/dependencies/protobuf-2.6.1/vsprojects/x64/Release" ^
+ -DProtobuf_PROTOC_EXECUTABLE="D:/WinSinga/dependencies/protoc-2.6.1-win32/protoc.exe" ^
+ ..
+ ```
+
+- 在Visual Studio中打开生成的解决方案。
+- 将构建设置改为Release和x64。
+- 将src/api中的singa_wrap.cxx文件添加到singa_objects项目中。
+- 在 singa_objects 项目中，打开 Additional Include Directories。
+- 添加Python的include路径。
+- 添加numpy的include路径。
+- 添加protobuf的include路径。
+- 在 singa_objects 项目的预处理程序定义中， 添加 USE_GLOG。
+- 构建singa_objects项目。
+
+- 在singa项目中:
+ - 将singa_wrap.obj添加到对象库。
+ - 将目标名称改为"_singa_wrap"。
+ - 将目标扩展名为.pyd。
+ - 将配置类型改为动态库(.dll)。
+ - 进入Additional Library Directories，添加路径到python、openblas、protobuf和glog库。
+ - 在Additional Dependencies中添加libopenblas.lib、libglog.lib和libprotobuf.lib。
+
+- 构建singa项目
+
+## 安装python模块
+
+
+- 将build/python/setup.py中的`_singa_wrap.so`改为`_singa_wrap.pyd`。
+- 将`src/proto/python_out`中的文件复制到`build/python/singa/proto`中。
+
+- （可选）创建并激活一个虚拟环境：
+ ```shell
+ mkdir SingaEnv
+ virtualenv SingaEnv
+ SingaEnv\Scripts\activate
+ ```
+
+- 进入build/python文件夹，运行:
+
+ ```shell
+ python setup.py install
+ ```
+
+- 将 _singa_wrap.pyd、libglog.dll 和 libopenblas.dll 添加到路径中，或者将它们复制到 python site-packages 中的 singa package 文件夹中。
+
+
+- 通过运行如下命令，来验证SINGA是否安装成功：
+
+ ```shell
+ python -c "from singa import tensor"
+ ```
+
+构建过程的视频教程可以在这里找到：
+
+[](https://www.youtube.com/watch?v=cteER7WeiGk)
+
+## 运行单元测试
+
+- 在测试文件夹中，生成Visual Studio解决方案：
+
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64"
+ ```
+
+- 在Visual Studio中打开生成的解决方案。
+
+- 更改build settings为Release和x64。
+
+- 构建glog项目。
+
+- 在test_singa项目中:
+ - 将 USE_GLOG 添加到Preprocessor Definitions中。
+ - 在 Additional Include Directories 中， 添加上面第 2 步中使用的 GLOG_INCLUDE_DIR、 CBLAS_INCLUDE_DIR 和 Protobuf_INCLUDE_DIR 的路径。同时添加build和build/include文件夹。
+ - 转到Additional Library Directories，添加openblas、protobuf和glog库的路径。同时添加 build/src/singa_objects.dir/Release。
+ - 转到 Additional Dependencies 并添加 libopenblas.lib、libglog.lib 和 libprotobuf.lib。修改两个库的名字：gtest.lib和singa_objects.lib。
+
+- 构建test_singa项目。
+
+- 将libglog.dll和libopenblas.dll添加到路径中，或者将它们复制到test/release文件夹中，使其可用。
+
+- 单元测试可以通过如下方式执行：
+
+ - 从命令行:
+
+ ```shell
+ test_singa.exe
+ ```
+
+ - 从Visual Studio:
+ - 右键点击test_singa项目，选择 "Set as StartUp Project"。
+ - 在Debug菜单中，选择'Start Without Debugging'。
+
+单元测试的视频教程可以在这里找到:
+
+[](https://www.youtube.com/watch?v=393gPtzMN1k)
+
+## 构建包含cuda的GPU支持
+
+在本节中，我们将扩展前面的步骤来启用GPU。
+
+### 安装依赖项
+
+除了上面第1节的依赖关系外，我们还需要以下内容：
+
+- CUDA
+
+ 从 https://developer.nvidia.com/cuda-downloads 下载一个合适的版本，如9.1。确保已经安装了Visual Studio集成模块。
+
+
+- cuDNN
+
+ 从 https://developer.nvidia.com/cudnn 下载一个合适的版本，如7.1。
+
+- cnmem:
+
+ - 从 https://github.com/NVIDIA/cnmem 下载最新版本。
+ - 构建Visual Studio解决方案：
+
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64"
+ ```
+
+ - 在Visual Studio中打开生成的解决方案。
+ - 将build settings改为Release和x64。
+ - 构建cnmem项目。
+
+### 构建SINGA源代码
+
+- 调用 cmake 并添加系统路径，类似于下面的例子：
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64" ^
+ -DGLOG_INCLUDE_DIR="D:/WinSinga/dependencies/glog-0.3.5/src/windows" ^
+ -DGLOG_LIBRARIES="D:/WinSinga/dependencies/glog-0.3.5/x64/Release" ^
+ -DCBLAS_INCLUDE_DIR="D:/WinSinga/dependencies/openblas-0.2.20/lapack-netlib/CBLAS/include" ^
+ -DCBLAS_LIBRARIES="D:/WinSinga/dependencies/openblas-0.2.20/lib/RELEASE" ^
+ -DProtobuf_INCLUDE_DIR="D:/WinSinga/dependencies/protobuf-2.6.1/src" ^
+ -DProtobuf_LIBRARIES="D:\WinSinga/dependencies/protobuf-2.6.1/vsprojects/x64/Release" ^
+ -DProtobuf_PROTOC_EXECUTABLE="D:/WinSinga/dependencies/protoc-2.6.1-win32/protoc.exe" ^
+ -DCUDNN_INCLUDE_DIR=D:\WinSinga\dependencies\cudnn-9.1-windows10-x64-v7.1\cuda\include ^
+ -DCUDNN_LIBRARIES=D:\WinSinga\dependencies\cudnn-9.1-windows10-x64-v7.1\cuda\lib\x64 ^
+ -DSWIG_DIR=D:\WinSinga\dependencies\swigwin-3.0.12 ^
+ -DSWIG_EXECUTABLE=D:\WinSinga\dependencies\swigwin-3.0.12\swig.exe ^
+ -DUSE_CUDA=YES ^
+ -DCUDNN_VERSION=7 ^
+ ..
+ ```
+
+* 为C++和Python生成swig接口。在src/api目录中：
+
+ ```shell
+ swig -python -c++ singa.i
+ ```
+
+* 在Visual Studio中打开生成的解决方案
+
+* 将build settings改为Release和x64
+
+#### 构建singa_objects
+
+- 将src/api中的singa_wrap.cxx文件添加到singa_objects项目中。
+- 在 singa_objects 项目中，打开 Additional Include Directories。
+- 添加Python的include路径
+- 添加numpy include路径
+- 添加protobuf包括路径
+- 增加CUDA、cuDNN和cnmem的包含路径。
+- 在 singa_objects 项目的预处理程序定义中， 加入 USE_GLOG、 USE_CUDA 和 USE_CUDNN。删除 DISABLE_WARNINGS。
+- 建立 singa_objects 项目
+
+#### 构建singa-kernel
+
+
+- 创建一个新的Visual Studio项目，类型为 "CUDA 9.1 Runtime"。给它起个名字，比如singa-kernel。
+- 该项目自带一个名为kernel.cu的初始文件，从项目中删除这个文件。
+- 添加这个文件：src/core/tensor/math_kernel.cu。
+- 在项目设置中。
+
+ - 将平台工具集设置为 "Visual Studio 2015 (v140)"
+ - 将 "配置类型 "设置为 "静态库(.lib)"
+ - 在include目录中，添加build/include。
+
+- 建立singa-kernel项目
+
+#### 构建singa
+
+- 在singa项目中：
+
+ - 将singa_wrap.obj添加到对象库中。
+ - 将目标名称改为"_singa_wrap"。
+ - 将目标扩展名为.pyd。
+ - 将配置类型改为动态库(.dll)。
+ - 到Additional Library Directories中添加python、openblas的路径。protobuf和glog库。
+ - 同时添加singa-kernel、cnmem、cuda和cudnn的library path。
+ - 到Additional Dependencies，并添加libopenblas.lib、libglog.lib和 libprotobuf.lib。
+ - 另外还要添加：singa-kernel.lib、cnmem.lib、cudnn.lib、cuda.lib、cublas.lib。curand.lib和cudart.lib。
+
+- 构建singa项目。
+
+### Install Python module
+
+- 将 build/python/setup.py 中的 _singa_wrap.so 改为 _singa_wrap.pyd。
+
+- 将 src/proto/python_out 中的文件复制到 build/python/singa/proto 中。
+
+- （可选） 创建并激活虚拟环境:
+
+ ```shell
+ mkdir SingaEnv
+ virtualenv SingaEnv
+ SingaEnv\Scripts\activate
+ ```
+
+- 进入build/python文件夹，运行:
+
+ ```shell
+ python setup.py install
+ ```
+
+- 将 _singa_wrap.pyd, libglog.dll, libopenblas.dll, cnmem.dll, CUDA Runtime (例如 cudart64_91.dll) 和 cuDNN (例如 cudnn64_7.dll) 添加到路径中，或者将它们复制到 python site-packages 中的 singa package 文件夹中。
+
+- 通过运行如下命令来验证SINGA是否已经安装：
+
+ ```shell
+ python -c "from singa import device; dev = device.create_cuda_gpu()"
+ ```
+
+这个部分的视频教程可以在这里找到：
+
+[](https://www.youtube.com/watch?v=YasKVjRtuDs)
+
+### 运行单元测试
+
+- 在测试文件夹中，生成Visual Studio解决方案：
+
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64"
+ ```
+
+- 在Visual Studio中打开生成的解决方案，或者将项目添加到步骤5.2中创建的singa解决方案中。
+
+- 将build settings改为Release和x64。
+
+- 构建 glog 项目。
+
+- 在test_singa项目中:
+
+ - 将 USE_GLOG; USE_CUDA; USE_CUDNN 添加到Preprocessor Definitions中。
+ - 在 Additional Include Directories 中， 添加上面 5.2 中使用的 GLOG_INCLUDE_DIR、 CBLAS_INCLUDE_DIR 和 Protobuf_INCLUDE_DIR 的路径。同时添加build、build/include、CUDA和cuDNN的include文件夹。
+ - 转到Additional Library Directories，添加openblas、protobuf和glog库的路径。同时添加 build/src/singa_objects.dir/Release、singa-kernel、cnmem、CUDA 和 cuDNN 库的路径。
+ - 在Additional Dependencies中添加libopenblas.lib; libglog.lib; libprotobuf.lib; cnmem.lib; cudnn.lib; cuda.lib; cublas.lib; curand.lib; cudart.lib; singa-kernel.lib。修正两个库的名字：gtest.lib和singa_objects.lib。
+
+* 构建.
+
+* 将libglog.dll、libopenblas.dll、cnmem.dll、cudart64_91.dll和cudnn64_7.dll添加到路径中，或将它们复制到test/release文件夹中，使其可用。
+
+* 单元测试可以通过如下方式执行：
+
+ - 从命令行:
+
+ ```shell
+ test_singa.exe
+ ```
+
+ - 从 Visual Studio:
+ - 右键点击test_singa项目，选择 'Set as StartUp Project'.
+ - 从Debug菜单，选择 'Start Without Debugging'
+
+运行单元测试的视频教程可以在这里找到：
+
+[](https://www.youtube.com/watch?v=YOjwtrvTPn4)
diff --git a/docs-site/docs_chs/installation.md b/docs-site/docs_chs/installation.md
new file mode 100644
index 0000000..49ce35e
--- /dev/null
+++ b/docs-site/docs_chs/installation.md
@@ -0,0 +1,150 @@
+---
+id: installation
+title: Installation
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 使用Conda
+
+Conda是一个Python、CPP等包的包管理器。
+
+目前，SINGA有Linux和MacOSX的conda包。推荐使用[Miniconda3](https://conda.io/miniconda.html)来配合SINGA使用，安装miniconda后，执行以下命令之一安装SINGA。
+
+1. 只使用CPU
+ [](https://colab.research.google.com/drive/1Ntkhi-Z6XTR8WYPXiLwujHd2dOm0772V?usp=sharing)
+
+```shell
+$ conda install -c nusdbsystem -c conda-forge singa-cpu
+```
+
+2. 使用带CUDA和cuDNN的GPU（需要CUDA驱动>=384.81）
+ [](https://colab.research.google.com/drive/1do_TLJe18IthLOnBOsHCEe-FFPGk1sPJ?usp=sharing)
+
+```shell
+$ conda install -c nusdbsystem -c conda-forge singa-gpu
+```
+
+3. 安装特定版本的SINGA，下面的命令列出了所有可用的SINGA软件包：
+
+```shell
+$ conda search -c nusdbsystem singa
+
+Loading channels: done
+# Name Version Build Channel
+singa 3.1.0.rc2 cpu_py36 nusdbsystem
+singa 3.1.0.rc2 cudnn7.6.5_cuda10.2_py36 nusdbsystem
+singa 3.1.0.rc2 cudnn7.6.5_cuda10.2_py37 nusdbsystem
+```
+
+<!--- > Please note that using the nightly built images is not recommended except for SINGA development and testing. Using stable releases is recommended. -->
+
+下面的命令将安装SINGA的特定版本：
+
+```shell
+$ conda install -c nusdbsystem -c conda-forge singa=X.Y.Z=cpu_py36
+```
+
+如果运行以下命令没有报错：
+```shell
+$ python -c "from singa import tensor"
+```
+
+那么SINGA就安装成功了。
+
+## 使用pip
+
+1. 只使用CPU
+ [](https://colab.research.google.com/drive/17RA056Brwk0vBQTFaZ-l9EbqwADO0NA9?usp=sharing)
+
+```bash
+pip install singa -f http://singa.apache.org/docs/next/wheel-cpu.html --trusted-host singa.apache.org
+```
+
+您可以通过`singa==<version>`安装特定版本的SINGA，其中`<version>`字段应被替换，例如`3.1.0`。可用的SINGA版本在链接中列出。
+
+要安装最新的开发版本，请将链接替换为
+http://singa.apache.org/docs/next/wheel-cpu-dev.html
+
+2. 使用CUDA和cuDNN的GPU
+ [](https://colab.research.google.com/drive/1W30IPCqj5fG8ADAQsFqclaCLyIclVcJL?usp=sharing)
+
+```bash
+pip install singa -f http://singa.apache.org/docs/next/wheel-gpu.html --trusted-host singa.apache.org
+```
+
+您也可以配置SINGA版本和CUDA版本，比如`singa==3.1.0+cuda10.2`，SINGA版本和CUDA版本的可用组合在链接中列出。
+
+要安装最新的开发版本，请将链接替换为
+http://singa.apache.org/docs/next/wheel-gpu-dev.html
+
+注意：你本地Python环境的Python版本将被用来寻找相应的wheel包。例如，如果你本地的Python是3.6，那么就会通过pip选择在Python 3.6上编译的wheel包并安装。事实上，wheel文件的名称包括SINGA版本、CUDA版本和Python版本。因此，`pip`知道要下载和安装哪个wheel文件。
+
+参考setup.py文件顶部的注释，了解如何构建wheel包。
+
+## 使用Docker
+
+按照[说明](https://docs.docker.com/install/)在你的本地主机上安装Docker。将您的用户添加到[docker组](https://docs.docker.com/install/linux/linux-postinstall/)中，以便在没有`sudo`的情况下运行docker命令。
+
+1. 仅使用CPU
+
+```shell
+$ docker run -it apache/singa:X.Y.Z-cpu-ubuntu16.04 /bin/bash
+```
+
+2. 要使用GPU，在安装Docker后安装
+ [Nvidia-Docker](https://github.com/NVIDIA/nvidia-docker)
+
+```shell
+$ nvidia-docker run -it apache/singa:X.Y.Z-cuda9.0-cudnn7.4.2-ubuntu16.04 /bin/bash
+```
+
+3. 关于SINGA Docker镜像（标签）的完整列表，请访问[docker hub site](https://hub.docker.com/r/apache/singa/)。对于每个docker镜像，标签的命名为：
+
+```shell
+version-(cpu|gpu)[-devel]
+```
+
+| Tag | Description | Example value |
+| --------- | -------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `version` | SINGA version | '2.0.0-rc0', '2.0.0', '1.2.0' |
+| `cpu` | the image cannot run on GPUs | 'cpu' |
+| `gpu` | the image can run on Nvidia GPUs | 'gpu', or 'cudax.x-cudnnx.x' e.g., 'cuda10.0-cudnn7.3' |
+| `devel` | indicator for development | if absent, SINGA Python package is installed for runtime only; if present, the building environment is also created, you can recompile SINGA from source at '/root/singa' |
+| `OS` | indicate OS version number | 'ubuntu16.04', 'ubuntu18.04' |
+
+## 从源码编译
+您可以使用本地构建工具或conda-build在本地主机操作系统上或在Docker容器中从源代码[构建和安装SINGA](build.md)。
+
+## FAQ
+
+- Q: `from singa import tensor`错误
+
+ A: 执行下面的命令，检查详细的错误：
+
+ ```shell
+ python -c "from singa import _singa_wrap"
+ # go to the folder of _singa_wrap.so
+ ldd path to _singa_wrap.so
+ python
+ >> import importlib
+ >> importlib.import_module('_singa_wrap')
+ ```
+ `_singa_wrap.so` 的文件夹是 `~/miniconda3/lib/python3.7/site-packages/singa`。通常情况下，这个错误是由于依赖的库不匹配或缺失造成的，例如 cuDNN 或 protobuf。解决方法是创建一个新的虚拟环境，并在该环境中安装SINGA，例如：
+
+
+ ```shell
+ conda create -n singa
+ conda activate singa
+ conda install -c nusdbsystem -c conda-forge singa-cpu
+ ```
+
+- Q: 使用虚拟环境时，每次安装SINGA时，都会重新安装numpy。但是，当我运行`import numpy`时，numpy没有被使用。
+
+ A:
+ 这可能是由`PYTHONPATH`环境变量引起的，在使用虚拟环境时，应将其设置为空，以避免与虚拟环境的路径冲突。
+
+- Q: 当我在Mac OS X中运行SINGA时，得到如下错误 "Fatal Python error:
+ PyThreadState_Get: no current thread Abort trap: 6"
+
+ A: 这个错误通常发生在系统中有多个 Python 版本的时候，例如，操作系统自带的版本和 Homebrew 安装的版本。SINGA链接的Python必须与Python解释器相同。您可以通过`which python`来检查解释器python版本并通过`otool -L <path to _singa_wrap.so>` 检查 SINGA 链接的 Python，如果通过conda安装SINGA，这个问题应该可以解决。
diff --git a/docs-site/docs_chs/issue-tracking.md b/docs-site/docs_chs/issue-tracking.md
new file mode 100644
index 0000000..9cc2afe
--- /dev/null
+++ b/docs-site/docs_chs/issue-tracking.md
@@ -0,0 +1,10 @@
+---
+id: issue-tracking
+title: Issue Tracking
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA使用[JIRA](https://issues.apache.org/jira/browse/singa)来管理问题，包括bug、新功能和讨论。
+
+我们现在正在转移到[Github Issues](https://github.com/apache/singa/issues)。
diff --git a/docs-site/docs_chs/mail-lists.md b/docs-site/docs_chs/mail-lists.md
new file mode 100644
index 0000000..a9d7f4b
--- /dev/null
+++ b/docs-site/docs_chs/mail-lists.md
@@ -0,0 +1,14 @@
+---
+id: mail-lists
+title: Project Mailing Lists
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+这些是为这个项目建立的邮件列表。每个列表都有一个订阅、退订和归档链接：
+
+| Name | Post | Subscribe | Unsubscribe | Archive |
+| ----------- | ------------------------------------ | ---------------------------------------------------------------- | -------------------------------------------------------------------- | ----------------------------------------------------------------------------------- |
+| Development | <dev@singa.incubator.apache.org> | [Subscribe](mailto:dev-subscribe@singa.incubator.apache.org) | [Unsubscribe](mailto:dev-unsubscribe@singa.incubator.apache.org.) | [mail-archives.apache.org](http://mail-archives.apache.org/mod_mbox/singa-dev/) |
+| Commits | <commits@singa.incubator.apache.org> | [Subscribe](mailto:commits-subscribe@singa.incubator.apache.org) | [Unsubscribe](mailto:commits-unsubscribe@singa.incubator.apache.org) | [mail-archives.apache.org](http://mail-archives.apache.org/mod_mbox/singa-commits/) |
+| Security | <security@singa.apache.org> | private | private | private |
diff --git a/docs-site/docs_chs/onnx.md b/docs-site/docs_chs/onnx.md
new file mode 100644
index 0000000..090ba28
--- /dev/null
+++ b/docs-site/docs_chs/onnx.md
@@ -0,0 +1,673 @@
+---
+id: onnx
+title: ONNX
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+[ONNX](https://onnx.ai/) 是机器学习模型的开放表示格式，它使AI开发人员能够在不同的库和工具中使用模型。SINGA支持加载ONNX格式模型用于训练和inference，并将使用SINGA API（如[Module](./module)）定义的模型保存为ONNX格式。
+
+SINGA在以下[版本](https://github.com/onnx/onnx/blob/master/docs/Versioning.md)中的ONNX中测试过。
+
+| ONNX version | File format version | Opset version ai.onnx | Opset version ai.onnx.ml | Opset version ai.onnx.training |
+| ------------ | ------------------- | --------------------- | ------------------------ | ------------------------------ |
+| 1.6.0 | 6 | 11 | 2 | - |
+
+## 通常用法
+
+### 从ONNX中读取一个Model到SINGA
+
+在通过 `onnx.load` 从磁盘加载 ONNX 模型后，您需要更新模型的batch_size，因为对于大多数模型来说，它们使用一个占位符来表示其批处理量。我们在这里举一个例子，若要 `update_batch_size`，你只需要更新输入和输出的 batch_size，内部的 tensors 的形状会自动推断出来。
+
+
+然后，您可以使用 `sonnx.prepare` 来准备 SINGA 模型。该函数将 ONNX 模型图中的所有节点迭代并翻译成 SINGA 运算符，加载所有存储的权重并推断每个中间张量的形状。
+
+```python3
+import onnx
+from singa import device
+from singa import sonnx
+
+# if the input has multiple tensors? can put this function inside prepare()?
+def update_batch_size(onnx_model, batch_size):
+ model_input = onnx_model.graph.input[0]
+ model_input.type.tensor_type.shape.dim[0].dim_value = batch_size
+ model_output = onnx_model.graph.output[0]
+ model_output.type.tensor_type.shape.dim[0].dim_value = batch_size
+ return onnx_model
+
+
+model_path = "PATH/To/ONNX/MODEL"
+onnx_model = onnx.load(model_path)
+
+# set batch size
+onnx_model = update_batch_size(onnx_model, 1)
+
+# convert onnx graph nodes into SINGA operators
+dev = device.create_cuda_gpu()
+sg_ir = sonnx.prepare(onnx_model, device=dev)
+```
+
+### Inference SINGA模型
+
+一旦创建了模型，就可以通过调用`sg_ir.run`进行inference。输入和输出必须是SINGA Tensor实例，由于SINGA模型以列表形式返回输出，如果只有一个输出，你只需要从输出中取第一个元素即可。
+
+```python3
+# can warp the following code in prepare()
+# and provide a flag training=True/False?
+
+class Infer:
+
+
+ def __init__(self, sg_ir):
+ self.sg_ir = sg_ir
+
+ def forward(self, x):
+ return sg_ir.run([x])[0]
+
+
+data = get_dataset()
+x = tensor.Tensor(device=dev, data=data)
+
+model = Infer(sg_ir)
+y = model.forward(x)
+```
+
+### 将SINGA模型保存成ONNX格式
+
+给定输入时序和输出时序，由运算符产生的模型，你可以追溯所有内部操作。因此，一个SINGA模型是由输入和输出张量定义的，要将 SINGA 模型导出为 ONNX 格式，您只需提供输入和输出张量列表。
+
+```python3
+# x is the input tensor, y is the output tensor
+sonnx.to_onnx([x], [y])
+```
+
+### 在ONNX模型上重新训练
+
+要使用 SINGA 训练（或改进）ONNX 模型，您需要设置内部的张量为可训练状态：
+
+```python3
+class Infer:
+
+ def __init__(self, sg_ir):
+ self.sg_ir = sg_ir
+ ## can wrap these codes in sonnx?
+ for idx, tens in sg_ir.tensor_map.items():
+ # allow the tensors to be updated
+ tens.requires_grad = True
+ tens.stores_grad = True
+
+ def forward(self, x):
+ return sg_ir.run([x])[0]
+
+autograd.training = False
+model = Infer(sg_ir)
+
+autograd.training = True
+# then you training the model like normal
+# give more details??
+```
+
+### 在ONNX模型上做迁移学习
+
+您也可以在ONNX模型的最后附加一些图层来进行转移学习。`last_layers` 意味着您从 [0, last_layers] 切断 ONNX 层。然后您可以通过普通的SINGA模型附加更多的层。
+
+```python3
+class Trans:
+
+ def __init__(self, sg_ir, last_layers):
+ self.sg_ir = sg_ir
+ self.last_layers = last_layers
+ self.append_linear1 = autograd.Linear(500, 128, bias=False)
+ self.append_linear2 = autograd.Linear(128, 32, bias=False)
+ self.append_linear3 = autograd.Linear(32, 10, bias=False)
+
+ def forward(self, x):
+ y = sg_ir.run([x], last_layers=self.last_layers)[0]
+ y = self.append_linear1(y)
+ y = autograd.relu(y)
+ y = self.append_linear2(y)
+ y = autograd.relu(y)
+ y = self.append_linear3(y)
+ y = autograd.relu(y)
+ return y
+
+autograd.training = False
+model = Trans(sg_ir, -1)
+
+# then you training the model like normal
+```
+
+## 一个完整示例
+
+本部分以mnist为例，介绍SINGA ONNX的使用方法。在这部分，将展示如何导出、加载、inference、再训练和迁移学习 mnist 模型的例子。您可以在[这里](https://colab.research.google.com/drive/1-YOfQqqw3HNhS8WpB8xjDQYutRdUdmCq)试用这部分内容。
+
+### 读取数据集
+
+首先，你需要导入一些必要的库，并定义一些辅助函数来下载和预处理数据集：
+
+```python
+import os
+import urllib.request
+import gzip
+import numpy as np
+import codecs
+
+from singa import device
+from singa import tensor
+from singa import opt
+from singa import autograd
+from singa import sonnx
+import onnx
+
+
+def load_dataset():
+ train_x_url = 'http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz'
+ train_y_url = 'http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz'
+ valid_x_url = 'http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz'
+ valid_y_url = 'http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz'
+ train_x = read_image_file(check_exist_or_download(train_x_url)).astype(
+ np.float32)
+ train_y = read_label_file(check_exist_or_download(train_y_url)).astype(
+ np.float32)
+ valid_x = read_image_file(check_exist_or_download(valid_x_url)).astype(
+ np.float32)
+ valid_y = read_label_file(check_exist_or_download(valid_y_url)).astype(
+ np.float32)
+ return train_x, train_y, valid_x, valid_y
+
+
+def check_exist_or_download(url):
+
+ download_dir = '/tmp/'
+
+ name = url.rsplit('/', 1)[-1]
+ filename = os.path.join(download_dir, name)
+ if not os.path.isfile(filename):
+ print("Downloading %s" % url)
+ urllib.request.urlretrieve(url, filename)
+ return filename
+
+
+def read_label_file(path):
+ with gzip.open(path, 'rb') as f:
+ data = f.read()
+ assert get_int(data[:4]) == 2049
+ length = get_int(data[4:8])
+ parsed = np.frombuffer(data, dtype=np.uint8, offset=8).reshape(
+ (length))
+ return parsed
+
+
+def get_int(b):
+ return int(codecs.encode(b, 'hex'), 16)
+
+
+def read_image_file(path):
+ with gzip.open(path, 'rb') as f:
+ data = f.read()
+ assert get_int(data[:4]) == 2051
+ length = get_int(data[4:8])
+ num_rows = get_int(data[8:12])
+ num_cols = get_int(data[12:16])
+ parsed = np.frombuffer(data, dtype=np.uint8, offset=16).reshape(
+ (length, 1, num_rows, num_cols))
+ return parsed
+
+
+def to_categorical(y, num_classes):
+ y = np.array(y, dtype="int")
+ n = y.shape[0]
+ categorical = np.zeros((n, num_classes))
+ categorical[np.arange(n), y] = 1
+ categorical = categorical.astype(np.float32)
+ return categorical
+```
+
+### MNIST模型
+
+然后你可以定义一个叫做**CNN**的类来构造mnist模型，这个模型由几个卷积层、池化层、全连接层和relu层组成。你也可以定义一个函数来计算我们结果的**准确性**。最后，你可以定义一个**训练函数**和一个**测试函数**来处理训练和预测的过程。
+
+```python
+class CNN:
+ def __init__(self):
+ self.conv1 = autograd.Conv2d(1, 20, 5, padding=0)
+ self.conv2 = autograd.Conv2d(20, 50, 5, padding=0)
+ self.linear1 = autograd.Linear(4 * 4 * 50, 500, bias=False)
+ self.linear2 = autograd.Linear(500, 10, bias=False)
+ self.pooling1 = autograd.MaxPool2d(2, 2, padding=0)
+ self.pooling2 = autograd.MaxPool2d(2, 2, padding=0)
+
+ def forward(self, x):
+ y = self.conv1(x)
+ y = autograd.relu(y)
+ y = self.pooling1(y)
+ y = self.conv2(y)
+ y = autograd.relu(y)
+ y = self.pooling2(y)
+ y = autograd.flatten(y)
+ y = self.linear1(y)
+ y = autograd.relu(y)
+ y = self.linear2(y)
+ return y
+
+
+def accuracy(pred, target):
+ y = np.argmax(pred, axis=1)
+ t = np.argmax(target, axis=1)
+ a = y == t
+ return np.array(a, "int").sum() / float(len(t))
+
+
+def train(model,
+ x,
+ y,
+ epochs=1,
+ batch_size=64,
+ dev=device.get_default_device()):
+ batch_number = x.shape[0] // batch_size
+
+ for i in range(epochs):
+ for b in range(batch_number):
+ l_idx = b * batch_size
+ r_idx = (b + 1) * batch_size
+
+ x_batch = tensor.Tensor(device=dev, data=x[l_idx:r_idx])
+ target_batch = tensor.Tensor(device=dev, data=y[l_idx:r_idx])
+
+ output_batch = model.forward(x_batch)
+ # onnx_model = sonnx.to_onnx([x_batch], [y])
+ # print('The model is:\n{}'.format(onnx_model))
+
+ loss = autograd.softmax_cross_entropy(output_batch, target_batch)
+ accuracy_rate = accuracy(tensor.to_numpy(output_batch),
+ tensor.to_numpy(target_batch))
+
+ sgd = opt.SGD(lr=0.001)
+ for p, gp in autograd.backward(loss):
+ sgd.update(p, gp)
+ sgd.step()
+
+ if b % 1e2 == 0:
+ print("acc %6.2f loss, %6.2f" %
+ (accuracy_rate, tensor.to_numpy(loss)[0]))
+ print("training completed")
+ return x_batch, output_batch
+
+def test(model, x, y, batch_size=64, dev=device.get_default_device()):
+ batch_number = x.shape[0] // batch_size
+
+ result = 0
+ for b in range(batch_number):
+ l_idx = b * batch_size
+ r_idx = (b + 1) * batch_size
+
+ x_batch = tensor.Tensor(device=dev, data=x[l_idx:r_idx])
+ target_batch = tensor.Tensor(device=dev, data=y[l_idx:r_idx])
+
+ output_batch = model.forward(x_batch)
+ result += accuracy(tensor.to_numpy(output_batch),
+ tensor.to_numpy(target_batch))
+
+ print("testing acc %6.2f" % (result / batch_number))
+```
+
+### 训练mnist模型并将其导出到onnx
+
+现在，你可以通过调用 **soonx.to_onnx** 函数来训练 mnist 模型并导出其 onnx 模型。
+
+```python
+def make_onnx(x, y):
+ return sonnx.to_onnx([x], [y])
+
+# create device
+dev = device.create_cuda_gpu()
+#dev = device.get_default_device()
+# create model
+model = CNN()
+# load data
+train_x, train_y, valid_x, valid_y = load_dataset()
+# normalization
+train_x = train_x / 255
+valid_x = valid_x / 255
+train_y = to_categorical(train_y, 10)
+valid_y = to_categorical(valid_y, 10)
+# do training
+autograd.training = True
+x, y = train(model, train_x, train_y, dev=dev)
+onnx_model = make_onnx(x, y)
+# print('The model is:\n{}'.format(onnx_model))
+
+# Save the ONNX model
+model_path = os.path.join('/', 'tmp', 'mnist.onnx')
+onnx.save(onnx_model, model_path)
+print('The model is saved.')
+```
+
+### Inference
+
+导出onnx模型后，可以在'/tmp'目录下找到一个名为**mnist.onnx**的文件，这个模型可以被其他库导入。现在，如果你想把这个onnx模型再次导入到singa中，并使用验证数据集进行推理，你可以定义一个叫做**Infer**的类，Infer的前向函数将被测试函数调用，对验证数据集进行推理。此外，你应该把训练的标签设置为**False**，以固定自变量算子的梯度。
+
+在导入onnx模型时，需要先调用**onnx.load**来加载onnx模型。然后将onnx模型输入到 **soonx.prepare**中进行解析，并启动到一个singa模型(代码中的**sg_ir**)。sg_ir里面包含了一个singa图，然后就可以通过输入到它的run函数中运行一步推理。
+
+```python
+class Infer:
+ def __init__(self, sg_ir):
+ self.sg_ir = sg_ir
+ for idx, tens in sg_ir.tensor_map.items():
+ # allow the tensors to be updated
+ tens.requires_grad = True
+ tens.stores_grad= True
+ sg_ir.tensor_map[idx] = tens
+
+ def forward(self, x):
+ return sg_ir.run([x])[0] # we can run one step of inference by feeding input
+
+# load the ONNX model
+onnx_model = onnx.load(model_path)
+sg_ir = sonnx.prepare(onnx_model, device=dev) # parse and initiate to a singa model
+
+# inference
+autograd.training = False
+print('The inference result is:')
+test(Infer(sg_ir), valid_x, valid_y, dev=dev)
+```
+
+### 重训练
+
+假设导入模型后，想再次对模型进行重新训练，我们可以定义一个名为**re_train**的函数。在调用这个re_train函数之前，我们应该将训练的标签设置为**True**，以使自变量运算符更新其梯度。而在完成训练后，我们再将其设置为**False**，以调用做推理的测试函数。
+
+```python
+def re_train(sg_ir,
+ x,
+ y,
+ epochs=1,
+ batch_size=64,
+ dev=device.get_default_device()):
+ batch_number = x.shape[0] // batch_size
+
+ new_model = Infer(sg_ir)
+
+ for i in range(epochs):
+ for b in range(batch_number):
+ l_idx = b * batch_size
+ r_idx = (b + 1) * batch_size
+
+ x_batch = tensor.Tensor(device=dev, data=x[l_idx:r_idx])
+ target_batch = tensor.Tensor(device=dev, data=y[l_idx:r_idx])
+
+ output_batch = new_model.forward(x_batch)
+
+ loss = autograd.softmax_cross_entropy(output_batch, target_batch)
+ accuracy_rate = accuracy(tensor.to_numpy(output_batch),
+ tensor.to_numpy(target_batch))
+
+ sgd = opt.SGD(lr=0.01)
+ for p, gp in autograd.backward(loss):
+ sgd.update(p, gp)
+ sgd.step()
+
+ if b % 1e2 == 0:
+ print("acc %6.2f loss, %6.2f" %
+ (accuracy_rate, tensor.to_numpy(loss)[0]))
+ print("re-training completed")
+ return new_model
+
+# load the ONNX model
+onnx_model = onnx.load(model_path)
+sg_ir = sonnx.prepare(onnx_model, device=dev)
+
+# re-training
+autograd.training = True
+new_model = re_train(sg_ir, train_x, train_y, dev=dev)
+autograd.training = False
+test(new_model, valid_x, valid_y, dev=dev)
+```
+
+### 迁移学习
+
+最后，如果我们想做迁移学习，我们可以定义一个名为**Trans**的函数，在onnx模型后追加一些层。为了演示，代码中只在onnx模型后追加了几个线性（全连接）和relu。可以定义一个transfer_learning函数来处理transfer-learning模型的训练过程，而训练的标签和前面一个的一样。
+
+```python
+class Trans:
+ def __init__(self, sg_ir, last_layers):
+ self.sg_ir = sg_ir
+ self.last_layers = last_layers
+ self.append_linear1 = autograd.Linear(500, 128, bias=False)
+ self.append_linear2 = autograd.Linear(128, 32, bias=False)
+ self.append_linear3 = autograd.Linear(32, 10, bias=False)
+
+ def forward(self, x):
+ y = sg_ir.run([x], last_layers=self.last_layers)[0]
+ y = self.append_linear1(y)
+ y = autograd.relu(y)
+ y = self.append_linear2(y)
+ y = autograd.relu(y)
+ y = self.append_linear3(y)
+ y = autograd.relu(y)
+ return y
+
+def transfer_learning(sg_ir,
+ x,
+ y,
+ epochs=1,
+ batch_size=64,
+ dev=device.get_default_device()):
+ batch_number = x.shape[0] // batch_size
+
+ trans_model = Trans(sg_ir, -1)
+
+ for i in range(epochs):
+ for b in range(batch_number):
+ l_idx = b * batch_size
+ r_idx = (b + 1) * batch_size
+
+ x_batch = tensor.Tensor(device=dev, data=x[l_idx:r_idx])
+ target_batch = tensor.Tensor(device=dev, data=y[l_idx:r_idx])
+ output_batch = trans_model.forward(x_batch)
+
+ loss = autograd.softmax_cross_entropy(output_batch, target_batch)
+ accuracy_rate = accuracy(tensor.to_numpy(output_batch),
+ tensor.to_numpy(target_batch))
+
+ sgd = opt.SGD(lr=0.07)
+ for p, gp in autograd.backward(loss):
+ sgd.update(p, gp)
+ sgd.step()
+
+ if b % 1e2 == 0:
+ print("acc %6.2f loss, %6.2f" %
+ (accuracy_rate, tensor.to_numpy(loss)[0]))
+ print("transfer-learning completed")
+ return trans_mode
+
+# load the ONNX model
+onnx_model = onnx.load(model_path)
+sg_ir = sonnx.prepare(onnx_model, device=dev)
+
+# transfer-learning
+autograd.training = True
+new_model = transfer_learning(sg_ir, train_x, train_y, dev=dev)
+autograd.training = False
+test(new_model, valid_x, valid_y, dev=dev)
+```
+
+## ONNX模型库
+
+[ONNX 模型库](https://github.com/onnx/models)是由社区成员贡献的 ONNX 格式的预先训练的最先进模型的集合。SINGA 现在已经支持了几个 CV 和 NLP 模型。将来会支持更多模型。
+
+### 图像分类
+
+这套模型以图像作为输入，然后将图像中的主要物体分为1000个物体类别，如键盘、鼠标、铅笔和许多动物。
+
+| Model Class | Reference | Description | Link |
+| --------------------------------------------------------------------------------------------------- | ------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <b>[MobileNet](https://github.com/onnx/models/tree/master/vision/classification/mobilenet)</b> | [Sandler et al.](https://arxiv.org/abs/1801.04381) | 最适合移动和嵌入式视觉应用的轻量级深度神经网络。 <br>Top-5 error from paper - ~10% | [](https://colab.research.google.com/drive/1HsixqJMIpKyEPhkbB8jy7NwNEFEAUWAf) |
+| <b>[ResNet18](https://github.com/onnx/models/tree/master/vision/classification/resnet)</b> | [He et al.](https://arxiv.org/abs/1512.03385) | 一个CNN模型（多达152层），在对图像进行分类时，使用shortcut来实现更高的准确性。 <br> Top-5 error from paper - ~3.6% | [](https://colab.research.google.com/drive/1u1RYefSsVbiP4I-5wiBKHjsT9L0FxLm9) |
+| <b>[VGG16](https://github.com/onnx/models/tree/master/vision/classification/vgg)</b> | [Simonyan et al.](https://arxiv.org/abs/1409.1556) | 深度CNN模型（多达19层）。类似于AlexNet，但使用多个较小的内核大小的滤波器，在分类图像时提供更高的准确性。 <br>Top-5 error from paper - ~8% | [](https://colab.research.google.com/drive/14kxgRKtbjPCKKsDJVNi3AvTev81Gp_Ds) |
+| <b>[ShuffleNet_V2](https://github.com/onnx/models/tree/master/vision/classification/shufflenet)</b> | [Simonyan et al.](https://arxiv.org/pdf/1707.01083.pdf) | 专门为移动设备设计的计算效率极高的CNN模型。这种网络架构设计考虑了速度等直接指标，而不是FLOP等间接指标。 Top-1 error from paper - ~30.6% | [](https://colab.research.google.com/drive/19HfRu3YHP_H2z3BcZujVFRp23_J5XsuA?usp=sharing) |
+
+### 目标检测
+
+目标检测模型可以检测图像中是否存在多个对象，并将图像中检测到对象的区域分割出来。
+
+| Model Class | Reference | Description | Link |
+| ----------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <b>[Tiny YOLOv2](https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/tiny_yolov2)</b> | [Redmon et al.](https://arxiv.org/pdf/1612.08242.pdf) | 一个用于目标检测的实时CNN，可以检测20个不同的类。一个更复杂的完整YOLOv2网络的小版本。 | [](https://colab.research.google.com/drive/11V4I6cRjIJNUv5ZGsEGwqHuoQEie6b1T) |
+
+### 面部识别
+
+人脸检测模型可以识别和/或识别给定图像中的人脸和情绪。
+
+| Model Class | Reference | Description | Link |
+| --------------------------------------------------------------------------------------------------------- | -------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <b>[ArcFace](https://github.com/onnx/models/tree/master/vision/body_analysis/arcface)</b> | [Deng et al.](https://arxiv.org/abs/1801.07698) | 一种基于CNN的人脸识别模型，它可以学习人脸的判别特征，并对输入的人脸图像进行分析。 | [](https://colab.research.google.com/drive/1qanaqUKGIDtifdzEzJOHjEj4kYzA9uJC) |
+| <b>[Emotion FerPlus](https://github.com/onnx/models/tree/master/vision/body_analysis/emotion_ferplus)</b> | [Barsoum et al.](https://arxiv.org/abs/1608.01041) | 基于人脸图像训练的情感识别深度CNN。 | [](https://colab.research.google.com/drive/1XHtBQGRhe58PDi4LGYJzYueWBeWbO23r) |
+
+### 机器理解
+
+这个自然语言处理模型的子集，可以回答关于给定上下文段落的问题。
+
+| Model Class | Reference | Description | Link |
+| ----------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <b>[BERT-Squad](https://github.com/onnx/models/tree/master/text/machine_comprehension/bert-squad)</b> | [Devlin et al.](https://arxiv.org/pdf/1810.04805.pdf) | 该模型根据给定输入段落的上下文回答问题。 | [](https://colab.research.google.com/drive/1kud-lUPjS_u-TkDAzihBTw0Vqr0FjCE-) |
+| <b>[RoBERTa](https://github.com/onnx/models/tree/master/text/machine_comprehension/roberta)</b> | [Devlin et al.](https://arxiv.org/pdf/1907.11692.pdf) | 一个基于大型变换器的模型，根据给定的输入文本预测情感。 | [](https://colab.research.google.com/drive/1F-c4LJSx3Cb2jW6tP7f8nAZDigyLH6iN?usp=sharing) |
+| <b>[GPT-2](https://github.com/onnx/models/tree/master/text/machine_comprehension/gpt-2)</b> | [Devlin et al.](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) | 一个基于大型变换器的语言模型，给定一些文本中的单词序列，预测下一个单词。 | [](https://colab.research.google.com/drive/1ZlXLSIMppPch6HgzKRillJiUcWn3PiK7?usp=sharing) |
+
+## 支持的操作符
+
+onnx支持下列运算:
+
+- Acos
+- Acosh
+- Add
+- And
+- Asin
+- Asinh
+- Atan
+- Atanh
+- AveragePool
+- BatchNormalization
+- Cast
+- Ceil
+- Clip
+- Concat
+- ConstantOfShape
+- Conv
+- Cos
+- Cosh
+- Div
+- Dropout
+- Elu
+- Equal
+- Erf
+- Expand
+- Flatten
+- Gather
+- Gemm
+- GlobalAveragePool
+- Greater
+- HardSigmoid
+- Identity
+- LeakyRelu
+- Less
+- Log
+- MatMul
+- Max
+- MaxPool
+- Mean
+- Min
+- Mul
+- Neg
+- NonZero
+- Not
+- OneHot
+- Or
+- Pad
+- Pow
+- PRelu
+- Reciprocal
+- ReduceMean
+- ReduceSum
+- Relu
+- Reshape
+- ScatterElements
+- Selu
+- Shape
+- Sigmoid
+- Sign
+- Sin
+- Sinh
+- Slice
+- Softmax
+- Softplus
+- Softsign
+- Split
+- Sqrt
+- Squeeze
+- Sub
+- Sum
+- Tan
+- Tanh
+- Tile
+- Transpose
+- Unsqueeze
+- Upsample
+- Where
+- Xor
+
+### 对ONNX后端的特别说明
+
+- Conv, MaxPool 以及 AveragePool
+
+ 输入必须是1d`(N*C*H)`和2d`(N*C*H*W)`的形状，`dilation`必须是1。
+
+- BatchNormalization
+
+ `epsilon` 设定为1e-05，不能改变
+
+- Cast
+
+ 只支持float32和int32，其他类型都会转向这两种类型。
+
+- Squeeze and Unsqueeze
+
+ 如果你在`Tensor`和Scalar之间`Squeeze`或`Unsqueeze`时遇到错误，请向我们报告。
+
+- Empty tensor
+
+ 空张量在SINGA是非法的。
+
+## 实现
+
+SINGA ONNX的代码在`python/singa/soonx.py`中，主要有三个类，`SingaFrontend`、`SingaBackend`和`SingaRep`。`SingaFrontend`将SINGA模型翻译成ONNX模型；`SingaBackend`将ONNX模型翻译成`SingaRep`对象，其中存储了所有的SINGA运算符和张量（本文档中的张量指SINGA Tensor）；`SingaRep`可以像SINGA模型一样运行。
+
+### SingaFrontend
+
+`SingaFrontend`的入口函数是`singa_to_onnx_model`，它也被称为`to_onnx`，`singa_to_onnx_model`创建了ONNX模型，它还通过`singa_to_onnx_graph`创建了一个ONNX图。
+
+
+`singa_to_onnx_graph`接受模型的输出，并从输出中递归迭代SINGA模型的图，得到所有的运算符，形成一个队列。SINGA模型的输入和中间张量，即可训练的权重，同时被获取。输入存储在`onnx_model.graph.input`中；输出存储在`onnx_model.graph.output`中；可训练权重存储在`onnx_model.graph.initializer`中。
+
+然后将队列中的SINGA运算符逐一翻译成ONNX运算符。`_rename_operators` 定义了 SINGA 和 ONNX 之间的运算符名称映射。`_special_operators` 定义了翻译运算符时要使用的函数。
+
+此外，SINGA 中的某些运算符与 ONNX 的定义不同，即 ONNX 将 SINGA 运算符的某些属性视为输入，因此 `_unhandled_operators` 定义了处理特殊运算符的函数。
+
+由于SINGA中的布尔类型被视为int32，所以`_bool_operators`定义了要改变的操作符为布尔类型。
+
+### SingaBackend
+
+`SingaBackend`的入口函数是`prepare`，它检查ONNX模型的版本，然后调用`_onnx_model_to_singa_net`。
+
+`_onnx_model_to_singa_net`的目的是获取SINGA的时序和运算符。tensors在ONNX中以其名称存储在字典中，而操作符则以`namedtuple('SingaOps', ['name', 'op', 'handle', 'forward'])`的形式存储在队列中。对于每个运算符，`name`是它的ONNX节点名称；`op`是ONNX节点；`forward`是SINGA运算符的转发函数；`handle`是为一些特殊的运算符准备的，如Conv和Pooling，它们有`handle`对象。
+
+`_onnx_model_to_singa_net`的第一步是调用`_init_graph_parameter`来获取模型内的所有tensors。对于可训练的权重，可以从`onnx_model.graph.initializer`中初始化`SINGA Tensor`。请注意，权重也可能存储在图的输入或称为`Constant`的ONNX节点中，SINGA也可以处理这些。
+
+虽然所有的权重都存储在ONNX模型中，但模型的输入是未知的，只有它的形状和类型。所以SINGA支持两种方式来初始化输入，1、根据其形状和类型生成随机张量，2、允许用户分配输入。第一种方法对大多数模型都很好，但是对于一些模型，比如BERT，矩阵的指数不能随机生成，否则会产生错误。
+
+然后，`_onnx_model_to_singa_net`迭代ONNX图中的所有节点，将其翻译成SIGNA运算符。另外，`_rename_operators` 定义了 SINGA 和 ONNX 之间的运算符名称映射。`_special_operators` 定义翻译运算符时要使用的函数。`_run_node`通过输入时序来运行生成的 SINGA 模型，并存储其输出时序，供以后的运算符使用。
+
+该类最后返回一个`SingaRep`对象，并在其中存储所有SINGA时序和运算符。
+
+### SingaRep
+
+`SingaBackend`存储所有的SINGA tensors和运算符。`run`接受模型的输入，并按照运算符队列逐个运行SINGA运算符。用户可以使用`last_layers`来决定是否将模型运行到最后几层。将 `all_outputs` 设置为 `False` 表示只得到最后的输出，设置为 `True` 表示也得到所有的中间输出。
diff --git a/docs-site/docs_chs/optimizer.md b/docs-site/docs_chs/optimizer.md
new file mode 100644
index 0000000..3c2da70
--- /dev/null
+++ b/docs-site/docs_chs/optimizer.md
@@ -0,0 +1,122 @@
+---
+id: optimizer
+title: Optimizer
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA支持各种流行的优化器，包括动量随机梯度下降、Adam、RMSProp和AdaGrad等。对于每一种优化器，它都支持使用衰减调度器来安排不同时间段的学习率。优化器和衰减调度器包含在`singa/opt.py`中。
+
+## 创建一个优化器
+
+1. 带动量的SGD
+
+```python
+# define hyperparameter learning rate
+lr = 0.001
+# define hyperparameter momentum
+momentum = 0.9
+# define hyperparameter weight decay
+weight_decay = 0.0001
+
+from singa import opt
+sgd = opt.SGD(lr=lr, momentum=momentum, weight_decay=weight_decay)
+```
+
+2. RMSProp
+
+```python
+# define hyperparameter learning rate
+lr = 0.001
+# define hyperparameter rho
+rho = 0.9
+# define hyperparameter epsilon
+epsilon = 1e-8
+# define hyperparameter weight decay
+weight_decay = 0.0001
+
+from singa import opt
+sgd = opt.RMSProp(lr=lr, rho=rho, epsilon=epsilon, weight_decay=weight_decay)
+```
+
+3. AdaGrad
+
+```python
+# define hyperparameter learning rate
+lr = 0.001
+# define hyperparameter epsilon
+epsilon = 1e-8
+# define hyperparameter weight decay
+weight_decay = 0.0001
+
+from singa import opt
+sgd = opt.AdaGrad(lr=lr, epsilon=epsilon, weight_decay=weight_decay)
+```
+
+4. Adam
+
+```python
+# define hyperparameter learning rate
+lr = 0.001
+# define hyperparameter beta 1
+beta_1= 0.9
+# define hyperparameter beta 2
+beta_1= 0.999
+# define hyperparameter epsilon
+epsilon = 1e-8
+# define hyperparameter weight decay
+weight_decay = 0.0001
+
+from singa import opt
+sgd = opt.Adam(lr=lr, beta_1=beta_1, beta_2=beta_2, epsilon=epsilon, weight_decay=weight_decay)
+```
+
+## 创建一个衰减调度器
+
+```python
+from singa import opt
+
+# define initial learning rate
+lr_init = 0.001
+# define the rate of decay in the decay scheduler
+decay_rate = 0.95
+# define whether the learning rate schedule is a staircase shape
+staircase=True
+# define the decay step of the decay scheduler (in this example the lr is decreased at every 2 steps)
+decay_steps = 2
+
+# create the decay scheduler, the schedule of lr becomes lr_init * (decay_rate ^ (step // decay_steps) )
+lr = opt.ExponentialDecay(0.1, 2, 0.5, True)
+# Use the lr to create an optimizer
+sgd = opt.SGD(lr=lr, momentum=0.9, weight_decay=0.0001)
+```
+
+## 使用模型API中的优化器
+
+当我们创建模型时，我们需要将优化器附加到模型上：
+
+```python
+# create a CNN using the Model API
+model = CNN()
+
+# initialize optimizer and attach it to the model
+sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
+model.set_optimizer(sgd)
+```
+
+然后，当我们调用模型时，它会运行利用优化器的 `train_one_batch` 方法。
+
+因此，一个迭代循环优化模型的例子是：
+
+```python
+for b in range(num_train_batch):
+ # generate the next mini-batch
+ x, y = ...
+
+ # Copy the data into input tensors
+ tx.copy_from_numpy(x)
+ ty.copy_from_numpy(y)
+
+ # Training with one batch
+ out, loss = model(tx, ty)
+```
diff --git a/docs-site/docs_chs/releases/RELEASE_NOTES_0.1.0.md b/docs-site/docs_chs/releases/RELEASE_NOTES_0.1.0.md
new file mode 100644
index 0000000..9d1dc09
--- /dev/null
+++ b/docs-site/docs_chs/releases/RELEASE_NOTES_0.1.0.md
@@ -0,0 +1,152 @@
+---
+id: RELEASE_NOTES_0.1.0
+title: singa-incubating-0.1.0 Release Notes
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA is a general distributed deep learning platform for training big deep
+learning models over large datasets. It is designed with an intuitive
+programming model based on the layer abstraction. SINGA supports a wide variety
+of popular deep learning models.
+
+This release includes following features:
+
+- Job management
+ - [SINGA-3](https://issues.apache.org/jira/browse/SINGA-3) Use Zookeeper to
+ check stopping (finish) time of the system
+ - [SINGA-16](https://issues.apache.org/jira/browse/SINGA-16) Runtime Process
+ id Management
+ - [SINGA-25](https://issues.apache.org/jira/browse/SINGA-25) Setup glog output
+ path
+ - [SINGA-26](https://issues.apache.org/jira/browse/SINGA-26) Run distributed
+ training in a single command
+ - [SINGA-30](https://issues.apache.org/jira/browse/SINGA-30) Enhance
+ easy-to-use feature and support concurrent jobs
+ - [SINGA-33](https://issues.apache.org/jira/browse/SINGA-33) Automatically
+ launch a number of processes in the cluster
+ - [SINGA-34](https://issues.apache.org/jira/browse/SINGA-34) Support external
+ zookeeper service
+ - [SINGA-38](https://issues.apache.org/jira/browse/SINGA-38) Support
+ concurrent jobs
+ - [SINGA-39](https://issues.apache.org/jira/browse/SINGA-39) Avoid ssh in
+ scripts for single node environment
+ - [SINGA-43](https://issues.apache.org/jira/browse/SINGA-43) Remove
+ Job-related output from workspace
+ - [SINGA-56](https://issues.apache.org/jira/browse/SINGA-56) No automatic
+ launching of zookeeper service
+ - [SINGA-73](https://issues.apache.org/jira/browse/SINGA-73) Refine the
+ selection of available hosts from host list
+
+* Installation with GNU Auto tool
+ - [SINGA-4](https://issues.apache.org/jira/browse/SINGA-4) Refine
+ thirdparty-dependency installation
+ - [SINGA-13](https://issues.apache.org/jira/browse/SINGA-13) Separate
+ intermediate files of compilation from source files
+ - [SINGA-17](https://issues.apache.org/jira/browse/SINGA-17) Add root
+ permission within thirdparty/install.
+ - [SINGA-27](https://issues.apache.org/jira/browse/SINGA-27) Generate python
+ modules for proto objects
+ - [SINGA-53](https://issues.apache.org/jira/browse/SINGA-53) Add lmdb
+ compiling options
+ - [SINGA-62](https://issues.apache.org/jira/browse/SINGA-62) Remove building
+ scrips and auxiliary files
+ - [SINGA-67](https://issues.apache.org/jira/browse/SINGA-67) Add singatest
+ into build targets
+
+- Distributed training
+ - [SINGA-7](https://issues.apache.org/jira/browse/SINGA-7) Implement shared
+ memory Hogwild algorithm
+ - [SINGA-8](https://issues.apache.org/jira/browse/SINGA-8) Implement
+ distributed Hogwild
+ - [SINGA-19](https://issues.apache.org/jira/browse/SINGA-19) Slice large Param
+ objects for load-balance
+ - [SINGA-29](https://issues.apache.org/jira/browse/SINGA-29) Update NeuralNet
+ class to enable layer partition type customization
+ - [SINGA-24](https://issues.apache.org/jira/browse/SINGA-24) Implement
+ Downpour training framework
+ - [SINGA-32](https://issues.apache.org/jira/browse/SINGA-32) Implement
+ AllReduce training framework
+ - [SINGA-57](https://issues.apache.org/jira/browse/SINGA-57) Improve
+ Distributed Hogwild
+
+* Training algorithms for different model categories
+ - [SINGA-9](https://issues.apache.org/jira/browse/SINGA-9) Add Support for
+ Restricted Boltzman Machine (RBM) model
+ - [SINGA-10](https://issues.apache.org/jira/browse/SINGA-10) Add Support for
+ Recurrent Neural Networks (RNN)
+
+- Checkpoint and restore
+ - [SINGA-12](https://issues.apache.org/jira/browse/SINGA-12) Support
+ Checkpoint and Restore
+
+* Unit test
+ - [SINGA-64](https://issues.apache.org/jira/browse/SINGA-64) Add the test
+ module for utils/common
+
+- Programming model
+ - [SINGA-36](https://issues.apache.org/jira/browse/SINGA-36) Refactor job
+ configuration, driver program and scripts
+ - [SINGA-37](https://issues.apache.org/jira/browse/SINGA-37) Enable users to
+ set parameter sharing in model configuration
+ - [SINGA-54](https://issues.apache.org/jira/browse/SINGA-54) Refactor job
+ configuration to move fields in ModelProto out
+ - [SINGA-55](https://issues.apache.org/jira/browse/SINGA-55) Refactor main.cc
+ and singa.h
+ - [SINGA-61](https://issues.apache.org/jira/browse/SINGA-61) Support user
+ defined classes
+ - [SINGA-65](https://issues.apache.org/jira/browse/SINGA-65) Add an example of
+ writing user-defined layers
+
+* Other features
+ - [SINGA-6](https://issues.apache.org/jira/browse/SINGA-6) Implement
+ thread-safe singleton
+ - [SINGA-18](https://issues.apache.org/jira/browse/SINGA-18) Update API for
+ displaying performance metric
+ - [SINGA-77](https://issues.apache.org/jira/browse/SINGA-77) Integrate with
+ Apache RAT
+
+Some bugs are fixed during the development of this release
+
+- [SINGA-2](https://issues.apache.org/jira/browse/SINGA-2) Check failed:
+ zsock_connect
+- [SINGA-5](https://issues.apache.org/jira/browse/SINGA-5) Server early
+ terminate when zookeeper singa folder is not initially empty
+- [SINGA-15](https://issues.apache.org/jira/browse/SINGA-15) Fixg a bug from
+ ConnectStub function which gets stuck for connecting layer*dealer*
+- [SINGA-22](https://issues.apache.org/jira/browse/SINGA-22) Cannot find
+ openblas library when it is installed in default path
+- [SINGA-23](https://issues.apache.org/jira/browse/SINGA-23) Libtool version
+ mismatch error.
+- [SINGA-28](https://issues.apache.org/jira/browse/SINGA-28) Fix a bug from
+ topology sort of Graph
+- [SINGA-42](https://issues.apache.org/jira/browse/SINGA-42) Issue when loading
+ checkpoints
+- [SINGA-44](https://issues.apache.org/jira/browse/SINGA-44) A bug when reseting
+ metric values
+- [SINGA-46](https://issues.apache.org/jira/browse/SINGA-46) Fix a bug in
+ updater.cc to scale the gradients
+- [SINGA-47](https://issues.apache.org/jira/browse/SINGA-47) Fix a bug in data
+ layers that leads to out-of-memory when group size is too large
+- [SINGA-48](https://issues.apache.org/jira/browse/SINGA-48) Fix a bug in
+ trainer.cc that assigns the same NeuralNet instance to workers from diff
+ groups
+- [SINGA-49](https://issues.apache.org/jira/browse/SINGA-49) Fix a bug in
+ HandlePutMsg func that sets param fields to invalid values
+- [SINGA-66](https://issues.apache.org/jira/browse/SINGA-66) Fix bugs in
+ Worker::RunOneBatch function and ClusterProto
+- [SINGA-79](https://issues.apache.org/jira/browse/SINGA-79) Fix bug in
+ singatool that can not parse -conf flag
+
+Features planned for the next release
+
+- [SINGA-11](https://issues.apache.org/jira/browse/SINGA-11) Start SINGA using
+ Mesos
+- [SINGA-31](https://issues.apache.org/jira/browse/SINGA-31) Extend Blob to
+ support xpu (cpu or gpu)
+- [SINGA-35](https://issues.apache.org/jira/browse/SINGA-35) Add random number
+ generators
+- [SINGA-40](https://issues.apache.org/jira/browse/SINGA-40) Support sparse
+ Param update
+- [SINGA-41](https://issues.apache.org/jira/browse/SINGA-41) Support single node
+ single GPU training
diff --git a/docs-site/docs_chs/releases/RELEASE_NOTES_0.2.0.md b/docs-site/docs_chs/releases/RELEASE_NOTES_0.2.0.md
new file mode 100644
index 0000000..77633f3
--- /dev/null
+++ b/docs-site/docs_chs/releases/RELEASE_NOTES_0.2.0.md
@@ -0,0 +1,81 @@
+---
+id: RELEASE_NOTES_0.2.0
+title: singa-incubating-0.2.0 Release Notes
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA is a general distributed deep learning platform for training big deep
+learning models over large datasets. It is designed with an intuitive
+programming model based on the layer abstraction. SINGA supports a wide variety
+of popular deep learning models.
+
+This release includes the following **major features**:
+
+- [Training on GPU](../docs/gpu.html) enables training of complex models on a
+ single node with multiple GPU cards.
+- [Hybrid neural net partitioning](../docs/hybrid.html) supports data and model
+ parallelism at the same time.
+- [Python wrapper](../docs/python.html) makes it easy to configure the job,
+ including neural net and SGD algorithm.
+- [RNN model and BPTT algorithm](../docs/general-rnn.html) are implemented to
+ support applications based on RNN models, e.g., GRU.
+- [Cloud software integration](../docs/distributed-training.md) includes Mesos,
+ Docker and HDFS.
+
+**More details** are listed as follows,
+
+- Programming model
+ - [SINGA-80] New Blob Level and Address Level Math Operation Interface
+ - [SINGA-82] Refactor input layers using data store abstraction
+ - [SINGA-87] Replace exclude field to include field for layer configuration
+ - [SINGA-110] Add Layer member datavec* and gradvec*
+ - [SINGA-120] Implemented GRU and BPTT (BPTTWorker)
+
+* Neuralnet layers
+ - [SINGA-91] Add SoftmaxLayer and ArgSortLayer
+ - [SINGA-106] Add dummy layer for test purpose
+ - [SINGA-120] Implemented GRU and BPTT (GRULayer and OneHotLayer)
+
+- GPU training support
+ - [SINGA-100] Implement layers using CUDNN for GPU training
+ - [SINGA-104] Add Context Class
+ - [SINGA-105] Update GUN make files for compiling cuda related code
+ - [SINGA-98] Add Support for AlexNet ImageNet Classification Model
+
+* Model/Hybrid partition
+ - [SINGA-109] Refine bridge layers
+ - [SINGA-111] Add slice, concate and split layers
+ - [SINGA-113] Model/Hybrid Partition Support
+
+- Python binding
+ - [SINGA-108] Add Python wrapper to singa
+
+* Predict-only mode
+ - [SINGA-85] Add functions for extracting features and test new data
+
+- Integrate with third-party tools
+ - [SINGA-11] Start SINGA on Apache Mesos
+ - [SINGA-78] Use Doxygen to generate documentation
+ - [SINGA-89] Add Docker support
+
+* Unit test
+ - [SINGA-95] Add make test after building
+
+- Other improvment
+ - [SINGA-84] Header Files Rearrange
+ - [SINGA-93] Remove the asterisk in the log tcp://169.254.12.152:\*:49152
+ - [SINGA-94] Move call to google::InitGoogleLogging() from Driver::Init() to
+ main()
+ - [SINGA-96] Add Momentum to Cifar10 Example
+ - [SINGA-101] Add ll (ls -l) command in .bashrc file when using docker
+ - [SINGA-114] Remove short logs in tmp directory
+ - [SINGA-115] Print layer debug information in the neural net graph file
+ - [SINGA-118] Make protobuf LayerType field id easy to assign
+ - [SIGNA-97] Add HDFS Store
+
+* Bugs fixed
+ - [SINGA-85] Fix compilation errors in examples
+ - [SINGA-90] Miscellaneous trivial bug fixes
+ - [SINGA-107] Error from loading pre-trained params for training stacked RBMs
+ - [SINGA-116] Fix a bug in InnerProductLayer caused by weight matrix sharing
diff --git a/docs-site/docs_chs/releases/RELEASE_NOTES_0.3.0.md b/docs-site/docs_chs/releases/RELEASE_NOTES_0.3.0.md
new file mode 100644
index 0000000..ed02198
--- /dev/null
+++ b/docs-site/docs_chs/releases/RELEASE_NOTES_0.3.0.md
@@ -0,0 +1,42 @@
+---
+id: RELEASE_NOTES_0.3.0
+title: singa-incubating-0.3.0 Release Notes
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA is a general distributed deep learning platform for training big deep
+learning models over large datasets. It is designed with an intuitive
+programming model based on the layer abstraction. SINGA supports a wide variety
+of popular deep learning models.
+
+This release includes following features:
+
+- GPU Support
+
+ - [SINGA-131] Implement and optimize hybrid training using both CPU and GPU
+ - [SINGA-136] Support cuDNN v4
+ - [SINGA-134] Extend SINGA to run over a GPU cluster
+ - [SINGA-157] Change the priority of cudnn library and install libsingagpu.so
+
+- Remove Dependences
+
+ - [SINGA-156] Remove the dependency on ZMQ for single process training
+ - [SINGA-155] Remove zookeeper for single-process training
+
+- Python Binding
+
+ - [SINGA-126] Python Binding for Interactive Training
+
+- Other Improvements
+
+ - [SINGA-80] New Blob Level and Address Level Math Operation Interface
+ - [SINGA-130] Data Prefetching
+ - [SINGA-145] New SGD based optimization Updaters: AdaDelta, Adam, AdamMax
+
+- Bugs Fixed
+ - [SINGA-148] Race condition between Worker threads and Driver
+ - [SINGA-150] Mesos Docker container failed
+ - [SIGNA-141] Undesired Hash collision when locating process id to worker…
+ - [SINGA-149] Docker build fail
+ - [SINGA-143] The compilation cannot detect libsingagpu.so file
diff --git a/docs-site/docs_chs/releases/RELEASE_NOTES_1.0.0.md b/docs-site/docs_chs/releases/RELEASE_NOTES_1.0.0.md
new file mode 100644
index 0000000..0577ede
--- /dev/null
+++ b/docs-site/docs_chs/releases/RELEASE_NOTES_1.0.0.md
@@ -0,0 +1,95 @@
+---
+id: RELEASE_NOTES_1.0.0
+title: singa-incubating-1.0.0 Release Notes
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA is a general distributed deep learning platform for training big deep
+learning models over large datasets. It is designed with an intuitive
+programming model based on the layer abstraction. SINGA supports a wide variety
+of popular deep learning models.
+
+This release includes following features:
+
+- Core abstractions including Tensor and Device
+ - [SINGA-207] Update Tensor functions for matrices
+ - [SINGA-205] Enable slice and concatenate operations for Tensor objects
+ - [SINGA-197] Add CNMem as a submodule in lib/
+ - [SINGA-196] Rename class Blob to Block
+ - [SINGA-194] Add a Platform singleton
+ - [SINGA-175] Add memory management APIs and implement a subclass using CNMeM
+ - [SINGA-173] OpenCL Implementation
+ - [SINGA-171] Create CppDevice and CudaDevice
+ - [SINGA-168] Implement Cpp Math functions APIs
+ - [SINGA-162] Overview of features for V1.x
+ - [SINGA-165] Add cross-platform timer API to singa
+ - [SINGA-167] Add Tensor Math function APIs
+ - [SINGA-166] light built-in logging for making glog optional
+ - [SINGA-164] Add the base Tensor class
+
+* IO components for file read/write, network and data pre-processing
+ - [SINGA-233] New communication interface
+ - [SINGA-215] Implement Image Transformation for Image Pre-processing
+ - [SINGA-214] Add LMDBReader and LMDBWriter for LMDB
+ - [SINGA-213] Implement Encoder and Decoder for CSV
+ - [SINGA-211] Add TextFileReader and TextFileWriter for CSV files
+ - [SINGA-210] Enable checkpoint and resume for v1.0
+ - [SINGA-208] Add DataIter base class and a simple implementation
+ - [SINGA-203] Add OpenCV detection for cmake compilation
+ - [SINGA-202] Add reader and writer for binary file
+ - [SINGA-200] Implement Encoder and Decoder for data pre-processing
+
+- Module components including layer classes, training algorithms and Python
+ binding
+ - [SINGA-235] Unify the engines for cudnn and singa layers
+ - [SINGA-230] OpenCL Convolution layer and Pooling layer
+ - [SINGA-222] Fixed bugs in IO
+ - [SINGA-218] Implementation for RNN CUDNN version
+ - [SINGA-204] Support the training of feed-forward neural nets
+ - [SINGA-199] Implement Python classes for SGD optimizers
+ - [SINGA-198] Change Layer::Setup API to include input Tensor shapes
+ - [SINGA-193] Add Python layers
+ - [SINGA-192] Implement optimization algorithms for SINGA v1 (nesterove,
+ adagrad, rmsprop)
+ - [SINGA-191] Add "autotune" for CudnnConvolution Layer
+ - [SINGA-190] Add prelu layer and flatten layer
+ - [SINGA-189] Generate python outputs of proto files
+ - [SINGA-188] Add Dense layer
+ - [SINGA-187] Add popular parameter initialization methods
+ - [SINGA-186] Create Python Tensor class
+ - [SINGA-184] Add Cross Entropy loss computation
+ - [SINGA-183] Add the base classes for optimizer, constraint and regularizer
+ - [SINGA-180] Add Activation layer and Softmax layer
+ - [SINGA-178] Add Convolution layer and Pooling layer
+ - [SINGA-176] Add loss and metric base classes
+ - [SINGA-174] Add Batch Normalization layer and Local Response Nomalization
+ layer.
+ - [SINGA-170] Add Dropout layer and CudnnDropout layer.
+ - [SINGA-169] Add base Layer class for V1.0
+
+* Examples
+
+ - [SINGA-232] Alexnet on Imagenet
+ - [SINGA-231] Batchnormlized VGG model for cifar-10
+ - [SINGA-228] Add Cpp Version of Convolution and Pooling layer
+ - [SINGA-227] Add Split and Merge Layer and add ResNet Implementation
+
+* Documentation
+
+ - [SINGA-239] Transfer documentation files of v0.3.0 to github
+ - [SINGA-238] RBM on mnist
+ - [SINGA-225] Documentation for installation and Cifar10 example
+ - [SINGA-223] Use Sphinx to create the website
+
+* Tools for compilation and some utility code
+ - [SINGA-229] Complete install targets
+ - [SINGA-221] Support for Travis-CI
+ - [SINGA-217] build python package with setup.py
+ - [SINGA-216] add jenkins for CI support
+ - [SINGA-212] Disable the compilation of libcnmem if USE_CUDA is OFF
+ - [SINGA-195] Channel for sending training statistics
+ - [SINGA-185] Add CBLAS and GLOG detection for singav1
+ - [SINGA-181] Add NVCC supporting for .cu files
+ - [SINGA-177] Add fully cmake supporting for the compilation of singa_v1
+ - [SINGA-172] Add CMake supporting for Cuda and Cudnn libs
diff --git a/docs-site/docs_chs/releases/RELEASE_NOTES_1.1.0.md b/docs-site/docs_chs/releases/RELEASE_NOTES_1.1.0.md
new file mode 100644
index 0000000..db79546
--- /dev/null
+++ b/docs-site/docs_chs/releases/RELEASE_NOTES_1.1.0.md
@@ -0,0 +1,56 @@
+---
+id: RELEASE_NOTES_1.1.0
+title: singa-incubating-1.1.0 Release Notes
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA is a general distributed deep learning platform for training big deep
+learning models over large datasets.
+
+This release includes following features:
+
+- Core components
+
+ - [SINGA-296] Add sign and to_host function for pysinga tensor module
+
+- Model components
+
+ - [SINGA-254] Implement Adam for V1
+ - [SINGA-264] Extend the FeedForwardNet to accept multiple inputs
+ - [SINGA-267] Add spatial mode in batch normalization layer
+ - [SINGA-271] Add Concat and Slice layers
+ - [SINGA-275] Add Cross Entropy Loss for multiple labels
+ - [SINGA-278] Convert trained caffe parameters to singa
+ - [SINGA-287] Add memory size check for cudnn convolution
+
+- Utility functions and CI
+
+ - [SINGA-242] Compile all source files into a single library.
+ - [SINGA-244] Separating swig interface and python binding files
+ - [SINGA-246] Imgtool for image augmentation
+ - [SINGA-247] Add windows support for singa
+ - [SINGA-251] Implement image loader for pysinga
+ - [SINGA-252] Use the snapshot methods to dump and load models for pysinga
+ - [SINGA-255] Compile mandatory dependent libaries together with SINGA code
+ - [SINGA-259] Add maven pom file for building java classes
+ - [SINGA-261] Add version ID into the checkpoint files
+ - [SINGA-266] Add Rafiki python toolkits
+ - [SINGA-273] Improve license and contributions
+ - [SINGA-284] Add python unittest into Jenkins and link static libs into whl
+ file
+ - [SINGA-280] Jenkins CI support
+ - [SINGA-288] Publish wheel of PySINGA generated by Jenkins to public servers
+
+- Documentation and usability
+
+ - [SINGA-263] Create Amazon Machine Image
+ - [SINGA-268] Add IPython notebooks to the documentation
+ - [SINGA-276] Create docker images
+ - [SINGA-289] Update SINGA website automatically using Jenkins
+ - [SINGA-295] Add an example of image classification using GoogleNet
+
+- Bugs fixed
+ - [SINGA-245] float as the first operand can not multiply with a tensor object
+ - [SINGA-293] Bug from compiling PySINGA on Mac OS X with multiple version of
+ Python
diff --git a/docs-site/docs_chs/releases/RELEASE_NOTES_1.2.0.md b/docs-site/docs_chs/releases/RELEASE_NOTES_1.2.0.md
new file mode 100644
index 0000000..aade8bd
--- /dev/null
+++ b/docs-site/docs_chs/releases/RELEASE_NOTES_1.2.0.md
@@ -0,0 +1,62 @@
+---
+id: RELEASE_NOTES_1.2.0
+title: singa-incubating-1.2.0 Release Notes
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA is a general distributed deep learning platform for training big deep
+learning models over large datasets.
+
+This release includes following features:
+
+- Core components
+
+ - [SINGA-290] Upgrade to Python 3
+ - [SINGA-341] Added stride functionality to tensors for CPP
+ - [SINGA-347] Create a function that supports einsum
+ - [SINGA-351] Added stride support and cudnn codes to cuda
+
+- Model components
+
+ - [SINGA-300] Add residual networks for imagenet classification
+ - [SINGA-312] Rename layer parameters
+ - [SINGA-313] Add L2 norm layer
+ - [SINGA-315] Reduce memory footprint by Python generator for parameter
+ - [SINGA-316] Add SigmoidCrossEntropy
+ - [SINGA-324] Extend RNN layer to accept variant seq length across batches
+ - [SINGA-326] Add Inception V4 for ImageNet classification
+ - [SINGA-328] Add VGG models for ImageNet classification
+ - [SINGA-329] Support layer freezing during training (fine-tuning)
+ - [SINGA-346] Update cudnn from V5 to V7
+ - [SINGA-349] Create layer operations for autograd
+ - [SINGA-363] Add DenseNet for Imagenet classification
+
+- Utility functions and CI
+
+ - [SINGA-274] Improve Debian packaging with CPack
+ - [SINGA-303] Create conda packages
+ - [SINGA-337] Add test cases for code
+ - [SINGA-348] Support autograd MLP Example
+ - [SINGA-345] Update Jenkins and fix bugs in compliation
+ - [SINGA-354] Update travis scripts to use conda-build for all platforms
+ - [SINGA-358] Consolidated RUN steps and cleaned caches in Docker containers
+ - [SINGA-359] Create alias for conda packages
+
+- Documentation and usability
+
+ - [SINGA-223] Fix side navigation menu in the website
+ - [SINGA-294] Add instructions to run CUDA unit tests on Windows
+ - [SINGA-305] Add jupyter notebooks for SINGA V1 tutorial
+ - [SINGA-319] Fix link errors on the index page
+ - [SINGA-352] Complete SINGA documentation in Chinese version
+ - [SINGA-361] Add git instructions for contributors and committers
+
+- Bugs fixed
+ - [SINGA-330] fix openblas building on i7 7700k
+ - [SINGA-331] Fix the bug of tensor division operation
+ - [SINGA-350] Error from python3 test
+ - [SINGA-356] Error using travis tool to build SINGA on mac os
+ - [SINGA-363] Fix some bugs in imagenet examples
+ - [SINGA-368] Fix the bug in Cifar10 examples
+ - [SINGA-369] the errors of examples in testing
diff --git a/docs-site/docs_chs/releases/RELEASE_NOTES_2.0.0.md b/docs-site/docs_chs/releases/RELEASE_NOTES_2.0.0.md
new file mode 100644
index 0000000..f574d69
--- /dev/null
+++ b/docs-site/docs_chs/releases/RELEASE_NOTES_2.0.0.md
@@ -0,0 +1,55 @@
+---
+id: RELEASE_NOTES_2.0.0
+title: singa-incubating-2.0.0 Release Notes
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA is a general distributed deep learning platform for training big deep
+learning models over large datasets.
+
+This release includes following features:
+
+- Core components
+
+ - [SINGA-434] Support tensor broadcasting
+ - [SINGA-370] Improvement to tensor reshape and various misc. changes related
+ to SINGA-341 and 351
+
+- Model components
+
+ - [SINGA-333] Add support for Open Neural Network Exchange (ONNX) format
+ - [SINGA-385] Add new python module for optimizers
+ - [SINGA-394] Improve the CPP operations via Intel MKL DNN lib
+ - [SINGA-425] Add 3 operators , Abs(), Exp() and leakyrelu(), for Autograd
+ - [SINGA-410] Add two function, set_params() and get_params(), for Autograd
+ Layer class
+ - [SINGA-383] Add Separable Convolution for autograd
+ - [SINGA-388] Develop some RNN layers by calling tiny operations like matmul,
+ addbias.
+ - [SINGA-382] Implement concat operation for autograd
+ - [SINGA-378] Implement maxpooling operation and its related functions for
+ autograd
+ - [SINGA-379] Implement batchnorm operation and its related functions for
+ autograd
+
+- Utility functions and CI
+
+ - [SINGA-432] Update depdent lib versions in conda-build config
+ - [SINGA-429] Update docker images for latest cuda and cudnn
+ - [SINGA-428] Move Docker images under Apache user name
+
+- Documentation and usability
+ - [SINGA-395] Add documentation for autograd APIs
+ - [SINGA-344] Add a GAN example
+ - [SINGA-390] Update installation.md
+ - [SINGA-384] Implement ResNet using autograd API
+ - [SINGA-352] Complete SINGA documentation in Chinese version
+
+* Bugs fixed
+ - [SINGA-431] Unit Test failed - Tensor Transpose
+ - [SINGA-422] ModuleNotFoundError: No module named "\_singa_wrap"
+ - [SINGA-418] Unsupportive type 'long' in python3.
+ - [SINGA-409] Basic `singa-cpu` import throws error
+ - [SINGA-408] Unsupportive function definition in python3
+ - [SINGA-380] Fix bugs from Reshape
diff --git a/docs-site/docs_chs/releases/RELEASE_NOTES_3.0.0.md b/docs-site/docs_chs/releases/RELEASE_NOTES_3.0.0.md
new file mode 100644
index 0000000..70a0ea2
--- /dev/null
+++ b/docs-site/docs_chs/releases/RELEASE_NOTES_3.0.0.md
@@ -0,0 +1,107 @@
+---
+id: RELEASE_NOTES_3.0.0
+title: singa-3.0.0 Release Notes
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+Release Notes - SINGA - Version singa-3.0.0.rc1
+
+SINGA is a distributed deep learning library.
+
+This release includes following changes:
+
+- Code quality has been promoted by introducing linting check in CI and auto
+ code formatter. For linting, the tools, `cpplint` and `pylint`, are used and
+ configured to comply
+ [google coding styles](http://google.github.io/styleguide/) details in
+ `tool/linting/`. Similarly, formatting tools, `clang-format` and `yapf`
+ configured with google coding styles, are the recommended one for developers
+ to clean code before submitting changes, details in `tool/code-format/`.
+ [LGTM](https://lgtm.com) is enabled on Github for code quality check; License
+ check is also enabled.
+
+- New Tensor APIs are added for naming consistency, and feature enhancement:
+
+ - size(), mem_size(), get_value(), to_proto(), l1(), l2(): added for the sake
+ of naming consistency
+ - AsType(): convert data type between `float` and `int`
+ - ceil(): perform element-wise ceiling of the input
+ - concat(): concatenate two tensor
+ - index selector: e.g. tensor1[:,:,1:,1:]
+ - softmax(in, axis): allow to perform softmax on a axis on a multi-dimensional
+ tensor
+
+- 14 new operators are added into the autograd module: Gemm, GlobalAveragePool,
+ ConstantOfShape, Dropout, ReduceSum, ReduceMean, Slice, Ceil, Split, Gather,
+ Tile, NonZero, Cast, OneHot. Their unit tests are added as well.
+
+- 14 new operators are added to sonnx module for both backend and frontend:
+ [Gemm](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Gemm),
+ [GlobalAveragePool](https://github.com/onnx/onnx/blob/master/docs/Operators.md#GlobalAveragePool),
+ [ConstantOfShape](https://github.com/onnx/onnx/blob/master/docs/Operators.md#ConstantOfShape),
+ [Dropout](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Dropout),
+ [ReduceSum](https://github.com/onnx/onnx/blob/master/docs/Operators.md#ReduceSum),
+ [ReduceMean](https://github.com/onnx/onnx/blob/master/docs/Operators.md#ReduceMean),
+ [Slice](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Slice),
+ [Ceil](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Ceil),
+ [Split](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Split),
+ [Gather](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Gather),
+ [Tile](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Tile),
+ [NonZero](https://github.com/onnx/onnx/blob/master/docs/Operators.md#NonZero),
+ [Cast](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Cast),
+ [OneHot](https://github.com/onnx/onnx/blob/master/docs/Operators.md#OneHot).
+ Their tests are added as well.
+
+- Some ONNX models are imported into SINGA, including
+ [Bert-squad](https://github.com/onnx/models/tree/master/text/machine_comprehension/bert-squad),
+ [Arcface](https://github.com/onnx/models/tree/master/vision/body_analysis/arcface),
+ [FER+ Emotion](https://github.com/onnx/models/tree/master/vision/body_analysis/emotion_ferplus),
+ [MobileNet](https://github.com/onnx/models/tree/master/vision/classification/mobilenet),
+ [ResNet18](https://github.com/onnx/models/tree/master/vision/classification/resnet),
+ [Tiny Yolov2](https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/tiny_yolov2),
+ [Vgg16](https://github.com/onnx/models/tree/master/vision/classification/vgg),
+ and Mnist.
+
+- Some operators now support
+ [multidirectional broadcasting](https://github.com/onnx/onnx/blob/master/docs/Broadcasting.md#multidirectional-broadcasting),
+ including Add, Sub, Mul, Div, Pow, PRelu, Gemm
+
+- [Distributed training with communication optimization].
+ [DistOpt](./python/singa/opt.py) has implemented multiple optimization
+ techniques, including gradient sparsification, chunk transmission, and
+ gradient compression.
+
+- Computational graph construction at the CPP level. The operations submitted to
+ the Device are buffered. After analyzing the dependency, the computational
+ graph is created, which is further analyzed for speed and memory optimization.
+ To enable this feature, use the [Module API](./python/singa/module.py).
+
+- New website based on Docusaurus. The documentation files are moved to a
+ separate repo [singa-doc](https://github.com/apache/singa-doc). The static
+ website files are stored at
+ [singa-site](https://github.com/apache/singa-site).
+
+- DNNL([Deep Neural Network Library](https://github.com/intel/mkl-dnn)), powered
+ by Intel, is integrated into
+ `model/operations/[batchnorm|pooling|convolution]`, the changes is opaque to
+ the end users. The current version is dnnl v1.1 which replaced previous
+ integration of mkl-dnn v0.18. The framework could boost the performance of dl
+ operations when executing on CPU. The dnnl dependency is installed through
+ conda.
+
+- Some Tensor APIs are marked as deprecated which could be replaced by
+ broadcast, and it can support better on multi-dimensional operations. These
+ APIs are add_column(), add_row(), div_column(), div_row(), mult_column(),
+ mult_row()
+
+- Conv and Pooling are enhanced to support fine-grained padding like (2,3,2,3),
+ and
+ [SAME_UPPER, SAME_LOWER](https://github.com/onnx/onnx/blob/master/docs/Operators.md#Conv)
+ pad mode and shape checking.
+
+- Reconstruct soonx,
+ - Support two types of weight value (Initializer and Constant Node);
+ - For some operators (BatchNorm, Reshape, Clip, Slice, Gather, Tile, OneHot),
+ move some inputs to its attributes;
+ - Define and implement the type conversion map.
diff --git a/docs-site/docs_chs/security.md b/docs-site/docs_chs/security.md
new file mode 100644
index 0000000..fed861e
--- /dev/null
+++ b/docs-site/docs_chs/security.md
@@ -0,0 +1,8 @@
+---
+id: security
+title: Security
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+用户可以向[SINGA安全团队邮件列表](mailto:security@singa.apache.org)报告安全漏洞。
diff --git a/docs-site/docs_chs/software-stack.md b/docs-site/docs_chs/software-stack.md
new file mode 100644
index 0000000..11f979b
--- /dev/null
+++ b/docs-site/docs_chs/software-stack.md
@@ -0,0 +1,84 @@
+---
+id: software-stack
+title: Software Stack
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA的软件栈包括两个主要层次，底层后端和Python接口层。图1展示了它们与硬件的关系。后端组件为深度学习模型提供基本的数据结构，为调度和执行操作提供硬件抽象，为分布式训练提供通信组件。Python接口封装了部分CPP数据结构，并为神经网络训练提供了额外的高级类，方便实现复杂的神经网络模型。
+
+接下来，我们以自下而上的方式介绍软件栈。
+
+ <br/> **Figure 1 - SINGA V3
+软件栈结构**
+
+## 底层后端
+
+### Device
+
+每个`Device`实例，即一个设备，是针对一个硬件设备创建的，例如GPU或CPU。`Device`管理数据结构的内存，并调度操作的执行，例如，在CUDA流或CPU线程上。根据硬件及其编程语言，SINGA实现了以下特定的设备类。
+
+- **CudaGPU** 代表Nvidia GPU卡,执行单元是CUDA流。
+- **CppCPU** 代表一个普通的CPU,执行单位是CPU线程。
+- **OpenclGPU** 代表Nvidia和AMD的普通GPU卡,执行单位是CommandQueue,鉴于OpenCL与许多硬件设备兼容，如FPGA和ARM，OpenclGPU可以扩展到其他设备。
+
+### Tensor
+
+`Tensor`类表示一个多维数组，它存储模型变量，例如，来自卷积层的输入图像和特征图。每个`Tensor`实例（即张量）都分配在一个设备上，它管理张量的内存，并针对张量调度（计算）操作。大多数机器学习算法可以使用（密集或稀疏）张量抽象及其操作来表达。因此，SINGA将能够运行广泛的模型，包括深度学习模型和其他传统的机器学习模型。
+
+### Operator
+
+针对张量的运算符有两类，一类是线性代数运算符，如矩阵乘法，另一类是神经网络特有的运算符，如卷积和池化。线性代数运算符以`Tensor`函数的形式提供，并针对不同的硬件设备分别实现。
+
+- CppMath (tensor_math_cpp.h) 为CppCPU使用Cpp实现张量运算。
+- CudaMath (tensor_math_cuda.h) 为CudaGPU使用CUDA实现张量运算。
+- OpenclMath(tensor_math_opencl.h) 使用OpenCL实现了OpenclGPU的张量运算。
+
+神经网络的具体运算符也是单独实现的，例如：
+
+- GpuConvFoward (convolution.h) 在Nvidia GPU上通过CuDNN实现卷积的前向函数。
+- CpuConvForward (convolution.h) 在CPU上使用CPP实现卷积的前向函数。
+
+通常情况下，用户会创建一个`Device`实例，并使用它来创建多个`Tensor`实例。当用户调用Tensor函数或神经网络操作时，会调用常驻设备的相应实现。换句话说，操作符的实现对用户是透明的。
+
+Tensor和Device抽象是可扩展的，以支持使用不同编程语言的各种硬件设备。一个新的硬件设备将通过添加一个新的Device子类和相应的操作符的实现来支持。
+
+在速度和内存方面的优化是由`Device`的`Scheduler`和`MemPool`完成的。例如，`Scheduler`根据运算符的依赖性创建一个[计算图](./graph)，然后它可以优化运算符的执行顺序，以实现并行性和内存共享。
+
+### Communicator
+
+`Communicator`是为了支持[分布式训练](./dist-train)。它使用套接字、MPI和NCCL实现通信协议。通常情况下，用户只需要调用`put()`和`get()`等高级API来发送和接收tensors。针对拓扑结构、消息大小等的通信优化在内部完成。
+
+## Python 接口层
+
+所有的后端组件都通过SWIG作为Python模块暴露出来。此外，还添加了以下类来支持复杂神经网络的实现：
+
+### Opt
+
+`Opt`及其子类实现了使用参数梯度更新模型参数值的方法（如SGD）。子类[DistOpt](./dist-train)通过调用`Communicator`中的方法来同步分布式训练的worker之间的梯度。
+
+### Operator
+
+`Operator`包装了多个使用后端Tensor或神经网络运算符实现的函数。例如，前向函数和后向函数`ReLU`组成`ReLU`运算符。
+
+### Layer
+
+`Layer`和它的子类用参数来包装运算符。例如，卷积和线性运算符有权重和偏置参数。 这些参数由相应的`Layer`类维护。
+
+### Autograd
+
+
+`Autograd`通过记录运算符的前向函数的执行情况，自动按照相反的顺序调用后向函数，实现了[反向模式的自动微分](https://rufflewind.com/2016-12-30/reverse-mode-automatic-differentiation)。所有的函数都可以通过`Scheduler`进行缓冲，建立[计算图](./graph)，以提高效率和内存优化。
+
+### Model
+
+
+[Model](./graph)提供了一个简单的接口来实现新的网络模型。你只需要继承`Model`，通过创建和调用层或运算符来定义模型的前向传播。当训练数据输入到`Model`中时，`Model`会自动进行autograd并通过`Opt`自动更新参数。通过`Model` API，SINGA充分利用了命令式编程和声明式编程的优势。用户使用[Model](./graph) API按照PyTorch那样的命令式编程风格来实现一个网络，PyTorch在每次迭代中都会重新创建操作，而SINGA则不同，它在第一次迭代后就会缓冲操作，隐式地创建一个计算图（当该功能被启用时）。该图类似于使用声明式编程的库所创建的图，例如TensorFlow。因此，SINGA可以在计算图上应用内存和速度优化技术。
+
+### ONNX
+
+
+为了支持ONNX，SINGA实现了一个[sonnx](./onnx)模块，其中包括。
+
+- SingaFrontend用于将SINGA模型保存为onnx格式。
+- SingaBackend 用于将 onnx 格式的模型加载到 SINGA 中进行训练和inference。
\ No newline at end of file
diff --git a/docs-site/docs_chs/source-repository.md b/docs-site/docs_chs/source-repository.md
new file mode 100644
index 0000000..154b2f6
--- /dev/null
+++ b/docs-site/docs_chs/source-repository.md
@@ -0,0 +1,23 @@
+---
+id: source-repository
+title: Source Repository
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+
+本项目使用[Git](http://git-scm.com/)来管理其源代码。关于Git的使用说明可以在 http://git-scm.com/documentation 找到。
+
+## Repository
+
+以下是在线资源库的链接：
+
+- https://gitbox.apache.org/repos/asf?p=singa.git
+
+Github镜像在：
+
+- https://github.com/apache/singa
+
+代码可以从任何一个repo中克隆，例如：
+
+ git clone https://github.com/apache/singa.git
diff --git a/docs-site/docs_chs/team-list.md b/docs-site/docs_chs/team-list.md
new file mode 100644
index 0000000..8bf4740
--- /dev/null
+++ b/docs-site/docs_chs/team-list.md
@@ -0,0 +1,56 @@
+---
+id: team-list
+title: The SINGA Team
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+一个成功的项目需要许多人扮演许多角色，有些成员编写代码或文档，而其他成员则作为测试人员，提交补丁和建议。
+
+SINGA社区的开发者主要来自新加坡国立大学、浙江大学、网易、大阪大学、yzBigData等。
+
+## PMC
+
+| Name | Email | Organization |
+| ------------- | ----------------------- | --------------------------------------------- |
+| Gang Chen | cg@apache.org | Zhejiang University |
+| Anh Dinh | dinhtta@apache.org | Singapore University of Technology and Design |
+| Ted Dunning | tdunning@apache.org | Apache Software Foundation |
+| Jinyang Gao | jinyang@apache.org | DAMO Academy, Alibaba Group |
+| Alan Gates | gates@apache.org | Apache Software Foundation |
+| Zhaojing Luo | zhaojing@apache.org | National University of Singapore |
+| Thejas Nair | thejas@apache.org | Apache Software Foundation |
+| Beng Chin Ooi | ooibc@apache.org | National University of Singapore |
+| Moaz Reyad | moaz@apache.org | Université Grenoble Alpes |
+| Kian-Lee Tan | tankianlee@apache.org | National University of Singapore |
+| Sheng Wang | wangsh@apache.org | DAMO Academy, Alibaba Group |
+| Wei Wang | wangwei@apache.org | National University of Singapore |
+| Zhongle Xie | zhongle@apache.org | National University of Singapore |
+| Sai Ho Yeung | chrishkchris@apache.org | National University of Singapore |
+| Meihui Zhang | meihuizhang@apache.org | Beijing Institute of Technology |
+| Kaiping Zheng | kaiping@apache.org | National University of Singapore |
+
+## Committers
+
+| Name | Email | Organization |
+| ------------ | ---------------------- | --------------------------------------------- |
+| Xiangrui Cai | caixr@apache.org | Nankai University |
+| Chonho Lee | chonho@apache.org | Osaka University |
+| Shicong Lin | shicong@apache.org | National University of Singapore |
+| Rulin Xing | rulin@apache.org | Huazhong University of Science and Technology |
+| Wanqi Xue | xuewanqi@apache.org | Nanyang Technological University |
+| Joddiy Zhang | joddiyzhang@apache.org | National University of Singapore |
+
+## Contributors
+
+| Name | Email | Organization |
+| ------------------ | ---------------------------- | -------------------------------- |
+| Haibo Chen | hzchenhaibo@corp.netease.com | NetEase |
+| Shicheng Chen | chengsc@comp.nus.edu.sg | National University of Singapore |
+| Xin Ji | vincent.j.xin@gmail.com | Visenze, Singapore |
+| Anthony K. H. Tung | atung@comp.nus.edu.sg | National University of Singapore |
+| Ji Wang | wangji@mzhtechnologies.com | Hangzhou MZH Technologies |
+| Yuan Wang | wangyuan@corp.netease.com | NetEase |
+| Wenfeng Wu | dcswuw@gmail.com | Freelancer, China |
+| Kaiyuan Yang | yangky@comp.nus.edu.sg | National University of Singapore |
+| Chang Yao | yaochang2009@gmail.com | Hangzhou MZH Technologies |
diff --git a/docs-site/docs_chs/tensor.md b/docs-site/docs_chs/tensor.md
new file mode 100644
index 0000000..7e05af7
--- /dev/null
+++ b/docs-site/docs_chs/tensor.md
@@ -0,0 +1,244 @@
+---
+id: tensor
+title: Tensor
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+每个Tensor实例都是分配在特定设备实例上的多维数组。Tensor实例可以存储变量，并在不同类型的硬件设备上提供线性代数运算，而无需用户察觉。需要注意的是，除了复制函数外，用户需要确保张量操作数分配在同一个设备上。
+
+## Tensor用法
+
+### 创建Tensor
+
+```python
+>>> import numpy as np
+>>> from singa import tensor
+>>> tensor.from_numpy( np.asarray([[1, 0, 0], [0, 1, 0]], dtype=np.float32) )
+[[1. 0. 0.]
+ [0. 1. 0.]]
+```
+
+### 转换到numpy
+
+```python
+>>> a = np.asarray([[1, 0, 0], [0, 1, 0]], dtype=np.float32)
+>>> tensor.from_numpy(a)
+[[1. 0. 0.]
+ [0. 1. 0.]]
+>>> tensor.to_numpy(tensor.from_numpy(a))
+array([[1., 0., 0.],
+ [0., 1., 0.]], dtype=float32)
+```
+
+### Tensor方法
+
+```python
+>>> t = tensor.from_numpy(a)
+>>> t.transpose([1,0])
+[[1. 0.]
+ [0. 1.]
+ [0. 0.]]
+```
+
+`Tensor`变换，最多支持6维。
+
+```python
+>>> a = tensor.random((2,3,4,5,6,7))
+>>> a.shape
+(2, 3, 4, 5, 6, 7)
+>>> a.reshape((2,3,4,5,7,6)).transpose((3,2,1,0,4,5)).shape
+(5, 4, 3, 2, 7, 6)
+```
+
+### Tensor算术方法
+
+`tensor`是实时计算的：
+
+```python
+>>> t + 1
+[[2. 1. 1.]
+ [1. 2. 1.]]
+>>> t / 5
+[[0.2 0. 0. ]
+ [0. 0.2 0. ]]
+```
+
+`tensor` broadcasting运算:
+
+```python
+>>> a
+[[1. 2. 3.]
+ [4. 5. 6.]]
+>>> b
+[[1. 2. 3.]]
+>>> a + b
+[[2. 4. 6.]
+ [5. 7. 9.]]
+>>> a * b
+[[ 1. 4. 9.]
+ [ 4. 10. 18.]]
+>>> a / b
+[[1. 1. 1. ]
+ [4. 2.5 2. ]]
+>>> a/=b # inplace operation
+>>> a
+[[1. 1. 1. ]
+ [4. 2.5 2. ]]
+```
+
+`tensor` broadcasting矩阵乘法：
+
+```python
+>>> from singa import tensor
+>>> a = tensor.random((2,2,2,3))
+>>> b = tensor.random((2,3,4))
+>>> tensor.mult(a,b).shape
+(2, 2, 2, 4)
+```
+
+### Tensor函数
+
+`singa.tensor`模块中的函数在应用函数中定义的变换后返回新的`Tensor`对象。
+
+```python
+>>> tensor.log(t+1)
+[[0.6931472 0. 0. ]
+ [0. 0.6931472 0. ]]
+```
+
+### Tensor在不同Devices上
+
+`tensor`默认在主机(CPU)上创建；也可以通过指定设备在不同的硬件`device`上创建。`tensor`可以通过`to_device()`函数在`device`之间移动。
+
+```python
+>>> from singa import device
+>>> x = tensor.Tensor((2, 3), device.create_cuda_gpu())
+>>> x.gaussian(1,1)
+>>> x
+[[1.531889 1.0128608 0.12691343]
+ [2.1674204 3.083676 2.7421203 ]]
+>>> # move to host
+>>> x.to_device(device.get_default_device())
+```
+
+### 使用Tensor训练MLP
+
+```python
+
+"""
+ code snipet from examples/mlp/module.py
+"""
+
+label = get_label()
+data = get_data()
+
+dev = device.create_cuda_gpu_on(0)
+sgd = opt.SGD(0.05)
+
+# define tensor for input data and label
+tx = tensor.Tensor((400, 2), dev, tensor.float32)
+ty = tensor.Tensor((400,), dev, tensor.int32)
+model = MLP(data_size=2, perceptron_size=3, num_classes=2)
+
+# attached model to graph
+model.set_optimizer(sgd)
+model.compile([tx], is_train=True, use_graph=True, sequential=False)
+model.train()
+
+for i in range(1001):
+ tx.copy_from_numpy(data)
+ ty.copy_from_numpy(label)
+ out, loss = model(tx, ty, 'fp32', spars=None)
+
+ if i % 100 == 0:
+ print("training loss = ", tensor.to_numpy(loss)[0])
+```
+
+输出:
+
+```bash
+$ python3 examples/mlp/module.py
+training loss = 0.6158037
+training loss = 0.52852553
+training loss = 0.4571422
+training loss = 0.37274635
+training loss = 0.30146334
+training loss = 0.24906921
+training loss = 0.21128304
+training loss = 0.18390492
+training loss = 0.16362564
+training loss = 0.148164
+training loss = 0.13589878
+```
+
+## Tensor实现
+
+上一节介绍了`Tensor`的一般用法，下面将介绍其底层的实现。首先，将介绍Python和C++ tensors的设计。后面会讲到前端（Python）和后端（C++）如何连接，如何扩展。
+
+### Python Tensor
+
+Python类`Tensor`，定义在`python/singa/tensor.py`中，提供了高层的张量操作，用于实现深度学习操作（通过[autograd](./autograd)），以及终端用户的数据管理。
+
+它主要是通过简单地封装C++张量方法来工作，包括算术方法（如`sum`）和非算术方法（如`reshape`）。一些高级的算术运算以后会引入，并使用纯Python的张量API来实现，如`tensordot`。Python Tensor API可以利用灵活的方法轻松实现复杂的神经网络操作。
+
+### C++ Tensor
+
+C++类`Tensor`，定义在`include/singa/core/tensor.h`中，主要是管理存放数据的内存，并提供低级的API用于张量操作。同时，它还通过封装不同的后端（CUDA、BLAS、cuBLAS等）提供各种算术方法（如`matmul`）。
+
+#### 执行背景和内存块
+
+Tensor的两个重要概念或者说数据结构是执行背景`device`，和内存块`Block`。
+
+每个`Tensor`物理上存储在一个硬件设备上，并由硬件设备管理，代表执行背景（CPU、GPU），Tensor的数学计算是在设备上执行的。
+
+Tensor数据在`Block`实例中，定义在`include/singa/core/common.h`中。`Block`拥有底层数据，而tensors则在描述tensor的元数据上拥有所有权，比如`shape`、`stride`。
+
+#### Tensor数学后端
+
+为了利用不同后端硬件设备提供的高效数学库，SINGA为每个支持的后端设备提供了一套Tensor函数的实现。
+
+- 'tensor_math_cpp.h'为CppCPU设备使用Cpp（带CBLAS）实现操作。
+- 'tensor_math_cuda.h'为CudaGPU设备实现了使用Cuda(带cuBLAS)的操作。
+- 'tensor_math_opencl.h'为OpenclGPU设备实现了使用OpenCL的操作。
+
+### 将C++ APIs暴露给Python
+
+
+SWIG(http://www.swig.org/)是一个可以自动将C++ API转换为Python API的工具。SINGA使用SWIG将C++ APIs公开到Python中。SWIG会生成几个文件，包括`python/singa/singa_wrap.py`。Python模块(如`tensor`、`device`和`autograd`)导入这个模块来调用C++ API来实现Python类和函数。
+
+```python
+import tensor
+
+t = tensor.Tensor(shape=(2, 3))
+```
+
+例如，当按上面的方法创建Python `Tensor`实例时，`Tensor`类的实现会创建一个在`singa_wrap.py`中定义的Tensor类的实例，它对应于C++ `Tensor`类。为了清楚起见，`singa_wrap.py`中的`Tensor`类在`tensor.py`中被称为`CTensor`。
+
+```python
+# in tensor.py
+from . import singa_wrap as singa
+
+CTensor = singa.Tensor
+```
+
+### 创建新的Tensor函数
+
+
+有了前面的描述所奠定的基础，扩展张量函数可以用自下而上的方式轻松完成。对于数学运算，其步骤是：
+
+- 在`tensor.h`中声明新的API。
+- 使用 `tensor.cc` 中预定义的宏生成代码，参考 `GenUnaryTensorFn(Abs);` 。
+- 在`tensor_math.h`中声明template 方法/函数。
+- 至少在CPU(`tensor_math_cpp.h`)和GPU(`tensor_math_cuda.h`)中进行真正的实现。
+- 将 API 加入到 `src/api/core_tensor.i` 中，通过 SWIG 公开 API。
+- 通过调用 `singa_wrap.py` 中自动生成的函数，在 `tensor.py` 中定义 Python Tensor API。
+- 在适当的地方编写单元测试
+
+## Python API
+
+_进行中_
+
+## CPP API
+
+_进行中_
diff --git a/docs-site/docs_chs/time-profiling.md b/docs-site/docs_chs/time-profiling.md
new file mode 100644
index 0000000..a3feca7
--- /dev/null
+++ b/docs-site/docs_chs/time-profiling.md
@@ -0,0 +1,153 @@
+---
+id: time-profiling
+title: Time Profiling
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA支持对图中缓冲的每个运算符进行时间分析。为了利用时间分析功能，我们首先调用`device.SetVerbosity`方法来设置时间分析器的verbosity，然后调用`device.PrintTimeProfiling`来打印出时间分析的结果。
+
+## 设置时间分析Verbosity
+
+要使用时间分析功能，我们需要设置verbosity。有三个级别的verbosity。在默认值`verbosity == 0`的情况下，它不会进行任何时间分析。当我们设置`verbosity == 1`时，它将对前向和后向传播时间进行分析。当`verbosity == 2`时，它将对图中每个缓冲操作所花费的时间进行分析。
+
+以下是设置时间分析功能的示例代码：
+
+```python
+# create a device
+from singa import device
+dev = device.create_cuda_gpu()
+# set the verbosity
+verbosity = 2
+dev.SetVerbosity(verbosity)
+# optional: skip the first 5 iterations when profiling the time
+dev.SetSkipIteration(5)
+```
+
+那么，当我们在程序的最后完成训练后，我们就可以通过调用`device.PrintTimeProfiling`方法来打印时间分析结果。
+
+```python
+dev.PrintTimeProfiling()
+```
+
+## 不同Verbosity的输出示例
+
+我们可以运行ResNet的[示例](https://github.com/apache/singa/blob/master/examples/cnn/benchmark.py)，看看不同的verbosity设置的输出。
+
+1. `verbosity == 1`
+
+```
+Time Profiling:
+Forward Propagation Time : 0.0409127 sec
+Backward Propagation Time : 0.114813 sec
+```
+
+2. `verbosity == 2`
+
+```
+Time Profiling:
+OP_ID0. SetValue : 1.73722e-05 sec
+OP_ID1. cudnnConvForward : 0.000612724 sec
+OP_ID2. GpuBatchNormForwardTraining : 0.000559449 sec
+OP_ID3. ReLU : 0.000375004 sec
+OP_ID4. GpuPoolingForward : 0.000240041 sec
+OP_ID5. SetValue : 3.4176e-06 sec
+OP_ID6. cudnnConvForward : 0.000115619 sec
+OP_ID7. GpuBatchNormForwardTraining : 0.000150415 sec
+OP_ID8. ReLU : 9.95494e-05 sec
+OP_ID9. SetValue : 3.22432e-06 sec
+OP_ID10. cudnnConvForward : 0.000648668 sec
+OP_ID11. GpuBatchNormForwardTraining : 0.000149793 sec
+OP_ID12. ReLU : 9.92118e-05 sec
+OP_ID13. SetValue : 3.37728e-06 sec
+OP_ID14. cudnnConvForward : 0.000400953 sec
+OP_ID15. GpuBatchNormForwardTraining : 0.000572181 sec
+OP_ID16. SetValue : 3.21312e-06 sec
+OP_ID17. cudnnConvForward : 0.000398698 sec
+OP_ID18. GpuBatchNormForwardTraining : 0.00056836 sec
+OP_ID19. Add : 0.000542246 sec
+OP_ID20. ReLU : 0.000372783 sec
+OP_ID21. SetValue : 3.25312e-06 sec
+OP_ID22. cudnnConvForward : 0.000260731 sec
+OP_ID23. GpuBatchNormForwardTraining : 0.000149041 sec
+OP_ID24. ReLU : 9.9072e-05 sec
+OP_ID25. SetValue : 3.10592e-06 sec
+OP_ID26. cudnnConvForward : 0.000637481 sec
+OP_ID27. GpuBatchNormForwardTraining : 0.000152577 sec
+OP_ID28. ReLU : 9.90518e-05 sec
+OP_ID29. SetValue : 3.28224e-06 sec
+OP_ID30. cudnnConvForward : 0.000404586 sec
+OP_ID31. GpuBatchNormForwardTraining : 0.000569679 sec
+OP_ID32. Add : 0.000542291 sec
+OP_ID33. ReLU : 0.00037211 sec
+OP_ID34. SetValue : 3.13696e-06 sec
+OP_ID35. cudnnConvForward : 0.000261219 sec
+OP_ID36. GpuBatchNormForwardTraining : 0.000148281 sec
+OP_ID37. ReLU : 9.89299e-05 sec
+OP_ID38. SetValue : 3.25216e-06 sec
+OP_ID39. cudnnConvForward : 0.000633644 sec
+OP_ID40. GpuBatchNormForwardTraining : 0.000150711 sec
+OP_ID41. ReLU : 9.84902e-05 sec
+OP_ID42. SetValue : 3.18176e-06 sec
+OP_ID43. cudnnConvForward : 0.000402752 sec
+OP_ID44. GpuBatchNormForwardTraining : 0.000571523 sec
+OP_ID45. Add : 0.000542435 sec
+OP_ID46. ReLU : 0.000372539 sec
+OP_ID47. SetValue : 3.24672e-06 sec
+OP_ID48. cudnnConvForward : 0.000493054 sec
+OP_ID49. GpuBatchNormForwardTraining : 0.000293142 sec
+OP_ID50. ReLU : 0.000190047 sec
+OP_ID51. SetValue : 3.14784e-06 sec
+OP_ID52. cudnnConvForward : 0.00148837 sec
+OP_ID53. GpuBatchNormForwardTraining : 8.34794e-05 sec
+OP_ID54. ReLU : 5.23254e-05 sec
+OP_ID55. SetValue : 3.40096e-06 sec
+OP_ID56. cudnnConvForward : 0.000292971 sec
+OP_ID57. GpuBatchNormForwardTraining : 0.00029174 sec
+OP_ID58. SetValue : 3.3248e-06 sec
+OP_ID59. cudnnConvForward : 0.000590154 sec
+OP_ID60. GpuBatchNormForwardTraining : 0.000294149 sec
+OP_ID61. Add : 0.000275119 sec
+OP_ID62. ReLU : 0.000189268 sec
+OP_ID63. SetValue : 3.2704e-06 sec
+OP_ID64. cudnnConvForward : 0.000341232 sec
+OP_ID65. GpuBatchNormForwardTraining : 8.3304e-05 sec
+OP_ID66. ReLU : 5.23667e-05 sec
+OP_ID67. SetValue : 3.19936e-06 sec
+OP_ID68. cudnnConvForward : 0.000542484 sec
+OP_ID69. GpuBatchNormForwardTraining : 8.60537e-05 sec
+OP_ID70. ReLU : 5.2479e-05 sec
+OP_ID71. SetValue : 3.41824e-06 sec
+OP_ID72. cudnnConvForward : 0.000291295 sec
+OP_ID73. GpuBatchNormForwardTraining : 0.000292795 sec
+OP_ID74. Add : 0.000274438 sec
+OP_ID75. ReLU : 0.000189689 sec
+OP_ID76. SetValue : 3.21984e-06 sec
+OP_ID77. cudnnConvForward : 0.000338776 sec
+OP_ID78. GpuBatchNormForwardTraining : 8.484e-05 sec
+OP_ID79. ReLU : 5.29408e-05 sec
+OP_ID80. SetValue : 3.18208e-06 sec
+OP_ID81. cudnnConvForward : 0.000545542 sec
+OP_ID82. GpuBatchNormForwardTraining : 8.40976e-05 sec
+OP_ID83. ReLU : 5.2256e-05 sec
+OP_ID84. SetValue : 3.36256e-06 sec
+OP_ID85. cudnnConvForward : 0.000293003 sec
+OP_ID86. GpuBatchNormForwardTraining : 0.0002989 sec
+OP_ID87. Add : 0.000275041 sec
+OP_ID88. ReLU : 0.000189867 sec
+OP_ID89. SetValue : 3.1184e-06 sec
+OP_ID90. cudnnConvForward : 0.000340417 sec
+OP_ID91. GpuBatchNormForwardTraining : 8.39395e-05 sec
+OP_ID92. ReLU : 5.26544e-05 sec
+OP_ID93. SetValue : 3.2336e-06 sec
+OP_ID94. cudnnConvForward : 0.000539787 sec
+OP_ID95. GpuBatchNormForwardTraining : 8.2753e-05 sec
+OP_ID96. ReLU : 4.86758e-05 sec
+OP_ID97. SetValue : 3.24384e-06 sec
+OP_ID98. cudnnConvForward : 0.000287108 sec
+OP_ID99. GpuBatchNormForwardTraining : 0.000293127 sec
+OP_ID100. Add : 0.000269478 sec
+.
+.
+.
+```
diff --git a/docs-site/docs_chs/wheel-cpu-dev.md b/docs-site/docs_chs/wheel-cpu-dev.md
new file mode 100644
index 0000000..9f8f7af
--- /dev/null
+++ b/docs-site/docs_chs/wheel-cpu-dev.md
@@ -0,0 +1,12 @@
+---
+id: wheel-cpu-dev
+title: CPU only Wheel Packages (develop version)
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 3.0.0.dev200720
+
+- [Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720-cp36-cp36m-manylinux2014_x86_64.whl)
+- [Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720-cp37-cp37m-manylinux2014_x86_64.whl)
+- [Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720-cp38-cp38-manylinux2014_x86_64.whl)
diff --git a/docs-site/docs_chs/wheel-cpu.md b/docs-site/docs_chs/wheel-cpu.md
new file mode 100644
index 0000000..430d9d1
--- /dev/null
+++ b/docs-site/docs_chs/wheel-cpu.md
@@ -0,0 +1,18 @@
+---
+id: wheel-cpu
+title: CPU only Wheel Packages
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 3.1.0
+
+- [Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0-cp36-cp36m-manylinux2014_x86_64.whl)
+- [Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0-cp37-cp37m-manylinux2014_x86_64.whl)
+- [Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0-cp38-cp38-manylinux2014_x86_64.whl)
+
+## 3.0.0
+
+- [Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0-cp36-cp36m-manylinux2014_x86_64.whl)
+- [Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0-cp37-cp37m-manylinux2014_x86_64.whl)
+- [Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0-cp38-cp38-manylinux2014_x86_64.whl)
diff --git a/docs-site/docs_chs/wheel-gpu-dev.md b/docs-site/docs_chs/wheel-gpu-dev.md
new file mode 100644
index 0000000..3c061e9
--- /dev/null
+++ b/docs-site/docs_chs/wheel-gpu-dev.md
@@ -0,0 +1,12 @@
+---
+id: wheel-gpu-dev
+title: Wheel Packages with CUDA enabled (develop version)
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 3.0.0.dev200720
+
+- [CUDA10.2, cuDNN 7.6.5, Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720%2Bcuda10.2-cp36-cp36m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720%2Bcuda10.2-cp37-cp37m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720%2Bcuda10.2-cp38-cp38-manylinux2014_x86_64.whl)
diff --git a/docs-site/docs_chs/wheel-gpu.md b/docs-site/docs_chs/wheel-gpu.md
new file mode 100644
index 0000000..2c0a1a9
--- /dev/null
+++ b/docs-site/docs_chs/wheel-gpu.md
@@ -0,0 +1,21 @@
+---
+id: wheel-gpu
+title: Wheel Packages with CUDA Enabled
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 3.1.0
+
+- [CUDA10.2, cuDNN 7.6.5, Python
+ 3.6]https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0%2Bcuda10.2-cp36-cp36m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python
+ 3.7]https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0%2Bcuda10.2-cp37-cp37m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python
+ 3.8]https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0%2Bcuda10.2-cp38-cp38-manylinux2014_x86_64.whl)
+
+## 3.0.0
+
+- [CUDA10.2, cuDNN 7.6.5, Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0%2Bcuda10.2-cp36-cp36m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0%2Bcuda10.2-cp37-cp37m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0%2Bcuda10.2-cp38-cp38-manylinux2014_x86_64.whl)
diff --git a/docs-site/website/pages/en/versions.js b/docs-site/website/pages/en/versions.js
index c322376..d27bd5a 100755
--- a/docs-site/website/pages/en/versions.js
+++ b/docs-site/website/pages/en/versions.js
@@ -30,6 +30,7 @@
const CWD = process.cwd()
// versions post docusaurus-upgrade Nov 2019
const versions = require(`${CWD}/versions.json`)
+const versions_otherlang = require(`${CWD}/versions_otherlang.json`)
// versions pre docusaurus-upgrade Nov 2019
const oldversions = require(`${CWD}/oldversions.json`)
@@ -54,7 +55,7 @@
<a
href={`${siteConfig.baseUrl}${siteConfig.docsUrl}/${
props.language ? props.language + "/" : ""
- }installation`}
+ }installation`}
>
Documentation
</a>
@@ -63,7 +64,7 @@
<a
href={`${siteConfig.baseUrl}${siteConfig.docsUrl}/${
props.language ? props.language + "/" : ""
- }releases/RELEASE_NOTES_${latestVersion}.html`}
+ }releases/RELEASE_NOTES_${latestVersion}.html`}
>
Release Notes
</a>
@@ -84,7 +85,7 @@
<a
href={`${siteConfig.baseUrl}${siteConfig.docsUrl}/${
props.language ? props.language + "/" : ""
- }next/installation`}
+ }next/installation`}
>
Documentation
</a>
@@ -102,7 +103,7 @@
<a
href={`${siteConfig.baseUrl}${siteConfig.docsUrl}/${
props.language ? props.language + "/" : ""
- }download-singa#incubating-v010-8-october-2015`}
+ }download-singa#incubating-v010-8-october-2015`}
>
this page
</a>{" "}
@@ -112,14 +113,14 @@
<tbody>
{versions.map(
version =>
- version !== latestVersion && (
+ version !== latestVersion && !versions_otherlang.includes(version) && (
<tr key={version}>
<th>{version}</th>
<td>
<a
href={`${siteConfig.baseUrl}${siteConfig.docsUrl}/${
props.language ? props.language + "/" : ""
- }${version}/installation`}
+ }${version}/installation`}
>
Documentation
</a>
@@ -128,7 +129,7 @@
<a
href={`${siteConfig.baseUrl}${siteConfig.docsUrl}/${
props.language ? props.language + "/" : ""
- }releases/RELEASE_NOTES_${version}.html`}
+ }releases/RELEASE_NOTES_${version}.html`}
>
Release Notes
</a>
@@ -151,7 +152,7 @@
<a
href={`${siteConfig.baseUrl}${siteConfig.docsUrl}/${
props.language ? props.language + "/" : ""
- }releases/RELEASE_NOTES_${version}.html`}
+ }releases/RELEASE_NOTES_${version}.html`}
>
Release Notes
</a>
@@ -165,6 +166,29 @@
You can find past versions of this project on{" "}
<a href={repoUrl}>GitHub</a>.
</p>
+ <h3 id="rc">Other Languages</h3>
+ <p>
+ Here you can find the documentation of Apache SINGA in other languages
+ </p>
+ <table className="versions">
+ <tbody>
+ {versions_otherlang.map(
+ version =>
+ <tr key={version}>
+ <th>{version}</th>
+ <td>
+ <a
+ href={`${siteConfig.baseUrl}${siteConfig.docsUrl}/${
+ props.language ? props.language + "/" : ""
+ }${version}/installation`}
+ >
+ Documentation
+ </a>
+ </td>
+ </tr>
+ )}
+ </tbody>
+ </table>
</div>
</Container>
</div>
diff --git a/docs-site/website/sidebars.json b/docs-site/website/sidebars.json
index 7bbe66a..34e497f 100644
--- a/docs-site/website/sidebars.json
+++ b/docs-site/website/sidebars.json
@@ -13,10 +13,11 @@
"graph",
"onnx",
"dist-train",
- "time-profiling"
+ "time-profiling",
+ "half-precision"
],
"Development": [
- "download-singa",
+ "downloads",
"build",
"contribute-code",
"contribute-docs",
diff --git a/docs-site/website/versioned_docs/version-2.0.0/download-singa.md b/docs-site/website/versioned_docs/version-2.0.0/download-singa.md
index c9d3dd3..c1fa985 100644
--- a/docs-site/website/versioned_docs/version-2.0.0/download-singa.md
+++ b/docs-site/website/versioned_docs/version-2.0.0/download-singa.md
@@ -1,7 +1,7 @@
---
-id: version-2.0.0-download-singa
+id: version-2.0.0-downloads
title: Download SINGA
-original_id: download-singa
+original_id: downloads
---
<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
diff --git a/docs-site/website/versioned_docs/version-3.0.0.rc1/download-singa.md b/docs-site/website/versioned_docs/version-3.0.0.rc1/download-singa.md
index 8bea604..c658a1b 100644
--- a/docs-site/website/versioned_docs/version-3.0.0.rc1/download-singa.md
+++ b/docs-site/website/versioned_docs/version-3.0.0.rc1/download-singa.md
@@ -1,7 +1,7 @@
---
-id: version-3.0.0.rc1-download-singa
+id: version-3.0.0.rc1-downloads
title: Download SINGA
-original_id: download-singa
+original_id: downloads
---
<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
diff --git a/docs-site/website/versioned_docs/version-3.0.0/download.md b/docs-site/website/versioned_docs/version-3.0.0/download.md
index b05fe49..c34aba3 100644
--- a/docs-site/website/versioned_docs/version-3.0.0/download.md
+++ b/docs-site/website/versioned_docs/version-3.0.0/download.md
@@ -1,7 +1,7 @@
---
-id: version-3.0.0-download-singa
+id: version-3.0.0-downloads
title: Download SINGA
-original_id: download-singa
+original_id: downloads
---
<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
diff --git a/docs-site/website/versioned_docs/version-3.1.0/autograd.md b/docs-site/website/versioned_docs/version-3.1.0/autograd.md
index e190cff..f2dbba5 100644
--- a/docs-site/website/versioned_docs/version-3.1.0/autograd.md
+++ b/docs-site/website/versioned_docs/version-3.1.0/autograd.md
@@ -193,8 +193,7 @@
### Using the Model API
-The following <<<<<<< HEAD
-[example](https://github.com/apache/singa/blob/master/examples/cnn/model/cnn.py)
+The following [example](https://github.com/apache/singa/blob/master/examples/cnn/model/cnn.py)
implements a CNN model using the [Model API](./graph).
#### Define the subclass of Model
diff --git a/docs-site/website/versioned_docs/version-3.1.0/contribute-docs.md b/docs-site/website/versioned_docs/version-3.1.0/contribute-docs.md
new file mode 100644
index 0000000..f91f2fb
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0/contribute-docs.md
@@ -0,0 +1,105 @@
+---
+id: version-3.1.0-contribute-docs
+title: How to Contribute to Documentation
+original_id: contribute-docs
+---
+
+<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+There are two types of documentation, namely markdown files and API usage
+reference. This guideline introduces some tools and instruction in preparing the
+source markdown files and API comments.
+
+The markdown files will be built into HTML pages via
+[Docusaurus](https://docusaurus.io/); The API comments (from the source code)
+will be used to generate API reference pages using Sphinx (for Python) and
+Doxygen (for CPP).
+
+## Markdown Files
+
+Try to follow the
+[Google Documentation style](https://developers.google.com/style). For example,
+
+1. Remove 'please' from an instruction. 'Please click...' VS 'Click ...'.
+2. Follow the
+ [standard captitalization rules](https://owl.purdue.edu/owl/general_writing/mechanics/help_with_capitals.html).
+3. Use 'you' instead of 'we' in the instructions.
+4. Use present tense and avoid 'will'
+5. Prefer active voice than passive voice.
+
+In addition, to make the documentation consistent,
+
+1. Keep the line short, e.g., length<=80
+2. Use the relative path assuming that we are in the root folder of the repo,
+ e.g., `doc-site/docs` refers to `singa-doc/docs-site/docs`
+3. Higlight the command, path, class function and variable using backticks,
+ e.g., `Tensor`, `singa-doc/docs-site/docs`.
+4. To hightlight other terms/concepts, use _graph_ or **graph**
+
+The [prettier tool](https://prettier.io/) used by this project will auto-format
+the code according to the
+[configuration](https://github.com/apache/singa-doc/blob/master/docs-site/.prettierrc)
+when we do `git commit`. For example, it will wrap the text in the markdown file
+to at most 80 characters (except the lines for comments).
+
+When introducing a concept (e.g., the `Tensor` class), provide the overview (the
+purpose and relation to other concepts), APIs and examples. Google colab can be
+used to demonstrate the usage.
+
+Refer to [this page](https://github.com/apache/singa-doc/tree/master/docs-site)
+for the details on how to edit the markdown files and build the website.
+
+## API References
+
+### CPP API
+
+Follow the
+[Google CPP Comments Style](https://google.github.io/styleguide/cppguide.html#Comments).
+
+To generate docs, run "doxygen" from the doc folder (Doxygen >= 1.8 recommended)
+
+### Python API
+
+Follow the
+[Google Python DocString Style](http://google.github.io/styleguide/pyguide.html#38-comments-and-docstrings).
+
+## Visual Studio Code (vscode)
+
+If you use vscode as the editor, the following plugins are useful.
+
+### Docstring Snippet
+
+[autoDocstring](https://marketplace.visualstudio.com/items?itemName=njpwerner.autodocstring)
+generates the docstring of functions, classes, etc. Choose the DocString Format
+to `google`.
+
+### Spell Check
+
+[Code Spell Checker](https://marketplace.visualstudio.com/items?itemName=streetsidesoftware.code-spell-checker)
+can be configured to check the comments of the code, or .md and .rst files.
+
+To do spell check only for comments of Python code, add the following snippet
+via `File - Preferences - User Snippets - python.json`
+
+ "cspell check" : {
+ "prefix": "cspell",
+ "body": [
+ "# Directives for doing spell check only for python and c/cpp comments",
+ "# cSpell:includeRegExp #.* ",
+ "# cSpell:includeRegExp (\"\"\"|''')[^\1]*\1",
+ "# cSpell: CStyleComment",
+ ],
+ "description": "# spell check only for python comments"
+ }
+
+To do spell check only for comments of Cpp code, add the following snippet via
+`File - Preferences - User Snippets - cpp.json`
+
+ "cspell check" : {
+ "prefix": "cspell",
+ "body": [
+ "// Directive for doing spell check only for cpp comments",
+ "// cSpell:includeRegExp CStyleComment",
+ ],
+ "description": "# spell check only for cpp comments"
+ }
diff --git a/docs-site/website/versioned_docs/version-3.1.0/device.md b/docs-site/website/versioned_docs/version-3.1.0/device.md
new file mode 100644
index 0000000..e50b01c
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0/device.md
@@ -0,0 +1,33 @@
+---
+id: version-3.1.0-device
+title: Device
+original_id: device
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+The Device abstract represents a hardware device with memory and computing
+units. All [Tensor operations](./tensor) are scheduled by the resident device
+for execution. Tensor memory is also managed by the device's memory manager.
+Therefore, optimization of memory and execution are implemented in the Device
+class.
+
+## Specific devices
+
+Currently, SINGA has three Device implmentations,
+
+1. CudaGPU for an Nvidia GPU card which runs Cuda code
+2. CppCPU for a CPU which runs Cpp code
+3. OpenclGPU for a GPU card which runs OpenCL code
+
+## Example Usage
+
+The following code provides examples of creating devices:
+
+```python
+from singa import device
+cuda = device.create_cuda_gpu_on(0) # use GPU card of ID 0
+host = device.get_default_device() # get the default host device (a CppCPU)
+ary1 = device.create_cuda_gpus(2) # create 2 devices, starting from ID 0
+ary2 = device.create_cuda_gpus([0,2]) # create 2 devices on ID 0 and 2
+```
diff --git a/docs-site/website/versioned_docs/version-3.1.0/download.md b/docs-site/website/versioned_docs/version-3.1.0/download.md
index a607672..7e5aefc 100644
--- a/docs-site/website/versioned_docs/version-3.1.0/download.md
+++ b/docs-site/website/versioned_docs/version-3.1.0/download.md
@@ -1,7 +1,7 @@
---
-id: version-3.1.0-download-singa
+id: version-3.1.0-downloads
title: Download SINGA
-original_id: download-singa
+original_id: downloads
---
<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
diff --git a/docs-site/website/versioned_docs/version-3.1.0/git-workflow.md b/docs-site/website/versioned_docs/version-3.1.0/git-workflow.md
new file mode 100644
index 0000000..e912c0e
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0/git-workflow.md
@@ -0,0 +1,131 @@
+---
+id: version-3.1.0-git-workflow
+title: Git Workflow
+original_id: git-workflow
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## For Developers
+
+1. Fork the [SINGA Github repository](https://github.com/apache/singa) to your
+ own Github account.
+
+2. Clone the **repo** (short for repository) from your Github
+
+ ```shell
+ git clone https://github.com/<Github account>/singa.git
+ git remote add upstream https://github.com/apache/singa.git
+ ```
+
+3. Create a new branch (e.g., `feature-foo` or `fixbug-foo`), work on it and
+ commit your code.
+
+ ```shell
+ git checkout dev
+ git checkout -b feature-foo
+ # write your code
+ git add <created/updated files>
+ git commit
+ ```
+
+ The commit message should include:
+
+ - A descriptive Title.
+ - A detailed description. If the commit is to fix a bug, the description
+ should ideally include a short reproduction of the problem. For new
+ features, it may include the motivation/purpose of this new feature.
+
+ If your branch has many small commits, you need to clean those commits via
+
+ ```shell
+ git rebase -i <commit id>
+ ```
+
+ You can
+ [squash and reword](https://help.github.com/en/articles/about-git-rebase) the
+ commits.
+
+4. When you are working on the code, the `dev` of SINGA may have been updated by
+ others; In this case, you need to pull the latest dev
+
+ ```shell
+ git checkout dev
+ git pull upstream dev:dev
+ ```
+
+5. [Rebase](https://git-scm.com/book/en/v2/Git-Branching-Rebasing) `feature-foo`
+ onto the `dev` branch and push commits to your own Github account (the new
+ branch). The rebase operation is to make the commit history clean. The
+ following git instructors should be executed after committing the current
+ work:
+
+ ```shell
+ git checkout feature-foo
+ git rebase dev
+ git push origin feature-foo:feature-foo
+ ```
+
+ The rebase command does the
+ [following steps](https://git-scm.com/book/en/v2/Git-Branching-Rebasing):
+ "This operation works by going to the common ancestor of the two branches
+ (the one you’re on and the one you’re rebasing onto), getting the diff
+ introduced by each commit of the branch you’re on, saving those diffs to
+ temporary files, resetting the current branch to the same commit as the
+ branch you are rebasing onto, and finally applying each change in turn."
+ Therefore, after executing it, you will be still on the feature branch, but
+ your own commit IDs/hashes are changed since the diffs are committed during
+ rebase; and your branch now has the latest code from the dev branch and your
+ own branch.
+
+6. Open a pull request (PR) against the dev branch of apache/singa on Github
+ website. If you want to inform other contributors who worked on the same
+ files, you can find the file(s) on Github and click "Blame" to see a
+ line-by-line annotation of who changed the code last. Then, you can add
+ @username in the PR description to ping them immediately. Please state that
+ the contribution is your original work and that you license the work to the
+ project under the project's open source license. Further commits (e.g., bug
+ fix) to your new branch will be added to this pull request automatically by
+ Github.
+
+7. Wait for committers to review the PR. During this time, the dev of SINGA may
+ have been updated by others, and then you need to
+ [merge the latest dev](https://docs.fast.ai/dev/git.html#how-to-keep-your-feature-branch-up-to-date)
+ to resolve conflicts. Some people
+ [rebase the PR onto the latest dev](https://github.com/edx/edx-platform/wiki/How-to-Rebase-a-Pull-Request)
+ instead of merging. However, if other developers fetch this PR to add new
+ features and then send PR, the rebase operation would introduce **duplicate
+ commits** (with different hash) in the future PR. See
+ [The Golden Rule of Rebasing](https://www.atlassian.com/git/tutorials/merging-vs-rebasing)
+ for the details of when to avoid using rebase. Another simple solution to
+ update the PR (to fix conflicts or commit errors) is to checkout a new branch
+ from the latest dev branch of Apache SINGAS repo; copy and paste the
+ updated/added code; commit and send a new PR.
+
+## For Committers
+
+Committers can merge the pull requests (PRs) into the dev branch of the upstream
+repo. Before merging each PR, the committer should
+
+- check the commit message (content and format)
+- check the changes to existing code. API changes should be recorded
+- check the Travis testing results for code/doc format and unit tests
+
+There are two approaches to merge a pull request:
+
+- On Github. Follow the [instructions](https://gitbox.apache.org/setup/) to
+ connect your Apache account with your Github account. After that you can
+ directly merge PRs on GitHub.
+- To merge pull request https://github.com/apache/singa/pull/xxx via command
+ line, the following instructions should be executed,
+
+ ```shell
+ git clone https://github.com/apache/singa.git
+ git remote add asf https://gitbox.apache.org/repos/asf/singa.git
+ git fetch origin pull/xxx/head:prxxx
+ git checkout dev
+ git merge --no-ff prxxx
+ git push asf dev:dev
+ ```
+
+ Do not use rebase to merge the PR; and disable fast forward.
diff --git a/docs-site/website/versioned_docs/version-3.1.0/history-singa.md b/docs-site/website/versioned_docs/version-3.1.0/history-singa.md
new file mode 100644
index 0000000..f408e8d
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0/history-singa.md
@@ -0,0 +1,42 @@
+---
+id: version-3.1.0-history-singa
+title: History of SINGA
+original_id: history-singa
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## History
+
+SINGA was initiated by the DB System Group at National University of Singapore
+in 2014, in collaboration with the database group of Zhejiang University. Please
+cite the following two papers if you use SINGA in your research:
+
+- B.C. Ooi, K.-L. Tan, S. Wang, W. Wang, Q. Cai, G. Chen, J. Gao, Z. Luo, A. K.
+ H. Tung, Y. Wang, Z. Xie, M. Zhang, and K. Zheng.
+ [SINGA: A distributed deep learning platform](http://www.comp.nus.edu.sg/~ooibc/singaopen-mm15.pdf).
+ ACM Multimedia (Open Source Software Competition) 2015
+
+- W. Wang, G. Chen, T. T. A. Dinh, B. C. Ooi, K.-L.Tan, J. Gao, and S. Wang.
+ [SINGA: putting deep learning in the hands of multimedia users](http://www.comp.nus.edu.sg/~ooibc/singa-mm15.pdf).
+ ACM Multimedia 2015.
+
+Rafiki is a sub module of SINGA. Please cite the following paper if you use
+Rafiki in your research:
+
+- Wei Wang, Jinyang Gao, Meihui Zhang, Sheng Wang, Gang Chen, Teck Khim Ng, Beng
+ Chin Ooi, Jie Shao, Moaz Reyad.
+ [Rafiki: Machine Learning as an Analytics Service System](http://www.vldb.org/pvldb/vol12/p128-wang.pdf).
+ [VLDB 2019](http://vldb.org/2019/)
+ ([BibTex](https://dblp.org/rec/bib2/journals/pvldb/WangWGZCNOS18.bib)).
+
+Companies like [NetEase](http://tech.163.com/17/0602/17/CLUL016I00098GJ5.html),
+[yzBigData](http://www.yzbigdata.com/en/index.html),
+[Shentilium](https://shentilium.com/), [Foodlg](http://www.foodlg.com/) and
+[Medilot](https://medilot.com/technologies) are using SINGA for their
+applications.
+
+## License
+
+SINGA is released under
+[Apache License Version 2.0](http://www.apache.org/licenses/LICENSE-2.0)
diff --git a/docs-site/website/versioned_docs/version-3.1.0/issue-tracking.md b/docs-site/website/versioned_docs/version-3.1.0/issue-tracking.md
new file mode 100644
index 0000000..27f9564
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0/issue-tracking.md
@@ -0,0 +1,12 @@
+---
+id: version-3.1.0-issue-tracking
+title: Issue Tracking
+original_id: issue-tracking
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA used [JIRA](https://issues.apache.org/jira/browse/singa) to manage issues
+including bugs, new features and discussions.
+
+We are now moving to [Github Issues](https://github.com/apache/singa/issues).
diff --git a/docs-site/website/versioned_docs/version-3.1.0/mail-lists.md b/docs-site/website/versioned_docs/version-3.1.0/mail-lists.md
new file mode 100644
index 0000000..4829d03
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0/mail-lists.md
@@ -0,0 +1,16 @@
+---
+id: version-3.1.0-mail-lists
+title: Project Mailing Lists
+original_id: mail-lists
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+These are the mailing lists that have been established for this project. For
+each list, there is a subscribe, unsubscribe, and an archive link.
+
+| Name | Post | Subscribe | Unsubscribe | Archive |
+| ----------- | ------------------------------------ | ---------------------------------------------------------------- | -------------------------------------------------------------------- | ----------------------------------------------------------------------------------- |
+| Development | <dev@singa.incubator.apache.org> | [Subscribe](mailto:dev-subscribe@singa.incubator.apache.org) | [Unsubscribe](mailto:dev-unsubscribe@singa.incubator.apache.org.) | [mail-archives.apache.org](http://mail-archives.apache.org/mod_mbox/singa-dev/) |
+| Commits | <commits@singa.incubator.apache.org> | [Subscribe](mailto:commits-subscribe@singa.incubator.apache.org) | [Unsubscribe](mailto:commits-unsubscribe@singa.incubator.apache.org) | [mail-archives.apache.org](http://mail-archives.apache.org/mod_mbox/singa-commits/) |
+| Security | <security@singa.apache.org> | private | private | private |
diff --git a/docs-site/website/versioned_docs/version-3.1.0/security.md b/docs-site/website/versioned_docs/version-3.1.0/security.md
new file mode 100644
index 0000000..bbfa924
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0/security.md
@@ -0,0 +1,10 @@
+---
+id: version-3.1.0-security
+title: Security
+original_id: security
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+Users can report security vulnerabilities to
+[SINGA Security Team Mail List](mailto:security@singa.apache.org)
diff --git a/docs-site/website/versioned_docs/version-3.1.0/source-repository.md b/docs-site/website/versioned_docs/version-3.1.0/source-repository.md
new file mode 100644
index 0000000..0edc829
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0/source-repository.md
@@ -0,0 +1,24 @@
+---
+id: version-3.1.0-source-repository
+title: Source Repository
+original_id: source-repository
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+This project uses [Git](http://git-scm.com/) to manage its source code.
+Instructions on Git use can be found at http://git-scm.com/documentation .
+
+## Repository
+
+The following is a link to the online source repository.
+
+- https://gitbox.apache.org/repos/asf?p=singa.git
+
+There is a Github mirror at
+
+- https://github.com/apache/singa
+
+The code can be cloned from either repo, e.g.,
+
+ git clone https://github.com/apache/singa.git
diff --git a/docs-site/website/versioned_docs/version-3.1.0/wheel-cpu-dev.md b/docs-site/website/versioned_docs/version-3.1.0/wheel-cpu-dev.md
index 8de7fb1..773323b 100644
--- a/docs-site/website/versioned_docs/version-3.1.0/wheel-cpu-dev.md
+++ b/docs-site/website/versioned_docs/version-3.1.0/wheel-cpu-dev.md
@@ -1,6 +1,6 @@
---
id: version-3.1.0-wheel-cpu-dev
-title: CPU only Wheel Packages (develop version)
+title: CPU only (dev version)
original_id: wheel-cpu-dev
---
diff --git a/docs-site/website/versioned_docs/version-3.1.0/wheel-cpu.md b/docs-site/website/versioned_docs/version-3.1.0/wheel-cpu.md
index 824d4a4..9e702ee 100644
--- a/docs-site/website/versioned_docs/version-3.1.0/wheel-cpu.md
+++ b/docs-site/website/versioned_docs/version-3.1.0/wheel-cpu.md
@@ -1,6 +1,6 @@
---
id: version-3.1.0-wheel-cpu
-title: CPU only Wheel Packages
+title: CPU only
original_id: wheel-cpu
---
diff --git a/docs-site/website/versioned_docs/version-3.1.0/wheel-gpu-dev.md b/docs-site/website/versioned_docs/version-3.1.0/wheel-gpu-dev.md
index 120125b..5dd149a 100644
--- a/docs-site/website/versioned_docs/version-3.1.0/wheel-gpu-dev.md
+++ b/docs-site/website/versioned_docs/version-3.1.0/wheel-gpu-dev.md
@@ -1,6 +1,6 @@
---
id: version-3.1.0-wheel-gpu-dev
-title: Wheel Packages with CUDA enabled (develop version)
+title: CUDA enabled (dev version)
original_id: wheel-gpu-dev
---
diff --git a/docs-site/website/versioned_docs/version-3.1.0/wheel-gpu.md b/docs-site/website/versioned_docs/version-3.1.0/wheel-gpu.md
index 7bd0221..32cf97c 100644
--- a/docs-site/website/versioned_docs/version-3.1.0/wheel-gpu.md
+++ b/docs-site/website/versioned_docs/version-3.1.0/wheel-gpu.md
@@ -1,6 +1,6 @@
---
id: version-3.1.0-wheel-gpu
-title: Wheel Packages with CUDA Enabled
+title: CUDA Enabled
original_id: wheel-gpu
---
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/autograd.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/autograd.md
new file mode 100644
index 0000000..c626ee3
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/autograd.md
@@ -0,0 +1,241 @@
+---
+id: version-3.1.0_Chinese-autograd
+title: Autograd
+original_id: autograd
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+实现autograd有两种典型的方式，一种是通过如[Theano](http://deeplearning.net/software/theano/index.html)的符号微分（symbolic differentiation）或通过如[Pytorch](https://pytorch.org/docs/stable/notes/autograd.html)的反向微分（reverse differentialtion）。SINGA遵循Pytorch方式，即通过记录计算图，并在正向传播后自动应用反向传播。自动传播算法的详细解释请参阅[这里](https://pytorch.org/docs/stable/notes/autograd.html)。我们接下来对SINGA中的相关模块进行解释，并举例说明其使用方法。
+
+## 相关模块
+
+在autograd中涉及三个类，分别是`singa.tensor.Tensor`，`singa.autograd.Operation`和`singa.autograd.Layer`。在本篇的后续部分中，我们使用Tensor、Operation和Layer来指代这三个类。
+
+### Tensor
+
+Tensor的三个属性被autograd使用：
+
+- `.creator`是一个`Operation`实例。它记录了产生Tensor实例的这个操作。
+- `.request_grad`是一个布尔变量。它用于指示autograd算法是否需要计算张量的梯度。例如，在反向传播的过程中，线性层的权重矩阵和卷积层（非底层）的特征图的张量梯度应该被计算。
+- `.store_grad`是一个布尔变量。它用于指示张量的梯度是否应该被存储并由后向函数输出。例如，特征图的梯度是在反向传播过程中计算出来的，但不包括在反向函数的输出中。
+
+开发者可以改变Tensor实例的`requires_grad`和`stores_grad`。例如，如果将后者设置为True，那么相应的梯度就会被包含在后向函数的输出。需要注意的是，如果`stores_grad`是True，那么 `requires_grad`一定是真，反之亦然。
+
+
+### Operation
+
+它将一个或多个`Tensor`实例作为输入，然后输出一个或多个`Tensor`实例。例如，ReLU可以作为一个具体的Operation子类来实现。当一个`Operation`实例被调用时（实例化后），会执行以下两个步骤。
+
+1.记录源操作，即输入张量的`创建者`。
+2.通过调用成员函数`.forward()`进行计算。
+
+有两个成员函数用于前向和反向传播，即`.forward()`和`.backward()`。它们以`Tensor.data`作为输入（类型为`CTensor`），并输出`Ctensor`s。要添加一个特定的操作，子类`Operation`应该实现自己的`.forward()`和`.backward()`函数。在后向传播过程中，autograd的`backward()`函数会自动调用`backward()`函数来计算输入的梯度（根据`require_grad`字段的参数和约束）。
+
+### Layer
+
+对于那些需要参数的Operation，我们把它们封装成一个新的类，`Layer`。例如，卷积操作被封装到卷积层(Convolution layer)中。`层`管理（存储）参数，并调用相应的`Operation`来实现变换。
+
+## 样例
+
+在[example folder](https://github.com/apache/singa/tree/master/examples/autograd)中提供了很多样例。在这里我我们分析两个最具代表性的例子。
+
+### 只使用Operation
+
+下一段代码展示了一个只使用`Operation`的多层感知机（MLP）模型：
+
+#### 调用依赖包
+
+```python
+from singa.tensor import Tensor
+from singa import autograd
+from singa import opt
+```
+
+#### 创建权重矩阵和偏置向量
+
+在将`requires_grad`和`stores_grad`都设置为`True`的情况下，创建参数张量。
+
+```python
+w0 = Tensor(shape=(2, 3), requires_grad=True, stores_grad=True)
+w0.gaussian(0.0, 0.1)
+b0 = Tensor(shape=(1, 3), requires_grad=True, stores_grad=True)
+b0.set_value(0.0)
+
+w1 = Tensor(shape=(3, 2), requires_grad=True, stores_grad=True)
+w1.gaussian(0.0, 0.1)
+b1 = Tensor(shape=(1, 2), requires_grad=True, stores_grad=True)
+b1.set_value(0.0)
+```
+
+#### 训练
+
+```python
+inputs = Tensor(data=data) # data matrix
+target = Tensor(data=label) # label vector
+autograd.training = True # for training
+sgd = opt.SGD(0.05) # optimizer
+
+for i in range(10):
+ x = autograd.matmul(inputs, w0) # matrix multiplication
+ x = autograd.add_bias(x, b0) # add the bias vector
+ x = autograd.relu(x) # ReLU activation operation
+
+ x = autograd.matmul(x, w1)
+ x = autograd.add_bias(x, b1)
+
+ loss = autograd.softmax_cross_entropy(x, target)
+
+ for p, g in autograd.backward(loss):
+ sgd.update(p, g)
+```
+
+### 使用Operation和Layer
+
+下面的[例子](https://github.com/apache/singa/blob/master/examples/autograd/mnist_cnn.py)使用autograd模块提供的层实现了一个CNN模型。
+
+#### 创建层
+
+```python
+conv1 = autograd.Conv2d(1, 32, 3, padding=1, bias=False)
+bn1 = autograd.BatchNorm2d(32)
+pooling1 = autograd.MaxPool2d(3, 1, padding=1)
+conv21 = autograd.Conv2d(32, 16, 3, padding=1)
+conv22 = autograd.Conv2d(32, 16, 3, padding=1)
+bn2 = autograd.BatchNorm2d(32)
+linear = autograd.Linear(32 * 28 * 28, 10)
+pooling2 = autograd.AvgPool2d(3, 1, padding=1)
+```
+
+#### 定义正向传播函数
+
+在正向传播中的operations会被自动记录，用于反向传播。
+
+```python
+def forward(x, t):
+ # x is the input data (a batch of images)
+ # t is the label vector (a batch of integers)
+ y = conv1(x) # Conv layer
+ y = autograd.relu(y) # ReLU operation
+ y = bn1(y) # BN layer
+ y = pooling1(y) # Pooling Layer
+
+ # two parallel convolution layers
+ y1 = conv21(y)
+ y2 = conv22(y)
+ y = autograd.cat((y1, y2), 1) # cat operation
+ y = autograd.relu(y) # ReLU operation
+ y = bn2(y)
+ y = pooling2(y)
+
+ y = autograd.flatten(y) # flatten operation
+ y = linear(y) # Linear layer
+ loss = autograd.softmax_cross_entropy(y, t) # operation
+ return loss, y
+```
+
+#### 训练
+
+```python
+autograd.training = True
+for epoch in range(epochs):
+ for i in range(batch_number):
+ inputs = tensor.Tensor(device=dev, data=x_train[
+ i * batch_sz:(1 + i) * batch_sz], stores_grad=False)
+ targets = tensor.Tensor(device=dev, data=y_train[
+ i * batch_sz:(1 + i) * batch_sz], requires_grad=False, stores_grad=False)
+
+ loss, y = forward(inputs, targets) # forward the net
+
+ for p, gp in autograd.backward(loss): # auto backward
+ sgd.update(p, gp)
+```
+
+### Using the Model API
+
+下面的[样例](https://github.com/apache/singa/blob/master/examples/cnn/model/cnn.py)使用[Model API](./graph)实现了一个CNN模型。.
+
+#### 定义Model的子类
+
+定义模型类，它应该是Model的子类。只有这样，在训练阶段使用的所有操作才会形成一个计算图以便进行分析。图中的操作将被按时序规划并有效执行，模型类中也可以包含层。
+
+```python
+class MLP(model.Model): # the model is a subclass of Model
+
+ def __init__(self, data_size=10, perceptron_size=100, num_classes=10):
+ super(MLP, self).__init__()
+
+ # init the operators, layers and other objects
+ self.relu = layer.ReLU()
+ self.linear1 = layer.Linear(perceptron_size)
+ self.linear2 = layer.Linear(num_classes)
+ self.softmax_cross_entropy = layer.SoftMaxCrossEntropy()
+
+ def forward(self, inputs): # define the forward function
+ y = self.linear1(inputs)
+ y = self.relu(y)
+ y = self.linear2(y)
+ return y
+
+ def train_one_batch(self, x, y):
+ out = self.forward(x)
+ loss = self.softmax_cross_entropy(out, y)
+ self.optimizer(loss)
+ return out, loss
+
+ def set_optimizer(self, optimizer): # attach an optimizer
+ self.optimizer = optimizer
+```
+
+#### 训练
+
+```python
+# create a model instance
+model = MLP()
+# initialize optimizer and attach it to the model
+sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
+model.set_optimizer(sgd)
+# input and target placeholders for the model
+tx = tensor.Tensor((batch_size, 1, IMG_SIZE, IMG_SIZE), dev, tensor.float32)
+ty = tensor.Tensor((batch_size, num_classes), dev, tensor.int32)
+# compile the model before training
+model.compile([tx], is_train=True, use_graph=True, sequential=False)
+
+# train the model iteratively
+for b in range(num_train_batch):
+ # generate the next mini-batch
+ x, y = ...
+
+ # Copy the data into input tensors
+ tx.copy_from_numpy(x)
+ ty.copy_from_numpy(y)
+
+ # Training with one batch
+ out, loss = model(tx, ty)
+```
+
+#### 保存模型checkpoint
+
+```python
+# define the path to save the checkpoint
+checkpointpath="checkpoint.zip"
+
+# save a checkpoint
+model.save_states(fpath=checkpointpath)
+```
+
+#### 加载模型checkpoint
+
+```python
+# define the path to load the checkpoint
+checkpointpath="checkpoint.zip"
+
+# load a checkpoint
+import os
+if os.path.exists(checkpointpath):
+ model.load_states(fpath=checkpointpath)
+```
+
+### Python API
+
+关于Python API的更多细节，请参考[这里](https://singa.readthedocs.io/en/latest/autograd.html#module-singa.autograd)。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/benchmark-train.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/benchmark-train.md
new file mode 100644
index 0000000..5290a32
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/benchmark-train.md
@@ -0,0 +1,17 @@
+---
+id: version-3.1.0_Chinese-benchmark-train
+title: Benchmark for Distributed Training
+original_id: benchmark-train
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+
+项目：我们使用深度卷积神经网络[ResNet-50](https://github.com/apache/singa/blob/master/examples/cnn/model/resnet.py)。它有50个卷积层，用于图像分类。它需要3.8个GFLOPs来通过网络处理一张图像（尺寸为224x224）。输入的图像大小为224x224。
+
+
+硬件方面：我们使用的是AWS的p2.8xlarge实例，每个实例有8个Nvidia Tesla K80 GPU，共96GB GPU内存，32个vCPU，488GB主内存，10Gbps网络带宽。
+
+衡量标准：我们衡量不同数量worker的每次迭代时间，以评估SINGA的可扩展性。Batch-size固定为每个GPU32个。采用同步训练方案。因此，有效的batch-size是`32N`，其中N是GPU的数量。我们与一个流行的开源系统进行比较，该系统采用参数服务器拓扑结构。选择第一个GPU作为服务器。
+
+ <br/> **可扩展性测试。条形为吞吐量，折线形为通信成本。**
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/build.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/build.md
new file mode 100644
index 0000000..35d174a
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/build.md
@@ -0,0 +1,431 @@
+---
+id: version-3.1.0_Chinese-build
+title: Build SINGA from Source
+original_id: build
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+源文件可以通过[tar.gz文件](https://dist.apache.org/repos/dist/dev/singa/)或git repo的形式下载：
+
+```shell
+$ git clone https://github.com/apache/singa.git
+$ cd singa/
+```
+
+如果您想为SINGA贡献代码，请参考[贡献代码](contribute-code.md)页面的步骤和要求。
+
+## 使用Conda构筑SINGA
+
+Conda-build 是一个构建工具，它从anaconda cloud安装依赖的库并执行构建脚本。
+
+安装conda-build(需要先安装conda)：
+
+```shell
+conda install conda-build
+```
+
+### 构建CPU版本
+
+构建SINGA的CPU版本：
+
+```shell
+conda build tool/conda/singa/
+```
+
+以上命令已经在Ubuntu（14.04，16.04和18.04）和macOS 10.11上测试过。更多信息请参考[Travis-CI page](https://travis-ci.org/apache/singa)页面。
+
+### 构建GPU版本
+
+要构建GPU版的SINGA，计算机必须装备有Nvida GPU，而且需要安装CUDA driver(>=384.81)、CUDA toolkit(>=9)和cuDNN(>=7)。以下两个Docker镜像提供了构建环境：
+
+1. apache/singa:conda-cuda9.0
+2. apache/singa:conda-cuda10.0
+
+构建环境准备好后，需要先导出CUDA版本，然后运行conda命令构建SINGA：
+
+```shell
+export CUDA=x.y (e.g. 9.0)
+conda build tool/conda/singa/
+```
+
+### 后处理
+
+生成的包文件的位置(`.tar.gz`)将打印在终端上，生成的包可以直接安装：
+
+```shell
+conda install -c conda-forge --use-local <path to the package file>
+```
+
+若要上传到anaconda云端供他人下载安装，需要在 anaconda 上注册一个账号，才能[上传包](https://docs.anaconda.com/anaconda-cloud/user-guide/getting-started/)：
+
+```shell
+conda install anaconda-client
+anaconda login
+anaconda upload -l main <path to the package file>
+```
+
+将包上传到云端后，您可以在[Anaconda Cloud](https://anaconda.org/)上看到，也可以通过以下命令查看：
+
+```shell
+conda search -c <anaconda username> singa
+```
+
+每个特定的SINGA软件包都由版本和构建字符串来标识。要安装一个特定的SINGA包，需要提供所有信息，例如：
+
+```shell
+conda install -c <anaconda username> -c conda-forge singa=2.1.0.dev=cpu_py36
+```
+
+为了使安装命令简单化，您可以创建以下附加包，这些包依赖于最新的CPU和GPU SINGA包：
+
+```console
+# for singa-cpu
+conda build tool/conda/cpu/ --python=3.6
+conda build tool/conda/cpu/ --python=3.7
+# for singa-gpu
+conda build tool/conda/gpu/ --python=3.6
+conda build tool/conda/gpu/ --python=3.7
+```
+
+因此，当您运行：
+
+```shell
+conda install -c <anaconda username> -c conda-forge singa-xpu
+```
+
+时(`xpu`表示'cpu' or 'gpu'), 相应的真正的SINGA包将作为依赖库被安装。
+
+## 使用本地工具在Ubuntu上构建SINGA
+
+请参阅 SINGA [Dockerfiles](https://github.com/apache/singa/blob/master/tool/docker/devel/ubuntu/cuda9/Dockerfile#L30)，了解在 Ubuntu 16.04 上安装依赖库的说明。您也可以使用 devel 映像创建一个 Docker 容器，并在容器中构建 SINGA。要使用GPU、DNNL、Python和单元测试来构建SINGA，请运行以下命令：
+
+```shell
+mkdir build # at the root of singa folder
+cd build
+cmake -DENABLE_TEST=ON -DUSE_CUDA=ON -DUSE_DNNL=ON -DUSE_PYTHON3=ON ..
+make
+cd python
+pip install .
+```
+CMake选项的详细内容在本页最后一节解释，上面最后一条命令是安装Python包。你也可以运行`pip install -e .`，它可以创建符号链接，而不是将 Python 文件复制到 site-package 文件夹中。
+
+如果SINGA在ENABLE_TEST=ON的情况下编译，您可以通过以下方式运行单元测试:
+
+```shell
+$ ./bin/test_singa
+```
+
+您可以看到所有的测试案例与测试结果。如果SINGA通过了所有测试，那么您就成功安装了SINGA。
+
+## 使用本地工具在Centos7上构建SINGA
+
+由于Centos7的软件包名称不同，因此从源码开始构建会有所不同。
+
+### 安装依赖项
+
+基础包和库文件：
+
+```shell
+sudo yum install freetype-devel libXft-devel ncurses-devel openblas-devel blas-devel lapack devel atlas-devel kernel-headers unzip wget pkgconfig zip zlib-devel libcurl-devel cmake curl unzip dh-autoreconf git python-devel glog-devel protobuf-devel
+```
+
+构建必需的包：
+
+```shell
+sudo yum group install "Development Tools"
+```
+
+若要安装swig：
+
+```shell
+sudo yum install pcre-devel
+wget http://prdownloads.sourceforge.net/swig/swig-3.0.10.tar.gz
+tar xvzf swig-3.0.10.tar.gz
+cd swig-3.0.10.tar.gz
+./configure --prefix=${RUN}
+make
+make install
+```
+
+安装gfortran：
+
+```shell
+sudo yum install centos-release-scl-rh
+sudo yum --enablerepo=centos-sclo-rh-testing install devtoolset-7-gcc-gfortran
+```
+
+安装pip和其他包：
+
+```shell
+sudo yum install epel-release
+sudo yum install python-pip
+pip install matplotlib numpy pandas scikit-learn pydot
+```
+
+### 安装SINGA
+
+按照《使用本地工具在Ubuntu上构建SINGA》的步骤1-5进行操作
+
+### 测试
+
+您可以通过如下方式进行
+
+```shell
+$ ./bin/test_singa
+```
+
+您可以看到所有的测试案例与测试结果。如果SINGA通过了所有测试，即表示安装成功。
+
+## 在Windows中编译SINGA
+在Windows上使用Python支持构建SINGA的说明可以在[install-win页面](install-win.md)找到。
+
+## 关于编译选项的更多细节
+
+### USE_MODULES (已过期废弃)
+
+如果没有安装protobuf和openblas，你可以用它们一起编译SINGA
+
+```shell
+$ In SINGA ROOT folder
+$ mkdir build
+$ cd build
+$ cmake -DUSE_MODULES=ON ..
+$ make
+```
+cmake 会下载 OpenBlas 和 Protobuf (2.6.1) 并与 SINGA 一起编译。
+
+您可以使用`ccmake ..`来配置编译选项。如果一些依赖的库不在系统默认路径中，则您需要手动导出以下环境变量：
+
+```shell
+export CMAKE_INCLUDE_PATH=<path to the header file folder>
+export CMAKE_LIBRARY_PATH=<path to the lib file folder>
+```
+
+### USE_PYTHON
+
+编译SINGA的Python封装器选项：
+
+```shell
+$ cmake -DUSE_PYTHON=ON ..
+$ make
+$ cd python
+$ pip install .
+```
+
+### USE_CUDA
+我们推荐用户安装CUDA和[cuDNN](https://developer.nvidia.com/cudnn)，以便在GPU上运行SINGA，以获得更好的性能。
+
+SINGA已经在CUDA 9/10和cuDNN 7上进行了测试。如果cuDNN安装在非系统文件夹中，例如 /home/bob/local/cudnn/，则需要执行以下命令来让cmake在编译时能够找到它们：
+
+```shell
+$ export CMAKE_INCLUDE_PATH=/home/bob/local/cudnn/include:$CMAKE_INCLUDE_PATH
+$ export CMAKE_LIBRARY_PATH=/home/bob/local/cudnn/lib64:$CMAKE_LIBRARY_PATH
+$ export LD_LIBRARY_PATH=/home/bob/local/cudnn/lib64:$LD_LIBRARY_PATH
+```
+
+CUDA和cuDNN的cmake选项应该设置成“ON”：
+
+```shell
+# Dependent libs are install already
+$ cmake -DUSE_CUDA=ON ..
+$ make
+```
+
+### USE_DNNL
+
+用户可以启用DNNL来提高CPU的计算性能，DNNL的安装指南可以在[这里](https://github.com/intel/mkl-dnn#installation)找到：
+
+
+SINGA在DNNL v1.1环境下已经进行过测试并通过，
+
+若要启用DNNL支持来编译SINGA:
+
+```shell
+# Dependent libs are installed already
+$ cmake -DUSE_DNNL=ON ..
+$ make
+```
+
+### USE_OPENCL
+
+SINGA使用opencl-headers和viennacl（版本1.7.1及以上）来支持OpenCL，它们可以通过如下方式安装：
+
+```shell
+# On Ubuntu 16.04
+$ sudo apt-get install opencl-headers, libviennacl-dev
+# On Fedora
+$ sudo yum install opencl-headers, viennacl
+```
+
+此外，你需要在你想运行OpenCL的平台安装OpenCL Installable Client Driver（ICD）。
+
+- 对于AMD和Nvidia的GPU，驱动包也应该安装与之匹配的OpenCL ICD。
+- 对于Intel的CPU和/或GPU，请从[Intel官方网站](https://software.intel.com/en-us/articles/opencl-drivers)上获取驱动程序。请注意，该网站上提供的驱动程序只支持最新的CPU和Iris GPU。
+- 对于旧的Intel CPU，你可以使用beignet-opencl-icd包。
+
+请注意，目前不建议在CPU上运行OpenCL，因为运行速度会很慢。内存传输是以整数秒为单位的（直观来说，CPU上是1000毫秒，而GPU上是1毫秒）。
+
+更多关于建立OpenCL工作环境的信息可以在[这里](https://wiki.tiker.net/OpenCLHowTo)找到。
+
+如果ViennaCL的软件包版本不是至少1.7.1，则需要从源码构建它：
+
+从[这个git repo](https://github.com/viennacl/viennacl-dev)clone版本库，切换（checkout）到release-1.7.1分支，然后构建它，并把项目路径添加到PATH，再把构建的库文件添加到LD_LIBRARY_PATH。
+
+构建支持OpenCL的SINGA（在SINGA 1.1上测试）：
+
+```shell
+$ cmake -DUSE_OPENCL=ON ..
+$ make
+```
+
+### PACKAGE
+
+这个设置是用来构建 Debian 软件包的。设置PACKAGE=ON，然后用make命令来编译软件包，如下所示：
+
+```shell
+$ cmake -DPACKAGE=ON
+$ make package
+```
+
+## FAQ
+
+- Q: 'import singa'阶段报错
+
+ A: 请检查`python -c "from singa import _singa_wrap`中的详细错误。有时是由依赖库引起的，比如protobuf有多个版本，cudnn缺失，numpy版本不匹配等问题。下面展示了不同情况下的解决方案
+
+ 1. 检查cudnn和cuda。如果cudnn缺失或与wheel包的版本不一致，你可以下载正确的cudnn版本到~/local/cudnn/，然后：
+
+ ```shell
+ $ echo "export LD_LIBRARY_PATH=/home/<yourname>/local/cudnn/lib64:$LD_LIBRARY_PATH" >> ~/.bashrc
+ ```
+
+ 2. 如果是protobuf的问题。你可以将protobuf (3.6.1)从源码安装到本地文件夹，比如 ~/local/，解压tar文件，然后：
+
+ ```shell
+ $ ./configure --prefix=/home/<yourname>local
+ $ make && make install
+ $ echo "export LD_LIBRARY_PATH=/home/<yourname>/local/lib:$LD_LIBRARY_PATH" >> ~/.bashrc
+ $ source ~/.bashrc
+ ```
+
+ 3. 如果找不到包括python在内的其他类库，则使用`pip`或`conda`创建虚拟环境.
+
+ 4. 如果不是上述原因造成的，请到`_singa_wrap.so`文件夹中查看：
+ ```shell
+ $ python
+ >> import importlib
+ >> importlib.import_module('_singa_wrap')
+ ```
+
+ 来检查错误信息。例如，如果是numpy的版本不匹配，错误信息会是：
+
+ ```shell
+ RuntimeError: module compiled against API version 0xb but this version of numpy is 0xa
+ ```
+
+ 那么你就需要更新numpy到更高版本。
+
+* Q: 运行`cmake ...`时出错，找不到依赖库。
+
+ A: 如果你还没有安装这些依赖库，请先安装它们。如果你在系统文件夹之外的文件夹中安装了库，例如/usr/local，那您需要手动导出以下变量:
+
+ ```shell
+ $ export CMAKE_INCLUDE_PATH=<path to your header file folder>
+ $ export CMAKE_LIBRARY_PATH=<path to your lib file folder>
+ ```
+
+- Q: 来自`make`的错误，例如linking阶段的错误.
+
+ A: 如果您的库文件在系统默认路径以外的其他文件夹中，则需要手动导出以下变量。
+
+ ```shell
+ $ export LIBRARY_PATH=<path to your lib file folder>
+ $ export LD_LIBRARY_PATH=<path to your lib file folder>
+ ```
+
+* Q: 来自头文件的错误，例如'cblas.h文件不存在'
+
+ A: 您需要手动将cblas.h的文件夹包含在CPLUS_INCLUDE_PATH中，例如：
+
+ ```shell
+ $ export CPLUS_INCLUDE_PATH=/opt/OpenBLAS/include:$CPLUS_INCLUDE_PATH
+ ```
+
+* Q: 在编译SINGA时，我收到错误信息`SSE2 instruction set not enabled`
+
+ A:您可以尝试如下指令:
+
+ ```shell
+ $ make CFLAGS='-msse2' CXXFLAGS='-msse2'
+ ```
+
+* Q:当试图导入.py文件时，从google.protobuf.internal收到`ImportError: cannot import name enum_type_wrapper`。
+
+ A: 您需要安装protobuf的Python绑定包，它可以通过如下方式安装：
+
+ ```shell
+ $ sudo apt-get install protobuf
+ ```
+
+ 或从源文件编译：
+
+ ```shell
+ $ cd /PROTOBUF/SOURCE/FOLDER
+ $ cd python
+ $ python setup.py build
+ $ python setup.py install
+ ```
+
+* Q: 当我从源码构建OpenBLAS时，被告知需要一个Fortran编译器。
+
+ A: 您可以通过如下方式编译OpenBLAS：
+
+ ```shell
+ $ make ONLY_CBLAS=1
+ ```
+ 或者通过如下方式安装：
+
+ ```shell
+ $ sudo apt-get install libopenblas-dev
+ ```
+
+* Q: 当我构建协议缓冲区时，它报告说在`/usr/lib64/libstdc++.so.6`中没有找到`GLIBC++_3.4.20`？
+
+ A: 这意味着链接器找到了libstdc++.so.6，但该库属于旧的GCC（用于编译和链接程序的GCC）版本。此程序依赖于新的libstdc++中定义的代码，而该代码属于较新版本的GCC，所以必须告诉链接器如何找到较新的libstdc++共享库。最简单的解决方法是找到正确的libstdc++，并将其导出到LD_LIBRARY_PATH。例如，如果下面命令的输出中列出了GLIBC++_3.4.20：
+
+ $ strings /usr/local/lib64/libstdc++.so.6|grep GLIBC++
+
+ 那么接下来需要设置环境变量为：
+
+ $ export LD_LIBRARY_PATH=/usr/local/lib64:$LD_LIBRARY_PATH
+
+* Q: 当构建glog时，报告说 "src/logging_unittest.cc:83:20: error: 'gflags' is not a namespace-name"。
+
+ A: 可能是由于安装的gflags用了不同的命名空间，比如 "google"，所以glog找不到'gflags'的命名空间。实际上建立glog并不需要gflags，所以你可以修改configure.ac文件来忽略gflags。
+
+ 1. cd to glog src directory
+ 2. change line 125 of configure.ac to "AC_CHECK_LIB(gflags, main, ac_cv_have_libgflags=0, ac_cv_have_libgflags=0)"
+ 3. autoreconf
+
+ 执行上述命令后，就可以重新构建glog了。
+
+* Q: 在使用虚拟环境时，每次运行pip install，都会重新安装numpy。但是，当运行`import numpy`时，numpy并没有被调用。
+
+ A: 可能是由于`PYTHONPATH`造成的，在使用虚拟环境时，应将`PYTHONPATH`设置为空，以避免与虚拟环境本身的路径冲突。
+
+* Q: 当从源代码编译PySINGA时，由于缺少<numpy/objectarray.h>，出现的编译错误。
+
+ A: 请安装numpy并导出numpy头文件的路径为
+
+ $ export CPLUS_INCLUDE_PATH=`python -c "import numpy; print numpy.get_include()"`:$CPLUS_INCLUDE_PATH
+
+* Q: 当我在Mac OS X中运行SINGA时，报错 "Fatal Python error:
+ PyThreadState_Get: no current thread Abort trap: 6"
+
+ A: 这个错误通常发生在系统上有多个版本的Python，并且是通过pip安装的SINGA (通过conda安装时不会出现这个问题)，例如，操作系统自带的版本和Homebrew安装的版本。PySINGA所链接的Python必须与Python解释器（interpreter）相同。 您可以通过 `which python` 检查您的解释器路径并通过`otool -L <path to _singa_wrap.so>`检查PySINGA链接的Python路径。要解决这个问题，请用正确的 Python 版本编译 SINGA。需要注意的是，如果您从源码编译 PySINGA，您需要在调用[cmake](http://stackoverflow.com/questions/15291500/i-have-2-versions-of-python-installed-but-cmake-is-using-older-version-how-do)时指定路径：
+
+
+ $ cmake -DPYTHON_LIBRARY=`python-config --prefix`/lib/libpython2.7.dylib -DPYTHON_INCLUDE_DIR=`python-config --prefix`/include/python2.7/ ..
+
+ 如果从二进制包中安装PySINGA，例如debian或wheel，那么你需要改变python解释器的路径，例如，重新设置$PATH，并把Python的正确路径放在前面的位置。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/contribute-code.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/contribute-code.md
new file mode 100644
index 0000000..656d5d6
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/contribute-code.md
@@ -0,0 +1,109 @@
+---
+id: version-3.1.0_Chinese-contribute-code
+title: How to Contribute Code
+original_id: contribute-code
+---
+
+<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 代码风格
+
+SINGA代码库在[CPP](http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml)和[Python](http://google.github.io/styleguide/pyguide.html)代码中都遵循Google风格。
+
+强制执行Google编码风格的一个简单方法是使用Visual Studio Code编辑器中的linting和格式化工具:
+
+- [C/C++扩展](https://marketplace.visualstudio.com/items?itemName=ms-vscode.cpptools)
+- [Python扩展](https://marketplace.visualstudio.com/items?itemName=ms-python.python)
+- [cpplint扩展](https://marketplace.visualstudio.com/items?itemName=mine.cpplint)
+- [Clang-Format](https://marketplace.visualstudio.com/items?itemName=xaver.clang-format)
+
+安装扩展后，编辑`settings.json`文件：
+
+```json
+{
+ "[cpp]": {
+ "editor.defaultFormatter": "xaver.clang-format"
+ },
+ "cpplint.cpplintPath": "path/to/cpplint",
+
+ "editor.formatOnSave": true,
+ "python.formatting.provider": "yapf",
+ "python.linting.enabled": true,
+ "python.linting.lintOnSave": true,
+ "clang-format.language.cpp.style": "google",
+ "python.formatting.yapfArgs": ["--style", "{based_on_style: google}"]
+}
+```
+
+根据您的操作系统，用户设置文件位于以下位置：
+
+1. Windows %APPDATA%\Code\User\settings.json
+2. macOS "\$HOME/Library/Application Support/Code/User/settings.json"
+3. Linux "\$HOME/.config/Code/User/settings.json"
+
+配置是在相应的配置文件中指定的。而这些工具会自动查找项目根目录下的配置文件，比如`.pylintrc`。
+
+#### 安装必要工具
+
+最理想的情况是所有贡献者都使用相同版本的代码格式化工具（clang-format 9.0.0和yapf 0.29.0），这样在不同PR中的代码格式化就会完全相同，从而摆脱github pull request冲突。
+
+首先，安装LLVM 9.0，它提供了clang-format 9.0.0版本，LLVM的下载页面如下:
+
+- [LLVM](http://releases.llvm.org/download.html#9.0.0)
+
+ - Ubuntu系统：
+
+ ```sh
+ sudo apt-get install clang-format-9
+ ```
+
+ - Windows系统，下载预编译包并安装。
+
+然后，安装cpplint, pylint和yapf
+
+- Ubuntu或OSX:
+
+ ```
+ $ sudo pip install cpplint
+ $ which cpplint
+ /path/to/cpplint
+
+ $ pip install yapf==0.29.0
+ $ pip install pylint
+ ```
+
+- Windows: 安装Anaconda进行包管理
+
+ ```
+ $ pip install cpplint
+ $ where cpplint
+ C:/path/to/cpplint.exe
+
+ $ pip install yapf==0.29.0
+ $ pip install pylint
+ ```
+
+#### 使用
+
+- 配置后，在编辑源代码文件时，linting会自动启用。错误和警告会在Visual Studio Code `PROBLEMS`面板中列出。
+- 代码格式化可以通过调出Command Palette(Windows中为`Shift+Ctrl+P`，OS X中为`Shift+Command+P`)并输入`Format Document`来完成。
+
+#### 提交
+
+修正格式错误以后就可以提交pull request了。
+
+## 开发环境
+
+推荐使用Visual Studio Code作为编辑器。可以安装Python、C/C++、代码拼写检查器、autoDocstring、vim、远程开发等扩展。这些扩展的参考配置（即`settings.json`）可以在[这里](https://gist.github.com/nudles/3d23cfb6ffb30ca7636c45fe60278c55)查看。
+
+如果更新CPP代码，需要从[源文件](./build.md)重新编译SINGA。建议使用`*-devel Docker`镜像中的原生构建工具或使用`conda build`。
+
+如果要只更新Python代码，您可以安装一次SINGA，然后复制更新后的Python文件来替换Python安装文件夹中的文件。
+
+```shell
+cp python/singa/xx.py <path to conda>/lib/python3.7/site-packages/singa/
+```
+
+## 工作流程
+
+请参阅[git工作流程页面](./git-workflow.md).
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/contribute-docs.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/contribute-docs.md
new file mode 100644
index 0000000..eecc42a
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/contribute-docs.md
@@ -0,0 +1,83 @@
+---
+id: version-3.1.0_Chinese-contribute-docs
+title: How to Contribute to Documentation
+original_id: contribute-docs
+---
+
+<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+文档有两种类型，即markdown文件和API使用参考。本指南介绍了一些工具，并指导如何准备md文件和API注释。
+
+md文件将通过[Docusaurus](https://docusaurus.io/)构建成HTML页面；API注释（来自源代码）将用于使用Sphinx（对应Python）和Doxygen（对应CPP）生成API参考页面。
+
+## Markdown文件
+
+请尽量遵循[Google Documentation style](https://developers.google.com/style)。例如：
+
+1. 删除指令中的"please"。如：'Please click...' VS 'Click...'。
+2. 遵循标准的[大小写规则](https://owl.purdue.edu/owl/general_writing/mechanics/help_with_capitals.html)。
+3. 在说明中用"you"代替"we"。
+4. 使用现在时态，避免使用`will`。
+5. 尽量使用主动语态而不是被动。
+
+此外，为了使文件一致：
+
+1. 句子不宜过长, e.g., 长度<=80
+2. 尽量使用相对路径，假设我们在repo的根目录下，例如，`doc-site/docs`指的是`singa-doc/docs-site/docs`。
+3. 使用背标将命令、路径、类函数和变量亮出来，例如，`Tensor`, `singa-doc/docs-site/docs`。
+4. 为了突出其他术语/概念，使用 _斜体_ or **加粗**
+
+
+本项目使用的[prettier tool](https://prettier.io/)会在我们进行git提交时，根据[配置](https://github.com/apache/singa-doc/blob/master/docs-site/.prettierrc)自动格式化代码。例如，它会将markdown文件中的文本最多折叠成80个字符（注释行除外）。
+
+在介绍一个概念（如`Tensor`类）时，要提供概述（目的和与其他概念的关系）、API和例子，还可以用Google colab来演示其用法。
+
+详细的编辑md文件和建立网站的方法请参考[本页面](https://github.com/apache/singa-doc/tree/master/docs-site)。
+
+## API References
+
+### CPP API
+
+请遵循[Google CPP注释风格](https://google.github.io/styleguide/cppguide.html#Comments).
+
+要生成文档，请从doc文件夹中运行 "doxygen"（推荐使用Doxygen >= 1.8）。
+
+### Python API
+
+请遵循[Google Python DocString风格](http://google.github.io/styleguide/pyguide.html#38-comments-and-docstrings).
+
+## Visual Studio Code (vscode)
+
+如果你使用vscode作为编辑器，我们推荐使用以下插件。
+
+### Docstring Snippet
+
+[autoDocstring](https://marketplace.visualstudio.com/items?itemName=njpwerner.autodocstring)生成函数、类等的docstring，要注意选择使用`google`的docString格式。
+
+### Spell Check
+
+[Code Spell Checker](https://marketplace.visualstudio.com/items?itemName=streetsidesoftware.code-spell-checker)可以用来检查代码的注释，或`.md`和`.rst`文件。
+
+要只对Python代码的注释进行拼写检查，可以在`File - Preferences - User Snippets - python.json`中添加以下代码段：
+
+ "cspell check" : {
+ "prefix": "cspell",
+ "body": [
+ "# Directives for doing spell check only for python and c/cpp comments",
+ "# cSpell:includeRegExp #.* ",
+ "# cSpell:includeRegExp (\"\"\"|''')[^\1]*\1",
+ "# cSpell: CStyleComment",
+ ],
+ "description": "# spell check only for python comments"
+ }
+
+如果要只对Cpp代码的注释进行拼写检查，可以在`File - Preferences - User Snippets - cpp.json`中添加以下代码段：
+
+ "cspell check" : {
+ "prefix": "cspell",
+ "body": [
+ "// Directive for doing spell check only for cpp comments",
+ "// cSpell:includeRegExp CStyleComment",
+ ],
+ "description": "# spell check only for cpp comments"
+ }
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/device.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/device.md
new file mode 100644
index 0000000..ff93723
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/device.md
@@ -0,0 +1,30 @@
+---
+id: version-3.1.0_Chinese-device
+title: Device
+original_id: device
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+Device代表一个有内存和计算单元的硬件设备。所有的[Tensor操作](./tensor)都由常驻设备安排执行的，Tensor内存也由设备的内存管理器管理。因此，需要在Device类中实现内存和执行的优化。
+
+
+## 特定设备
+
+目前，SINGA支持三种设备：
+
+1. CudaGPU：用于运行Cuda代码的Nvidia GPU。
+2. CppCPU：用于运行Cpp代码的CPU。
+3. OpenclGPU：用于运行OpenCL代码的GPU卡。
+
+## 用法示例
+
+下面的代码提供了创建设备的例子：
+
+```python
+from singa import device
+cuda = device.create_cuda_gpu_on(0) # use GPU card of ID 0
+host = device.get_default_device() # get the default host device (a CppCPU)
+ary1 = device.create_cuda_gpus(2) # create 2 devices, starting from ID 0
+ary2 = device.create_cuda_gpus([0,2]) # create 2 devices on ID 0 and 2
+```
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/dist-train.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/dist-train.md
new file mode 100644
index 0000000..6c77a6f
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/dist-train.md
@@ -0,0 +1,379 @@
+---
+id: version-3.1.0_Chinese-dist-train
+title: Distributed Training
+original_id: dist-train
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA支持跨多个GPU的数据并行训练（在单个节点上或跨不同节点）。下图说明了数据并行训练的情况：
+
+
+
+在分布式训练中，每个进程(称为worker)在单个GPU上运行一个训练脚本，每个进程都有一个单独的通信等级，训练数据被分发给各个worker，模型在每个worker上被复制。在每次迭代中，worker从其分区中读取数据的一个mini-batch（例如，256张图像），并运行BackPropagation算法来计算权重的梯度，通过all-reduce（由[NCCL](https://developer.nvidia.com/nccl)提供）进行平均，按照随机梯度下降算法（SGD）进行权重更新。
+
+NCCL的all-reduce操作可以用来减少和同步不同GPU的梯度。假设我们使用4个GPU进行训练，如下图所示。一旦计算出4个GPU的梯度，all-reduce将返回GPU的梯度之和，并使其在每个GPU上可用，然后就可以轻松计算出平均梯度。
+
+
+
+## 使用
+
+SINGA提供了一个名为`DistOpt`（`Opt`的一个子类）的模块，用于分布式训练。它封装了一个普通的SGD优化器，并调用`Communicator`进行梯度同步。下面的例子说明了如何使用`DistOpt`在MNIST数据集上训练一个CNN模型。源代码在[这里](https://github.com/apache/singa/blob/master/examples/cnn/)，或者可以使用[Colab notebook]()。
+
+### 代码示例
+
+1. 定义神经网络模型：
+
+```python
+class CNN(model.Model):
+
+ def __init__(self, num_classes=10, num_channels=1):
+ super(CNN, self).__init__()
+ self.conv1 = layer.Conv2d(num_channels, 20, 5, padding=0, activation="RELU")
+ self.conv2 = layer.Conv2d(20, 50, 5, padding=0, activation="RELU")
+ self.linear1 = layer.Linear(500)
+ self.linear2 = layer.Linear(num_classes)
+ self.pooling1 = layer.MaxPool2d(2, 2, padding=0)
+ self.pooling2 = layer.MaxPool2d(2, 2, padding=0)
+ self.relu = layer.ReLU()
+ self.flatten = layer.Flatten()
+ self.softmax_cross_entropy = layer.SoftMaxCrossEntropy()
+
+ def forward(self, x):
+ y = self.conv1(x)
+ y = self.pooling1(y)
+ y = self.conv2(y)
+ y = self.pooling2(y)
+ y = self.flatten(y)
+ y = self.linear1(y)
+ y = self.relu(y)
+ y = self.linear2(y)
+ return y
+
+ def train_one_batch(self, x, y, dist_option='fp32', spars=0):
+ out = self.forward(x)
+ loss = self.softmax_cross_entropy(out, y)
+
+ # Allow different options for distributed training
+ # See the section "Optimizations for Distributed Training"
+ if dist_option == 'fp32':
+ self.optimizer(loss)
+ elif dist_option == 'fp16':
+ self.optimizer.backward_and_update_half(loss)
+ elif dist_option == 'partialUpdate':
+ self.optimizer.backward_and_partial_update(loss)
+ elif dist_option == 'sparseTopK':
+ self.optimizer.backward_and_sparse_update(loss,
+ topK=True,
+ spars=spars)
+ elif dist_option == 'sparseThreshold':
+ self.optimizer.backward_and_sparse_update(loss,
+ topK=False,
+ spars=spars)
+ return out, loss
+
+# create model
+model = CNN()
+```
+
+2. 创建`DistOpt`实例并将其应用到创建的模型上：
+
+```python
+sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
+sgd = opt.DistOpt(sgd)
+model.set_optimizer(sgd)
+dev = device.create_cuda_gpu_on(sgd.local_rank)
+```
+
+下面是关于代码中一些变量的解释：
+
+(i) `dev`
+
+dev代表`Device`实例，在设备中加载数据并运行CNN模型。
+
+(ii)`local_rank`
+
+Local rank表示当前进程在同一节点中使用的GPU数量。例如，如果你使用的节点有2个GPU，`local_rank=0`表示这个进程使用的是第一个GPU，而`local_rank=1`表示使用的是第二个GPU。使用MPI或多进程，你能够运行相同的训练脚本，唯一的区别`local_rank`的值不同。
+
+(iii)`global_rank`
+
+global中的rank代表了你使用的所有节点中所有进程的全局排名。让我们考虑这样的情况：你有3个节点，每个节点有两个GPU， `global_rank=0`表示使用第1个节点的第1个GPU的进程， `global_rank=2`表示使用第2个节点的第1个GPU的进程， `global_rank=4`表示使用第3个节点的第1个GPU的进程。
+
+3. 加载和分割训练/验证数据：
+
+```python
+def data_partition(dataset_x, dataset_y, global_rank, world_size):
+ data_per_rank = dataset_x.shape[0] // world_size
+ idx_start = global_rank * data_per_rank
+ idx_end = (global_rank + 1) * data_per_rank
+ return dataset_x[idx_start:idx_end], dataset_y[idx_start:idx_end]
+
+train_x, train_y, test_x, test_y = load_dataset()
+train_x, train_y = data_partition(train_x, train_y,
+ sgd.global_rank, sgd.world_size)
+test_x, test_y = data_partition(test_x, test_y,
+ sgd.global_rank, sgd.world_size)
+```
+
+这个`dev`的数据集的一个分区被返回。
+
+
+这里，`world_size`代表你用于分布式训练的所有节点中的进程总数。
+
+4. 初始化并同步所有worker的模型参数:
+
+```python
+#Synchronize the initial parameter
+tx = tensor.Tensor((batch_size, 1, IMG_SIZE, IMG_SIZE), dev, tensor.float32)
+ty = tensor.Tensor((batch_size, num_classes), dev, tensor.int32)
+model.compile([tx], is_train=True, use_graph=graph, sequential=True)
+...
+#Use the same random seed for different ranks
+seed = 0
+dev.SetRandSeed(seed)
+np.random.seed(seed)
+```
+
+5. 运行BackPropagation和分布式SGD
+
+```python
+for epoch in range(max_epoch):
+ for b in range(num_train_batch):
+ x = train_x[idx[b * batch_size: (b + 1) * batch_size]]
+ y = train_y[idx[b * batch_size: (b + 1) * batch_size]]
+ tx.copy_from_numpy(x)
+ ty.copy_from_numpy(y)
+ # Train the model
+ out, loss = model(tx, ty)
+```
+
+### 执行命令
+
+有两种方式可以启动训练，MPI或Python multiprocessing。
+
+#### Python multiprocessing
+
+它可以在一个节点上使用多个GPU，其中，每个GPU是一个worker。
+
+1. 将上述训练用的代码打包进一个函数：
+
+```python
+def train_mnist_cnn(nccl_id=None, local_rank=None, world_size=None):
+ ...
+```
+
+2. 创建`mnist_multiprocess.py`。
+
+```python
+if __name__ == '__main__':
+ # Generate a NCCL ID to be used for collective communication
+ nccl_id = singa.NcclIdHolder()
+
+ # Define the number of GPUs to be used in the training process
+ world_size = int(sys.argv[1])
+
+ # Define and launch the multi-processing
+ import multiprocessing
+ process = []
+ for local_rank in range(0, world_size):
+ process.append(multiprocessing.Process(target=train_mnist_cnn,
+ args=(nccl_id, local_rank, world_size)))
+
+ for p in process:
+ p.start()
+```
+
+下面是关于上面所创建的变量的一些说明：
+
+(i) `nccl_id`
+
+需要注意的是，我们在这里需要生成一个NCCL ID，用于集体通信，然后将其传递给所有进程。NCCL ID就像一个门票，只有拥有这个ID的进程才能加入到all-reduce操作中。(如果我们接下来使用MPI，NCCL ID的传递就没有必要了，因为在我们的代码中，这个ID会由MPI自动广播。)
+
+(ii) `world_size`
+
+world_size是您想用于训练的GPU数量。
+
+(iii) `local_rank`
+
+local_rank决定分布式训练的本地顺序，以及在训练过程中使用哪个gpu。在上面的代码中，我们使用for循环来运行训练函数，其中参数local_rank从0迭代到world_size。在这种情况下，不同的进程可以使用不同的GPU进行训练。
+
+创建`DistOpt`实例的参数应按照如下方式更新：
+
+```python
+sgd = opt.DistOpt(sgd, nccl_id=nccl_id, local_rank=local_rank, world_size=world_size)
+```
+
+3. 运行`mnist_multiprocess.py`：
+
+```sh
+python mnist_multiprocess.py 2
+```
+
+与单GPU训练相比，它最主要的意义是速度提升：
+
+```
+Starting Epoch 0:
+Training loss = 408.909790, training accuracy = 0.880475
+Evaluation accuracy = 0.956430
+Starting Epoch 1:
+Training loss = 102.396790, training accuracy = 0.967415
+Evaluation accuracy = 0.977564
+Starting Epoch 2:
+Training loss = 69.217010, training accuracy = 0.977915
+Evaluation accuracy = 0.981370
+Starting Epoch 3:
+Training loss = 54.248390, training accuracy = 0.982823
+Evaluation accuracy = 0.984075
+Starting Epoch 4:
+Training loss = 45.213406, training accuracy = 0.985560
+Evaluation accuracy = 0.985276
+Starting Epoch 5:
+Training loss = 38.868435, training accuracy = 0.987764
+Evaluation accuracy = 0.986278
+Starting Epoch 6:
+Training loss = 34.078186, training accuracy = 0.989149
+Evaluation accuracy = 0.987881
+Starting Epoch 7:
+Training loss = 30.138697, training accuracy = 0.990451
+Evaluation accuracy = 0.988181
+Starting Epoch 8:
+Training loss = 26.854443, training accuracy = 0.991520
+Evaluation accuracy = 0.988682
+Starting Epoch 9:
+Training loss = 24.039650, training accuracy = 0.992405
+Evaluation accuracy = 0.989083
+```
+
+#### MPI
+
+只要有多个GPU，MPI既适用于单节点，也适用于多节点。
+
+1. 创建`mnist_dist.py`。
+
+```python
+if __name__ == '__main__':
+ train_mnist_cnn()
+```
+
+2. 为MPI生成一个hostfile，例如下面的hostfile在一个节点上使用了2个进程（即2个GPU）：
+
+```txt
+localhost:2
+```
+
+3. 通过`mpiexec`启动训练：
+
+```sh
+mpiexec --hostfile host_file python mnist_dist.py
+```
+
+与单GPU训练相比，它同样可以带来速度的提升：
+```
+Starting Epoch 0:
+Training loss = 383.969543, training accuracy = 0.886402
+Evaluation accuracy = 0.954327
+Starting Epoch 1:
+Training loss = 97.531479, training accuracy = 0.969451
+Evaluation accuracy = 0.977163
+Starting Epoch 2:
+Training loss = 67.166870, training accuracy = 0.978516
+Evaluation accuracy = 0.980769
+Starting Epoch 3:
+Training loss = 53.369656, training accuracy = 0.983040
+Evaluation accuracy = 0.983974
+Starting Epoch 4:
+Training loss = 45.100403, training accuracy = 0.985777
+Evaluation accuracy = 0.986078
+Starting Epoch 5:
+Training loss = 39.330826, training accuracy = 0.987447
+Evaluation accuracy = 0.987179
+Starting Epoch 6:
+Training loss = 34.655270, training accuracy = 0.988799
+Evaluation accuracy = 0.987780
+Starting Epoch 7:
+Training loss = 30.749735, training accuracy = 0.989984
+Evaluation accuracy = 0.988281
+Starting Epoch 8:
+Training loss = 27.422146, training accuracy = 0.991319
+Evaluation accuracy = 0.988582
+Starting Epoch 9:
+Training loss = 24.548153, training accuracy = 0.992171
+Evaluation accuracy = 0.988682
+```
+
+## 针对分布式训练的优化
+
+SINGA为分布式训练提供了多种优化策略，以降低模块间的通信成本。参考`DistOpt`的API，了解每种策略的配置。
+
+
+当我们使用`model.Model`建立模型时，我们需要在`training_one_batch`方法中启用分布式训练的选项，请参考本页顶部的示例代码。我们也可以直接复制这些选项的代码，然后在其他模型中使用。
+
+有了定义的选项，我们可以在使用`model(tx, ty, dist_option, spars)`开始训练时，设置对应的参数`dist_option`和`spars`。
+
+### 不采取优化手段
+
+```python
+out, loss = model(tx, ty)
+```
+
+`loss`是损失函数的输出张量，例如分类任务中的交叉熵。
+
+### 半精度梯度（Half-precision Gradients）
+
+```python
+out, loss = model(tx, ty, dist_option = 'fp16')
+```
+
+在调用all-reduce之前，它将每个梯度值转换为16-bit表示（即半精度）。
+
+### 部分同步（Partial Synchronization）
+
+```python
+out, loss = model(tx, ty, dist_option = 'partialUpdate')
+```
+
+在每一次迭代中，每个rank都做本地SGD更新。然后，只对一个部分的参数进行平均同步，从而节省了通信成本。分块大小是在创建`DistOpt`实例时配置的。
+
+### 梯度稀疏化（Gradient Sparsification）
+
+该策略应用稀疏化方案来选择梯度的子集进行all-reduce，有两种方式：
+
+- 选择前k大的元素，spars是被选择的元素的一部分（比例在0 - 1之间）。
+
+```python
+out, loss = model(tx, ty, dist_option = 'sparseTopK', spars = spars)
+```
+
+- 所有绝对值大于预定义阈值的梯度都会被选中。
+
+```python
+out, loss = model(tx, ty, dist_option = 'sparseThreshold', spars = spars)
+```
+超参数在创建`DistOpt`实例时配置。
+
+## 实现
+
+本节主要是让开发者了解分布训练模块的代码是如何实现的。
+
+### NCCL communicator的C接口
+
+
+首先，通信层是用C语言[communicator.cc](https://github.com/apache/singa/blob/master/src/io/communicator.cc)编写的，它调用用NCCL库进行集体通信。
+
+communicator有两个构造器，一个是MPI的，另一个是Multiprocess的。
+
+(i) MPI构造器
+
+构造器首先先获取全局rank和world_size，计算出本地rank，然后由rank 0生成NCCL ID并广播给每个rank。之后，它调用setup函数来初始化NCCL communicator、cuda流和缓冲区。
+
+(ii) Python multiprocess构造器
+
+构造器首先从输入参数中获取rank、world_size和NCCL ID。之后，调用setup函数来初始化NCCL communicator、cuda流和缓冲区。
+
+在初始化之后，它提供了all-reduce功能来同步模型参数或梯度。例如，synch接收一个输入张量，通过NCCL例程进行all-reduce，在我们调用synch之后，需要调用wait函数来等待all-reduce操作的完成。
+
+### DistOpt的Python接口
+
+然后，python接口提供了一个[DistOpt](https://github.com/apache/singa/blob/master/python/singa/opt.py)类来封装一个[optimizer](https://github.com/apache/singa/blob/master/python/singa/opt.py)对象，以执行基于MPI或Multiprocess的分布式训练。在初始化过程中，它创建了一个NCCL communicator对象（来自于上面小节提到的C接口），然后，`DistOpt`中的每一次all-reduce操作都会用到这个communicator对象。
+
+在MPI或Multiprocess中，每个进程都有一个独立的rank，它给出了各个进程使用的GPU的信息。训练数据是被拆分的，因此每个进程可以根据一部分训练数据来评估子梯度。一旦每个进程的子梯度被计算出来，就可以将所有进程计算出的子梯度做all-reduce，得到总体随机梯度。
\ No newline at end of file
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/download.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/download.md
new file mode 100644
index 0000000..45eb152
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/download.md
@@ -0,0 +1,159 @@
+---
+id: version-3.1.0_Chinese-downloads
+title: Download SINGA
+original_id: downloads
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## Verify
+
+要验证下载的tar.gz文件，下载[KEYS](https://www.apache.org/dist/singa/KEYS)和ASC文件，然后执行以下命令:
+
+```shell
+% gpg --import KEYS
+% gpg --verify downloaded_file.asc downloaded_file
+```
+
+你也可以检查SHA512或MD5值判断下载是否完成。
+
+## V3.0.0 (18 April 2020):
+
+- [Apache SINGA 3.0.0](http://www.apache.org/dyn/closer.cgi/singa/3.0.0/apache-singa-3.0.0.tar.gz)
+ [\[SHA512\]](https://www.apache.org/dist/singa/3.0.0/apache-singa-3.0.0.tar.gz.sha512)
+ [\[ASC\]](https://www.apache.org/dist/singa/3.0.0/apache-singa-3.0.0.tar.gz.asc)
+- [Release Notes 3.0.0](releases/RELEASE_NOTES_3.0.0)
+- 新特性及重要更新：
+ - 增强了ONNX。在SINGA中测试了多种ONNX模型。
+ - 使用 MPI 和 NCCL Communication进行分布式训练，通过梯度稀疏化和压缩以及分块传输进行了优化。
+ - 计算图的构建，利用图优化了速度和内存。
+ - 新的文档网站（singa.apache.org）和API参考网站（apache-singa.rtfd.io）。
+ - 使用CI实现代码质量检查。
+ - 将MKLDNN替换为DNNL。
+ - 更新Tensor API以支持广播操作。
+ - 实现了支持ONNX模型的新的autograd操作符。
+
+## 孵化版本（incubating） v2.0.0 (20 April 2019):
+
+- [Apache SINGA 2.0.0 (incubating)](http://www.apache.org/dyn/closer.cgi/incubator/singa/2.0.0/apache-singa-incubating-2.0.0.tar.gz)
+ [\[SHA512\]](https://www.apache.org/dist/incubator/singa/2.0.0/apache-singa-incubating-2.0.0.tar.gz.sha512)
+ [\[ASC\]](https://www.apache.org/dist/incubator/singa/2.0.0/apache-singa-incubating-2.0.0.tar.gz.asc)
+- [Release Notes 2.0.0 (incubating)](releases/RELEASE_NOTES_2.0.0.html)
+- 新特性及重要更新：
+ - 增强了autograd功能(适用于卷积网络和循环网络)。
+ - 支持ONNX。
+ - 通过英特尔MKL DNN lib改进CPP操作。
+ - 实现tensor广播。
+ - 在Apache用户名下移动Docker镜像。
+ - 更新conda-build配置中依赖的lib版本。
+
+## 孵化版本（incubating） v1.2.0 (6 June 2018):
+
+- [Apache SINGA 1.2.0 (incubating)](https://archive.apache.org/dist/incubator/singa/1.2.0/apache-singa-incubating-1.2.0.tar.gz)
+ [\[SHA512\]](https://archive.apache.org/dist/incubator/singa/1.2.0/apache-singa-incubating-1.2.0.tar.gz.sha512)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/1.2.0/apache-singa-incubating-1.2.0.tar.gz.asc)
+- [Release Notes 1.2.0 (incubating)](releases/RELEASE_NOTES_1.2.0.html)
+- 新特性及重要更新：
+ - 实现了autograd（目前支持MLP模式）。
+ - 升级PySinga以支持Python 3
+ - 改进Tensor类的stride范围。
+ - 将cuDNN从V5升级到V7。
+ - 增加VGG、Inception V4、ResNet和DenseNet进行ImageNet分类。
+ - 为conda包创建别名
+ - 完整的中文文档
+ - 添加在Windows上运行Singa的说明
+ - 更新编译，CI
+ - 修复一些错误
+
+## 孵化版本（incubating） v1.1.0 (12 February 2017):
+
+- [Apache SINGA 1.1.0 (incubating)](https://archive.apache.org/dist/incubator/singa/1.1.0/apache-singa-incubating-1.1.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/1.1.0/apache-singa-incubating-1.1.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/1.1.0/apache-singa-incubating-1.1.0.tar.gz.asc)
+- [Release Notes 1.1.0 (incubating)](releases/RELEASE_NOTES_1.1.0.html)
+- 新特性和重要更新：
+ - 创建Docker镜像(CPU和GPU版本)
+ - 为SINGA创建Amazon AMI（CPU版）。
+ - 集成Jenkins，自动生成Wheel和Debian包（用于安装），并更新网站。
+ - 增强FeedFowardNet，例如，多输入和用于调试的verbose模式。
+ - 添加Concat和Slice层。
+ - 优化CrossEntropyLoss以接受带有多个标签的实例。
+ - 添加包含图像增强方法的image_tool.py。
+ - 支持通过快照API加载和保存模型。
+ - 在Windows上编译SINGA源代码。
+ - 将必要依赖库与SINGA代码一起编译。
+ - 为SINGA启用Java binding（基本）。
+ - 在checkpoint文件中添加版本ID。
+ - 增加Rafiki工具包以提供RESTFul APIs。
+ - 添加了从Caffe预训练的例子，包括GoogleNet。
+
+## 孵化版本（incubating） v1.0.0 (8 September 2016):
+
+- [Apache SINGA 1.0.0 (incubating)](https://archive.apache.org/dist/incubator/singa/1.0.0/apache-singa-incubating-1.0.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/1.0.0/apache-singa-incubating-1.0.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/1.0.0/apache-singa-incubating-1.0.0.tar.gz.asc)
+- [Release Notes 1.0.0 (incubating)](releases/RELEASE_NOTES_1.0.0.html)
+- 新特性和重要更新：
+ - 创建Tensor概念，用于支持更多的机器学习模型。
+ - 创建Device概念，用于运行在不同的硬件设备上，包括CPU，(Nvidia/AMD) GPU 和 FPGA (将在以后的版本中测试)。
+ - 用 cmake 取代 GNU autotool 进行编译。
+ - 支持 Mac OS。
+ - 改进Python binding，包括安装和编程。
+ - 更多的深度学习模型，包括VGG和ResNet。
+ - 更多的IO类用于读取/写入文件和编码/解码数据。
+ - 直接基于Socket的新网络通信组件。
+ - Cudnn V5，包含Dropout和RNN层。
+ - 将网页构建工具从maven更换为Sphinx。
+ - 整合Travis-CI。
+
+## 孵化版本（incubating） v0.3.0 (20 April 2016):
+
+- [Apache SINGA 0.3.0 (incubating)](https://archive.apache.org/dist/incubator/singa/0.3.0/apache-singa-incubating-0.3.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/0.3.0/apache-singa-incubating-0.3.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/0.3.0/apache-singa-incubating-0.3.0.tar.gz.asc)
+- [Release Notes 0.3.0 (incubating)](releases/RELEASE_NOTES_0.3.0.html)
+- 新特性和重要更新：
+ - 在GPU集群上进行训练，可以在GPU集群上进行深度学习模型的训练。
+ - 改进Python wrapper简化配置工作，包括神经网络和SGD算法。
+ - 新增SGD更新器，包括Adam、AdaDelta和AdaMax。
+ - 安装时减少了单节点训练的依赖库。
+ - 使用CPU和GPU进行异构训练。
+ - 支持 cuDNN V4。
+ - 数据预取。
+ - 修复一些bug。
+
+## 孵化版本（incubating） v0.2.0 (14 January 2016):
+
+- [Apache SINGA 0.2.0 (incubating)](https://archive.apache.org/dist/incubator/singa/0.2.0/apache-singa-incubating-0.2.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/0.2.0/apache-singa-incubating-0.2.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/0.2.0/apache-singa-incubating-0.2.0.tar.gz.asc)
+- [Release Notes 0.2.0 (incubating)](releases/RELEASE_NOTES_0.2.0.html)
+- 新特性和重要更新：
+ - 在GPU上进行训练，可以在一个节点上用多个GPU卡训练复杂的模型。
+ - 混合神经网分区支持数据和模型同时并行。
+ - Python wrapper简化配置，包括神经网络和SGD算法。
+ - 实现了RNN模型和BPTT算法，支持基于RNN模型的应用，如GRU。
+ - 云软件集成，包括Mesos、Docker和HDFS。
+ - 可视化神经网结构和层信息，以便优化调试。
+ - 针对Blobs和原始数据指针的线性代数函数和随机函数。
+ - 添加新的层，包括SoftmaxLayer、ArgSortLayer、DummyLayer、RNN层和cuDNN层。
+ - 更新Layer类以携带多个数据/梯度Blobs。
+ - 通过加载之前训练的模型参数，提取特征并测试新数据的性能。
+ - 为IO操作添加Store类。
+
+## Incubating v0.1.0 (8 October 2015):
+
+- [Apache SINGA 0.1.0 (incubating)](https://archive.apache.org/dist/incubator/singa/apache-singa-incubating-0.1.0.tar.gz)
+ [\[MD5\]](https://archive.apache.org/dist/incubator/singa/apache-singa-incubating-0.1.0.tar.gz.md5)
+ [\[ASC\]](https://archive.apache.org/dist/incubator/singa/apache-singa-incubating-0.1.0.tar.gz.asc)
+- [Amazon EC2 image](https://console.aws.amazon.com/ec2/v2/home?region=ap-southeast-1#LaunchInstanceWizard:ami=ami-b41001e6)
+- [Release Notes 0.1.0 (incubating)](releases/RELEASE_NOTES_0.1.0.html)
+- 新特性和重要更新：
+ - 允许使用GNU构建工具进行安装。
+ - 实现用zookeeper进行工作管理的脚本。
+ - 实现基于NeuralNet和Layer的编程模型。
+ - 实现基于Worker、Server和Stub的系统架构。
+ - 训练模型来自三种不同的模型类别，即前馈模型、能量模型和RNN模型。
+ - 使用CPU的同步和异步分布式训练框架。
+ - checkpoint文件生成和恢复。
+ - 使用gtest进行单元测试。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/examples.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/examples.md
new file mode 100644
index 0000000..f50d700
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/examples.md
@@ -0,0 +1,63 @@
+---
+id: version-3.1.0_Chinese-examples
+title: Examples
+original_id: examples
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+本页列出了一些使用SINGA的深度学习任务示例。源代码维护在 [Github](https://github.com/apache/singa/tree/master/examples) 上的 SINGA repo 内。对于使用SINGA Python API在CPU或单GPU上运行的例子，它们也可以在[Google Colab](https://colab.research.google.com/)上获得。你可以直接在谷歌云上运行它们，而无需在本地设置环境。下面给出了每个例子的链接。
+
+## 图像分类
+
+| 网络模型 | 数据集 | 链接 |
+| ----------- | --------------------------------- | ------------------------------------------------------------------------------------------------------- |
+| Simple CNN | MNIST, CIFAR10, CIFAR100 | [Colab](https://colab.research.google.com/drive/1fbGUs1AsoX6bU5F745RwQpohP4bHTktq) |
+| AlexNet | ImageNet | [Cpp]() |
+| VGG | ImageNet | [Cpp](), [Python](), [Colab](https://colab.research.google.com/drive/14kxgRKtbjPCKKsDJVNi3AvTev81Gp_Ds) |
+| XceptionNet | MNIST, CIFAR10, CIFAR100 | [Python]() |
+| ResNet | MNIST, CIFAR10, CIFAR100, CIFAR10 | [Python](), [Colab](https://colab.research.google.com/drive/1u1RYefSsVbiP4I-5wiBKHjsT9L0FxLm9) |
+| MobileNet | ImageNet | [Colab](https://colab.research.google.com/drive/1HsixqJMIpKyEPhkbB8jy7NwNEFEAUWAf) |
+
+## 目标检测
+
+| 网络模型 | 数据集 | 链接 |
+| ----------- | ---------- | ---------------------------------------------------------------------------------- |
+| Tiny YOLOv2 | Pascal VOC | [Colab](https://colab.research.google.com/drive/11V4I6cRjIJNUv5ZGsEGwqHuoQEie6b1T) |
+
+## 面部及表情识别
+
+| 模型 | 数据集 | 链接 |
+| --------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------- |
+| ArcFace | Refined MS-Celeb-1M | [Colab](https://colab.research.google.com/drive/1qanaqUKGIDtifdzEzJOHjEj4kYzA9uJC) |
+| Emotion FerPlus | [Facial Expression Recognition Challenge](https://www.kaggle.com/c/challenges-in-representation-learning-facial-expression-recognition-challenge/data) | [Colab](https://colab.research.google.com/drive/1XHtBQGRhe58PDi4LGYJzYueWBeWbO23r) |
+
+## 图像生成
+
+| 模型 | 数据集 | 链接 |
+| ----- | ------- | ---------------------------------------------------------------------------------- |
+| GAN | MNIST | [Colab](https://colab.research.google.com/drive/1f86MNDW47DJqHoIqWD1tOxcyx2MWys8L) |
+| LSGAN | MNIST | [Colab](https://colab.research.google.com/drive/1C6jNRf28vnFOI9JVM4lpkJPqxsnhxdol) |
+
+## 机器理解
+
+| 模型 | 数据集 | 链接 |
+| ---------- | ------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- |
+| Bert-Squad | [SQuAD v1.1](https://rajpurkar.github.io/SQuAD-explorer/explore/1.1/dev/) | [Colab](https://colab.research.google.com/drive/1kud-lUPjS_u-TkDAzihBTw0Vqr0FjCE-) |
+
+## 文字识别
+
+| 模型 | 数据集 | 链接 |
+| ----------- | ------- | ---------- |
+| Simple LSTM | IMDB | [python]() |
+
+## 文本排序
+
+| 模型 | 数据集 | 链接 |
+| ------ | ----------- | ---------- |
+| BiLSTM | InsuranceQA | [python]() |
+
+## 其他
+
+- MNIST数据集的有限玻尔兹曼机, [source](),
+ [Colab](https://colab.research.google.com/drive/19996noGu9JyHHkVmp4edBGu7PJSRQKsd).
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/git-workflow.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/git-workflow.md
new file mode 100644
index 0000000..24910da
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/git-workflow.md
@@ -0,0 +1,86 @@
+---
+id: version-3.1.0_Chinese-git-workflow
+title: Git Workflow
+original_id: git-workflow
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 对于开发者
+
+1. 将[SINGA Github repository](https://github.com/apache/singa) fork到你自己的Github账户。
+
+2. 从你自己的git仓库中clone **repo** (short for repository):
+
+ ```shell
+ git clone https://github.com/<Github account>/singa.git
+ git remote add upstream https://github.com/apache/singa.git
+ ```
+
+3. 创建一个新的分支（例如 `feature-foo` 或 `fixbug-foo`），在这个分支上工作并提交你的代码:
+
+ ```shell
+ git checkout dev
+ git checkout -b feature-foo
+ # write your code
+ git add <created/updated files>
+ git commit
+ ```
+
+ commit信息应包括：
+
+ - 一个概括性的标题。
+ - 详细的描述。如果提交是为了修复bug，描述中最好包括问题的简短复现；如果是新功能，可以描述新功能的动机/目的。
+
+ 如果您的分支有很多小的commit，您需要通过:
+
+ ```shell
+ git rebase -i <commit id>
+ ```
+ 你可以[压制和重写](https://help.github.com/en/articles/about-git-rebase)提交的内容。
+
+4. 当你在写代码的时候，SINGA的`dev`分支可能已经被别人更新了；在这种情况下，你需要拉取最新的`dev`分支：
+
+ ```shell
+ git checkout dev
+ git pull upstream dev:dev
+ ```
+
+5. 将`feature-foo` [rebase](https://git-scm.com/book/en/v2/Git-Branching-Rebasing)到`dev`分支上，并将提交的内容推送到自己的Github账户（你刚刚创建的新分支），rebase操作是为了清理提交历史。提交当前工作后，应执行以下 git 指令：
+
+ ```shell
+ git checkout feature-foo
+ git rebase dev
+ git push origin feature-foo:feature-foo
+ ```
+
+ Rebase命令的[操作步骤](https://git-scm.com/book/en/v2/Git-Branching-Rebasing)如下: "这个操作的工作原理是进入到两个分支（你所在的分支和你要rebase的分支）的共同来源 -> 获取你所在分支的每次commit所引入的差异 -> 将这些差异保存到临时文件中 -> 将当前分支重置为与你要rebase的分支相同的commit -> 最后依次修改每个差异。"
+
+ 因此，执行后，你还是在特性分支上，但你自己的提交ID/hash被改变了，因为diffs是在rebase时提交的；而且你的分支现在有来自`dev`分支和你自己分支的最新代码。
+
+6. 在 Github 网站上创建一个针对 apache/singa `dev`分支的pull request（PR）。如果你想通知其他在相同文件上工作过的贡献者，你可以在Github上找到文件，然后点击 "Blame"，就可以看到最后修改代码的人的逐行注释。然后，你可以在PR描述中加上@username，就可以立即ping到他们。请说明该贡献是你的原创作品，并且你在项目的开源许可下将该作品授权给项目。你的新分支的进一步提交（例如，bug修复）将由Github自动添加到这个pull request中。
+
+7. 接下来等待committer审核PR。在这段时间里，SINGA的`dev`可能已经被其他人更新了，这时你需要[合并](https://docs.fast.ai/dev/git.html#how-to-keep-your-feature-branch-up-to-date)最新的`dev`来解决冲突。有些人将PR重新[rebase到最新的dev](https://github.com/edx/edx-platform/wiki/How-to-Rebase-a-Pull-Request)上，而不是合并。但是，如果其他开发者获取这个PR来添加新的功能，然后再发送PR，那么rebase操作会在未来的PR中引入**重复的提交**（不同的哈希）。关于何时避免使用rebase的细节，请参见[The Golden Rule of Rebasing](https://www.atlassian.com/git/tutorials/merging-vs-rebasing)。另一种简单的更新PR的方法（修复冲突或提交错误）是，从Apache SINGAS repo的最新开发分支中checkout出一个新的分支，复制并粘贴更新/添加的代码，然后commit并发送一个新的PR。
+
+## 对于Committers
+
+Committer可以将PR合并到上游 repo 的 dev 分支。在合并每一个PR之前，提交者应该做到：
+
+- 检查commit信息(内容和格式)
+- 检查对现有代码的修改，API的变化应该被记录下来
+- 检查Travis测试结果，检查代码/文档格式和单元测试。
+
+合并PR的方式有两种:
+
+- 在Github上，按照[说明](https://gitbox.apache.org/setup/)将你的Apache账户与Github账户链接，之后你就可以直接在GitHub上合并PR了。
+- 通过命令行合并pull request到https://github.com/apache/singa/pull/xxx，应执行以下指令：
+
+ ```shell
+ git clone https://github.com/apache/singa.git
+ git remote add asf https://gitbox.apache.org/repos/asf/singa.git
+ git fetch origin pull/xxx/head:prxxx
+ git checkout dev
+ git merge --no-ff prxxx
+ git push asf dev:dev
+ ```
+ 不要使用rebase来合并PR，并禁用fast forward。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/graph.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/graph.md
new file mode 100644
index 0000000..a999a5a
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/graph.md
@@ -0,0 +1,465 @@
+---
+id: version-3.1.0_Chinese-graph
+title: Model
+original_id: graph
+---
+
+<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+神经网络中的前向和反向传播可以用一组操作来表示，比如卷积和池化。每个操作都需要一些输入的[tensors](./tensor)，并应用一个[operator](./autograd)来生成输出的张量。通过将每个运算符表示为一个节点，将每个张量表示为一条边，所有的运算就形成了一个计算图。有了计算图，可以通过调度运算的执行和内存的智能分配/释放来进行速度和内存优化。在SINGA中，用户只需要使用[Model](https://github.com/apache/singa/blob/master/python/singa/model.py) API定义神经网络模型，计算图则会在C++后台自动构建和优化。
+
+
+这样，一方面，用户使用[Model](./graph) API按照PyTorch那样的命令式编程风格实现网络。而与PyTorch在每次迭代中重新创建操作不同的是，SINGA在第一次迭代后就会缓冲操作，隐式地创建计算图（当该功能被启用时）。因此，另一方面，SINGA的计算图与使用声明式编程的库（如TensorFlow）创建的计算图类似，因而它可以享受在图上进行的优化。
+
+## 样例
+
+下面的代码说明了`Model`API的用法：
+
+1. 将新模型实现为Model类的子类：
+
+```Python
+class CNN(model.Model):
+
+ def __init__(self, num_classes=10, num_channels=1):
+ super(CNN, self).__init__()
+ self.conv1 = layer.Conv2d(num_channels, 20, 5, padding=0, activation="RELU")
+ self.conv2 = layer.Conv2d(20, 50, 5, padding=0, activation="RELU")
+ self.linear1 = layer.Linear(500)
+ self.linear2 = layer.Linear(num_classes)
+ self.pooling1 = layer.MaxPool2d(2, 2, padding=0)
+ self.pooling2 = layer.MaxPool2d(2, 2, padding=0)
+ self.relu = layer.ReLU()
+ self.flatten = layer.Flatten()
+ self.softmax_cross_entropy = layer.SoftMaxCrossEntropy()
+
+ def forward(self, x):
+ y = self.conv1(x)
+ y = self.pooling1(y)
+ y = self.conv2(y)
+ y = self.pooling2(y)
+ y = self.flatten(y)
+ y = self.linear1(y)
+ y = self.relu(y)
+ y = self.linear2(y)
+ return y
+
+ def train_one_batch(self, x, y):
+ out = self.forward(x)
+ loss = self.softmax_cross_entropy(out, y)
+ self.optimizer(loss)
+ return out, loss
+```
+
+2. 创建model、optimizer、device等的实例。编译模型：
+
+```python
+model = CNN()
+
+# initialize optimizer and attach it to the model
+sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
+model.set_optimizer(sgd)
+
+# initialize device
+dev = device.create_cuda_gpu()
+
+# input and target placeholders for the model
+tx = tensor.Tensor((batch_size, 1, IMG_SIZE, IMG_SIZE), dev, tensor.float32)
+ty = tensor.Tensor((batch_size, num_classes), dev, tensor.int32)
+
+# compile the model before training
+model.compile([tx], is_train=True, use_graph=True, sequential=False)
+```
+
+3. 迭代训练：
+
+```python
+for b in range(num_train_batch):
+ # generate the next mini-batch
+ x, y = ...
+
+ # Copy the data into input tensors
+ tx.copy_from_numpy(x)
+ ty.copy_from_numpy(y)
+
+ # Training with one batch
+ out, loss = model(tx, ty)
+```
+
+这个例子的Google Colab notebook可以在[这里](https://colab.research.google.com/drive/1fbGUs1AsoX6bU5F745RwQpohP4bHTktq)找到。
+
+
+更多例子：
+
+- [MLP](https://github.com/apache/singa/blob/master/examples/mlp/model.py)
+- [CNN](https://github.com/apache/singa/blob/master/examples/cnn/model/cnn.py)
+- [ResNet](https://github.com/apache/singa/blob/master/examples/cnn/model/resnet.py)
+
+## 实现
+
+### 图的构建
+
+SINGA分三步构建计算图：
+
+1. 将操作保存在缓冲区。
+2. 分析操作的依赖性。
+3. 根据依赖关系创建节点和边。
+
+以MLP模型的dense层的矩阵乘法运算为例，该操作会在[MLP model](https://github.com/apache/singa/blob/master/examples/mlp/model.py)的前向函数中被调用：
+
+```python
+class MLP(model.Model):
+
+ def __init__(self, data_size=10, perceptron_size=100, num_classes=10):
+ super(MLP, self).__init__()
+ self.linear1 = layer.Linear(perceptron_size)
+ ...
+
+ def forward(self, inputs):
+ y = self.linear1(inputs)
+ ...
+```
+
+`线性`层由`mutmul`运算符组成，`autograd`通过SWIG调用CPP中提供的`Mult`函数来实现`matmul`运算符。
+
+```python
+# implementation of matmul()
+singa.Mult(inputs, w)
+```
+
+At the backend, the `Mult` function is implemented by calling `GEMV` a CBLAS
+function. Instead of calling `GEMV` directly, `Mult` submits `GEMV` and the
+arguments to the device as follows,
+在后端，`Mult`函数是通过调用`GEMV`一个CBLAS函数来实现的。但`Mult`没有直接调用`GEMV`，而是将`GEMV`和参数提交给设备，具体如下。
+
+```c++
+// implementation of Mult()
+C->device()->Exec(
+ [a, A, b, B, CRef](Context *ctx) mutable {
+ GEMV<DType, Lang>(a, A, B, b, &CRef, ctx);
+ },
+ read_blocks, {C->block()});
+```
+
+`Device`的`Exec`函数对函数及其参数进行缓冲。此外，它还拥有这个函数要读写的块的信息（块是指张量的内存块）。
+
+一旦`Model.forward()`被执行一次，所有的操作就会被`Device`缓冲。接下来，对所有操作的读写信息进行分析，用来建立计算图。例如，如果一个块`b`被一个操作O1写入，之后又被另一个操作O2读出，我们就会知道O2依赖于O1并且有一条从A到B的有向边，它代表了块`b`（或其张量）。之后我们就构建了一个有向无环图，如下图所示。该图会构建一次。
+
+
+
+<br/>**Figure 1 - MLP例子的计算图**
+
+### 优化
+
+目前，基于计算图进行了以下优化：
+
+**惰性分配** 当创建张量/块时，设备不会立即为它们分配内存。相反，是在第一次访问块时，才会分配内存。
+
+**自动回收** 每个张量/块的参考计数是根据图计算出来的。在执行操作之前，参考计数是读取这个块的操作次数。在执行过程中，一旦执行了一个操作，每一个输入块的参考数就会减少1，如果一个块的参考数达到了0，就意味着这个块在剩下的操作中不会再被读取。因此，它的内存可以被安全释放。此外，SINGA还会跟踪图外的块的使用情况。如果一个块被Python代码使用（而不是被autograd操作符使用），它将不会被回收。
+
+**内存共享** SINGA使用内存池，如[CnMem](https://github.com/NVIDIA/cnmem)来管理CUDA内存。有了自动回收和内存池，SINGA就可以在张量之间共享内存。考虑两个操作`c=a+b`和`d=2xc`。在执行第二个操作之前，根据惰性分配原则，应该分配d的内存。假设`a`在其余操作中没有使用。根据自动回收，`a`的块将在第一次操作后被释放。因此，SINGA会向CUDA流提交四个操作：加法、释放`a`、分配`b`和乘法。这样，内存池就可以将`a`释放的内存与`b`共享，而不是要求GPU为`b`做真正的malloc。
+
+其他的优化技术，如来自编译器的优化技术，如常见的子表达式消除和不同CUDA流上的并行化操作也可以应用。
+
+## 新的操作符
+
+`autograd`模块中定义的每个运算符都实现了两个功能：前向和反向，通过在后台调用运算符来实现。如果要在`autograd`中添加一个新的运算符，需要在后台添加多个运算符。
+
+以[Conv2d](https://github.com/apache/singa/blob/master/python/singa/autograd.py)运算符为例，在Python端，根据设备类型，从后台调用运算符来实现前向和反向功能：
+
+```python
+class _Conv2d(Operation):
+
+ def forward(self, x, W, b=None):
+ ......
+ if training:
+ if self.handle.bias_term:
+ self.inputs = (x, W, b) # record x, W, b
+ else:
+ self.inputs = (x, W)
+
+ if (type(self.handle) != singa.ConvHandle):
+ return singa.GpuConvForward(x, W, b, self.handle)
+ else:
+ return singa.CpuConvForward(x, W, b, self.handle)
+
+ def backward(self, dy):
+ if (type(self.handle) != singa.ConvHandle):
+ dx = singa.GpuConvBackwardx(dy, self.inputs[1], self.inputs[0],
+ self.handle)
+ dW = singa.GpuConvBackwardW(dy, self.inputs[0], self.inputs[1],
+ self.handle)
+ db = singa.GpuConvBackwardb(
+ dy, self.inputs[2],
+ self.handle) if self.handle.bias_term else None
+ else:
+ dx = singa.CpuConvBackwardx(dy, self.inputs[1], self.inputs[0],
+ self.handle)
+ dW = singa.CpuConvBackwardW(dy, self.inputs[0], self.inputs[1],
+ self.handle)
+ db = singa.CpuConvBackwardb(
+ dy, self.inputs[2],
+ self.handle) if self.handle.bias_term else None
+ if db:
+ return dx, dW, db
+ else:
+ return dx, dW
+```
+
+对于后台的每一个操作符，应按以下方式实现：
+
+- 假设操作符是`foo()`，它的真正实现应该包装在另一个函数中，例如`_foo()`。`foo()`将`_foo`和参数一起作为lambda函数传递给`Device`的`Exec`函数进行缓冲，要读和写的块也同时被传递给`Exec`。
+
+- lambda表达式中使用的所有参数都需要根据以下规则获取：
+
+ - `值捕获`: 如果参数变量是一个局部变量，或者将被立刻释放（例如，中间时序）。否则，一旦`foo()`存在，这些变量将被销毁。
+ - `引用捕获`：如果变量是记录在python端或者是一个持久变量（例如Conv2d类中的参数W和ConvHand）。
+
+ - `可变捕获`: 如果在`_foo()`中修改了由值捕获的变量，则lambda表达式应带有mutable（可变）标签。
+
+下面是一个在后台实现的操作的[例子](https://github.com/apache/singa/blob/master/src/model/operation/convolution.cc)：
+
+```c++
+Tensor GpuConvBackwardx(const Tensor &dy, const Tensor &W, const Tensor &x,
+ const CudnnConvHandle &cch) {
+ CHECK_EQ(dy.device()->lang(), kCuda);
+
+ Tensor dx;
+ dx.ResetLike(x);
+
+ dy.device()->Exec(
+ /*
+ * dx is a local variable so it's captured by value
+ * dy is an intermediate tensor and isn't recorded on the python side
+ * W is an intermediate tensor but it's recorded on the python side
+ * chh is a variable and it's recorded on the python side
+ */
+ [dx, dy, &W, &cch](Context *ctx) mutable {
+ Block *wblock = W.block(), *dyblock = dy.block(), *dxblock = dx.block();
+ float alpha = 1.f, beta = 0.f;
+ cudnnConvolutionBackwardData(
+ ctx->cudnn_handle, &alpha, cch.filter_desc, wblock->data(),
+ cch.y_desc, dyblock->data(), cch.conv_desc, cch.bp_data_alg,
+ cch.workspace.block()->mutable_data(),
+ cch.workspace_count * sizeof(float), &beta, cch.x_desc,
+ dxblock->mutable_data());
+ },
+ {dy.block(), W.block()}, {dx.block(), cch.workspace.block()});
+ /* the lambda expression reads the blocks of tensor dy and w
+ * and writes the blocks of tensor dx and chh.workspace
+ */
+
+ return dx;
+}
+```
+
+## Benchmark
+
+### 单节点
+
+- 实验设定
+ - 模型：
+ - 使用层: ResNet50 in
+ [resnet.py](https://github.com/apache/singa/blob/master/examples/cnn/autograd/resnet_cifar10.py)
+ - 使用模型: ResNet50 in
+ [resnet.py](https://github.com/apache/singa/blob/master/examples/cnn/model/resnet.py)
+ - GPU: NVIDIA RTX 2080Ti
+- 注释：
+ - `s` ：second，秒
+ - `it` ： iteration，迭代次数
+ - `Mem`：peak memory usage of single GPU，单GPU显存峰值
+ - `Throughout`：number of images processed per second，每秒处理的图像数
+ - `Time`：total time，总时间
+ - `Speed`：iterations per second。每秒迭代次数
+ - `Reduction`：the memory usage reduction rate compared with that using layer，与使用层的内存使用率相比，内存使用率降低了多少
+ - `Speedup`: speedup ratio compared with dev branch，与dev分支相比的加速率
+- 结果：
+ <table style="text-align: center">
+ <tr>
+ <th style="text-align: center">Batchsize</th>
+ <th style="text-align: center">Cases</th>
+ <th style="text-align: center">Mem(MB)</th>
+ <th style="text-align: center">Time(s)</th>
+ <th style="text-align: center">Speed(it/s)</th>
+ <th style="text-align: center">Throughput</th>
+ <th style="text-align: center">Reduction</th>
+ <th style="text-align: center">Speedup</th>
+ </tr>
+ <tr>
+ <td rowspan="4">16</td>
+ <td nowrap>layer</td>
+ <td>4975</td>
+ <td>14.1952</td>
+ <td>14.0893</td>
+ <td>225.4285</td>
+ <td>0.00%</td>
+ <td>1.0000</td>
+ </tr>
+ <tr>
+ <td nowrap>model:disable graph</td>
+ <td>4995</td>
+ <td>14.1264</td>
+ <td>14.1579</td>
+ <td>226.5261</td>
+ <td>-0.40%</td>
+ <td>1.0049</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, bfs</td>
+ <td>3283</td>
+ <td>13.7438</td>
+ <td>14.5520</td>
+ <td>232.8318</td>
+ <td>34.01%</td>
+ <td>1.0328</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, serial</td>
+ <td>3265</td>
+ <td>13.7420</td>
+ <td>14.5540</td>
+ <td>232.8635</td>
+ <td>34.37%</td>
+ <td>1.0330</td>
+ </tr>
+ <tr>
+ <td rowspan="4">32</td>
+ <td nowrap>layer</td>
+ <td>10119</td>
+ <td>13.4587</td>
+ <td>7.4302</td>
+ <td>237.7649</td>
+ <td>0.00%</td>
+ <td>1.0000</td>
+ </tr>
+ <tr>
+ <td nowrap>model:disable graph</td>
+ <td>10109</td>
+ <td>13.2952</td>
+ <td>7.5315</td>
+ <td>240.6875</td>
+ <td>0.10%</td>
+ <td>1.0123</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, bfs</td>
+ <td>6839</td>
+ <td>13.1059</td>
+ <td>7.6302</td>
+ <td>244.1648</td>
+ <td>32.41%</td>
+ <td>1.0269</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, serial</td>
+ <td>6845</td>
+ <td>13.0489</td>
+ <td>7.6635</td>
+ <td>245.2312</td>
+ <td>32.35%</td>
+ <td>1.0314</td>
+ </tr>
+ </table>
+
+### 多线程
+
+- 实验设置：
+ - API：
+ - 使用层: ResNet50 in
+ [resnet_dist.py](https://github.com/apache/singa/blob/master/examples/cnn/autograd/resnet_dist.py)
+ - 使用模型: ResNet50 in
+ [resnet.py](https://github.com/apache/singa/blob/master/examples/cnn/model/resnet.py)
+ - GPU: NVIDIA RTX 2080Ti \* 2
+ - MPI: 在同一节点上的两个MPI processes
+- 注释: 与上面相同
+- 结果：
+ <table style="text-align: center">
+ <tr>
+ <th style="text-align: center">Batchsize</th>
+ <th style="text-align: center">Cases</th>
+ <th style="text-align: center">Mem(MB)</th>
+ <th style="text-align: center">Time(s)</th>
+ <th style="text-align: center">Speed(it/s)</th>
+ <th style="text-align: center">Throughput</th>
+ <th style="text-align: center">Reduction</th>
+ <th style="text-align: center">Speedup</th>
+ </tr>
+ <tr>
+ <td rowspan="4">16</td>
+ <td nowrap>layer</td>
+ <td>5439</td>
+ <td>17.3323</td>
+ <td>11.5391</td>
+ <td>369.2522</td>
+ <td>0.00%</td>
+ <td>1.0000</td>
+ </tr>
+ <tr>
+ <td nowrap>model:disable graph</td>
+ <td>5427</td>
+ <td>17.8232</td>
+ <td>11.2213</td>
+ <td>359.0831</td>
+ <td>0.22%</td>
+ <td>0.9725</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, bfs</td>
+ <td>3389</td>
+ <td>18.2310</td>
+ <td>10.9703</td>
+ <td>351.0504</td>
+ <td>37.69%</td>
+ <td>0.9507</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, serial</td>
+ <td>3437</td>
+ <td>17.0389</td>
+ <td>11.7378</td>
+ <td>375.6103</td>
+ <td>36.81%</td>
+ <td>1.0172</td>
+ </tr>
+ <tr>
+ <td rowspan="4">32</td>
+ <td nowrap>layer</td>
+ <td>10547</td>
+ <td>14.8635</td>
+ <td>6.7279</td>
+ <td>430.5858</td>
+ <td>0.00%</td>
+ <td>1.0000</td>
+ </tr>
+ <tr>
+ <td nowrap>model:disable graph</td>
+ <td>10503</td>
+ <td>14.7746</td>
+ <td>6.7684</td>
+ <td>433.1748</td>
+ <td>0.42%</td>
+ <td>1.0060</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, bfs</td>
+ <td>6935</td>
+ <td>14.8553</td>
+ <td>6.7316</td>
+ <td>430.8231</td>
+ <td>34.25%</td>
+ <td>1.0006</td>
+ </tr>
+ <tr>
+ <td nowrap>model:enable graph, serial</td>
+ <td>7027</td>
+ <td>14.3271</td>
+ <td>6.9798</td>
+ <td>446.7074</td>
+ <td>33.37%</td>
+ <td>1.0374</td>
+ </tr>
+ </table>
+
+### 结论
+
+- 在启用计算图的情况下进行训练，可以显著减少内存占用。
+- 目前，在速度上有一点改进。在效率方面还可以做更多的优化。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/history-singa.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/history-singa.md
new file mode 100644
index 0000000..b649472
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/history-singa.md
@@ -0,0 +1,38 @@
+---
+id: version-3.1.0_Chinese-history-singa
+title: History of SINGA
+original_id: history-singa
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 历史
+
+SINGA由新加坡国立大学DB System Group于2014年发起，与浙江大学数据库组合作。如果您在研究中使用SINGA，请引用以下两篇论文。
+
+- B.C. Ooi, K.-L. Tan, S. Wang, W. Wang, Q. Cai, G. Chen, J. Gao, Z. Luo, A. K.
+ H. Tung, Y. Wang, Z. Xie, M. Zhang, and K. Zheng.
+ [SINGA: A distributed deep learning platform](http://www.comp.nus.edu.sg/~ooibc/singaopen-mm15.pdf).
+ ACM Multimedia (Open Source Software Competition) 2015
+
+- W. Wang, G. Chen, T. T. A. Dinh, B. C. Ooi, K.-L.Tan, J. Gao, and S. Wang.
+ [SINGA: putting deep learning in the hands of multimedia users](http://www.comp.nus.edu.sg/~ooibc/singa-mm15.pdf).
+ ACM Multimedia 2015.
+
+Rafiki是SINGA的一个子模块，如果您在研究中使用Rafiki，请引用以下论文。
+
+- Wei Wang, Jinyang Gao, Meihui Zhang, Sheng Wang, Gang Chen, Teck Khim Ng, Beng
+ Chin Ooi, Jie Shao, Moaz Reyad.
+ [Rafiki: Machine Learning as an Analytics Service System](http://www.vldb.org/pvldb/vol12/p128-wang.pdf).
+ [VLDB 2019](http://vldb.org/2019/)
+ ([BibTex](https://dblp.org/rec/bib2/journals/pvldb/WangWGZCNOS18.bib)).
+
+[NetEase](http://tech.163.com/17/0602/17/CLUL016I00098GJ5.html),
+[yzBigData](http://www.yzbigdata.com/en/index.html),
+[Shentilium](https://shentilium.com/), [Foodlg](http://www.foodlg.com/) 以及
+[Medilot](https://medilot.com/technologies) 等公司在他们的工作和应用中使用了SINGA。
+
+## License
+
+SINGA以[Apache License Version 2.0](http://www.apache.org/licenses/LICENSE-2.0)
+版本发布。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/how-to-release.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/how-to-release.md
new file mode 100644
index 0000000..2656b63
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/how-to-release.md
@@ -0,0 +1,173 @@
+---
+id: version-3.1.0_Chinese-how-to-release
+title: How to Prepare a Release
+original_id: how-to-release
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+这是SINGA的[发布准备流程](http://www.apache.org/dev/release-publishing.html)指南。
+
+1. 选择一个发布管理者。发布管理者（RM）是发布过程的协调者，他的签名文件（.asc）将会与发布一起上传。RM 生成 KEY (RSA 4096 位)并将其上传到公钥服务器，首先需要得到其他Apache用户对他的密钥的认可（签名），才能连接到信任网，RM需要先求助其他项目管理者帮忙认证他的密钥。[如何生成密钥？](http://www.apache.org/dev/release-signing.html)
+
+2. 检查license。 [FAQ](https://www.apache.org/legal/src-headers.html#faq-docs);
+ [SINGA Issue](https://issues.apache.org/jira/projects/SINGA/issues/SINGA-447)
+
+ - 代码库不能包含与APL不兼容的第三方代码。
+ - 依赖项与APL兼容，GNU类license不兼容。
+ - 我们编写的所有源文件都必须包含Apache license头：http://www.apache.org/legal/src-headers.html.
+ 链接中有一个脚本可以帮助将这个头同步到所有文件。
+ - 更新LICENSE文件。如果我们在发行包中包含了任何非APL的第三方代码，必须要在NOTICE文件的最后注明。
+
+3. 复查版本。检查代码和文档。
+
+ - 编译过程无错误。
+ - (尽可能地)包含进单元测试。
+ - Conda包运行无误。
+ - Apache网站上的在线文档是最新的。
+
+4. 准备好RELEASE_NOTES文件。包括以下项目，介绍，特性，错误（链接到JIRA或Github PR），变更，依赖列表，不兼容问题，可以按照这个[例子]((http://commons.apache.org/proper/commons-digester/commons-digester-3.0/RELEASE-NOTES.txt))来写。
+
+5. 打包候选版本。该版本应该打包成：apache-singa-VERSION.tar.gz。这个版本不应该包含任何二进制文件，包括git文件。但是CMake的编译依赖于git标签来获取版本号；要删除这个依赖，你需要手动更新CMakeLists.txt文件来设置版本号：
+
+ ```
+ # remove the following lines
+ include(GetGitRevisionDescription)
+ git_describe(VERSION --tags --dirty=-d)
+ string(REGEX REPLACE "^([0-9]+)\\..*" "\\1" VERSION_MAJOR "${VERSION}")
+ string(REGEX REPLACE "^[0-9]+\\.([0-9]+).*" "\\1" VERSION_MINOR "${VERSION}")
+ string(REGEX REPLACE "^[0-9]+\\.[0-9]+\\.([0-9]+).*" "\\1" VERSION_PATCH "${VERSION}")
+
+ # set the numbers manually
+ SET(PACKAGE_VERSION 3.0.0)
+ SET(VERSION 3.0.0)
+ SET(SINGA_MAJOR_VERSION 3) # 0 -
+ SET(SINGA_MINOR_VERSION 0) # 0 - 9
+ SET(SINGA_PATCH_VERSION 0) # 0 - 99
+ ```
+
+ 将软件包上传到[stage repo](https://dist.apache.org/repos/dist/dev/singa/)。应包括tar文件、签名、KEY和SHA256校验和文件。MD5不再使用，详细规则在[这里](https://dist.apache.org/repos/dist/dev/singa/)。阶段文件夹应该包括：
+
+ - apache-singa-VERSION.tar.gz
+ - apache-singa-VERSION.acs
+ - apache-singa-VERSION.SHA256
+
+ 创建这些文件并上传到stage svn repo的命令如下：
+
+ ```sh
+ # in singa repo
+ rm -rf .git
+ rm -rf rafiki/*
+ cd ..
+ tar -czvf apache-singa-VERSION.tar.gz singa/
+
+ mkdir stage
+ cd stage
+ svn co https://dist.apache.org/repos/dist/dev/singa/
+ cd singa
+ # copy the KEYS file from singa repo to this folder if it is not here
+ cp ../../singa/KEYS .
+ mkdir VERSION
+ # copy the tar.gz file
+ mv ../../apache-singa-VERSION.tar.gz VERSION/
+ cd VERSION
+ sha512sum apache-singa-VERSION.tar.gz > apache-singa-VERSION.tar.gz.sha512
+ gpg --armor --output apache-singa-VERSION.tar.gz.asc --detach-sig apache-singa-VERSION.tar.gz
+ cd ..
+ svn add VERSION
+ svn commit
+ ```
+
+6) 通过发送电子邮件的方式进行投票。现举例如下：
+
+```
+ To: dev@singa.apache.org
+ Subject: [VOTE] Release apache-singa-X.Y.Z (release candidate N)
+
+ Hi all,
+
+ I have created a build for Apache SINGA 3.1.0, release candidate 2.
+
+ The release note is at
+ https://github.com/apache/singa/blob/master/RELEASE_NOTES.
+
+ The artifacts to be voted on are located here:
+ https://dist.apache.org/repos/dist/dev/singa/3.1.0.rc2/apache-singa-3.1.0.rc2.tar.gz
+
+ The hashes of the artifacts are as follows:
+ SHA512: 84545499ad36da108c6a599edd1d853f82d331bc03273b5278515554866f0c698e881f956b2eabcb6b29c07fa9fa4ff1add5a777b58db8a6a2362cf383b5c04d
+
+ Release artifacts are signed with the followingkey:
+ https://dist.apache.org/repos/dist/dev/singa/KEYS
+
+ The signature file is:
+ https://dist.apache.org/repos/dist/dev/singa/3.1.0.rc2/apache-singa-3.1.0.rc2.tar.gz.asc
+
+ The Github tag is at:
+ https://github.com/apache/singa/releases/tag/3.1.0.rc2
+
+ The documentation website is at
+ http://singa.apache.org/docs/next/installation/
+
+ Some examples are available for testing:
+ https://github.com/apache/singa/tree/master/examples
+
+
+
+ Please vote on releasing this package. The vote is open for at least 72 hours and passes if a majority of at least three +1 votes are cast.
+
+ [ ] +1 Release this package as Apache SINGA X.Y.Z
+
+ [ ] 0 I don't feel strongly about it, but I'm okay with the release
+
+ [ ] -1 Do not release this package because...
+
+ Here is my vote: +1
+
+```
+
+7) 等待至少48小时的测试回复。任何PMC、提交者或贡献者都可以测试发布的功能，以及反馈。大家在投票+1之前应该检查这些。如果投票通过，则发送如下的结果邮件，否则，从头开始重复刚刚的步骤。
+
+```
+
+To: dev@singa.apache.org Subject: [RESULT][vote] Release apache-singa-X.Y.Z
+(release candidate N)
+
+Thanks to everyone who has voted and given their comments. The tally is as
+follows.
+
+N binding +1s: <names>
+
+N non-binding +1s: <names>
+
+No 0s or -1s.
+
+I am delighted to announce that the proposal to release Apache SINGA X.Y.Z has
+passed.
+
+````
+
+8) 将软件包上传至 https://dist.apache.org/repos/dist/release/singa/，以便[distribution](http://www.apache.org/dev/release-publishing.html#distribution)。
+
+9) 更新SINGA网站的下载页面。tar.gz 文件必须从镜像下载，使用 closer.cgi 脚本；其他工件必须从 Apache 主站点下载。更多细节请看[这里](http://www.apache.org/dev/release-download-pages.html)。我们在之前的版本中得到的一些反馈。“下载页面必须只链接到正式发布的版本，所以不能包含到GitHub的链接”，“链接到KEYS, sig和Hash的链接不能使用dist.apache.org而应该使用 https://www.apache.org/dist/singa/...”“而且你只需要一个KEYS链接，而且应该描述如何使用KEYS+sig或Hash来验证下载。”
+
+10) 删除RC标签并编译conda包。
+
+11) 发布release信息：
+
+ ```
+ To: announce@apache.org, dev@singa.apache.org
+ Subject: [ANNOUNCE] Apache SINGA X.Y.Z released
+
+ We are pleased to announce that SINGA X.Y.Z is released.
+
+ SINGA is a general distributed deep learning platform
+ for training big deep learning models over large datasets.
+ The release is available at: http://singa.apache.org/downloads.html
+ The main features of this release include XXX
+ We look forward to hearing your feedback, suggestions,
+ and contributions to the project.
+
+ On behalf of the SINGA team, {SINGA Team Member Name}
+ ```
+
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/install-win.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/install-win.md
new file mode 100644
index 0000000..3b29ccd
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/install-win.md
@@ -0,0 +1,360 @@
+---
+id: version-3.1.0_Chinese-install-win
+title: Build SINGA on Windows
+original_id: install-win
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+在Microsoft Windows上从源码构建SINGA的过程有四个部分：安装依赖关系、构建SINGA源码、（可选择）安装python模块和（可选择）运行单元测试。
+
+## 安装依赖项
+
+你可以创建一个文件夹来构建依赖关系。
+
+使用到的依赖项有：
+
+- 编译器和IDE：
+ - Visual Studio，社区版是免费的，可以用来构建SINGA。
+ https://www.visualstudio.com/
+- CMake
+ - 可以从 http://cmake.org/ 下载
+ - 确保 cmake 可执行文件的路径在系统路径中，或者在调用 cmake 时使用完整路径。
+- SWIG
+ - 可以从 http://swig.org/ 下载
+ - 确保swig可执行文件的路径在系统路径中，或者在调用swig时使用完整路径。请使用最新的版本，如3.0.12。
+
+- Protocol Buffers
+ - 下载一个合适的版本，如2.6.1:
+ https://github.com/google/protobuf/releases/tag/v2.6.1 。
+ - 下载 protobuf-2.6.1.zip 和 protoc-2.6.1-win32.zip。
+ - 将这两个文件解压到dependencies文件夹中，将protoc可执行文件的路径添加到系统路径中，或者在调用它时使用完整路径。
+ - 打开Visual Studio solution，它可以在vsproject文件夹中找到。
+ - 将build settings改为Release和x64。
+ - 构建libprotobuf项目。
+
+- Openblas
+ - 从 http://www.openblas.net 下载合适的源码，如0.2.20。
+ - 将源码解压到dependencies文件夹中。
+ - 如果你没有安装Perl，请下载一个perl环境，如Strawberry Perl (http://strawberryperl.com/)。
+ - 在源文件夹中运行此命令来构建Visual Studio解决方案：
+
+ ```bash
+ cmake -G "Visual Studio 15 2017 Win64"
+ ```
+
+ - 打开Visual Studio解决方案并将build settings更改为Release和x64。
+ - 构建libopenblas项目。
+
+- Google glog
+ - 下载一个合适的版本，如0.3.5:
+ https://github.com/google/glog/releases
+ - 将源码解压到dependencies文件夹中。
+ - 打开Visual Studio solution.
+ - 将build settings改为Release and x64.
+ - 构建libglog项目。
+
+## 构建SINGA源代码
+
+- 下载SINGA源代码
+- 编译protobuf文件:
+
+ - 在src/proto目录中：
+
+ ```shell
+ mkdir python_out
+ protoc.exe *.proto --python_out python_out
+ ```
+
+- 为C++和Python生成swig接口：在src/api目录中：
+
+ ```shell
+ swig -python -c++ singa.i
+ ```
+
+- 生成SINGA的Visual Studio解决方案：在SINGA源码根目录中：
+
+ ```shell
+ mkdir build
+ cd build
+ ```
+
+- 调用 cmake 并添加系统路径，类似于下面的例子:
+
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64" ^
+ -DGLOG_INCLUDE_DIR="D:/WinSinga/dependencies/glog-0.3.5/src/windows" ^
+ -DGLOG_LIBRARIES="D:/WinSinga/dependencies/glog-0.3.5/x64/Release" ^
+ -DCBLAS_INCLUDE_DIR="D:/WinSinga/dependencies/openblas-0.2.20/lapack-netlib/CBLAS/include" ^
+ -DCBLAS_LIBRARIES="D:/WinSinga/dependencies/openblas-0.2.20/lib/RELEASE" ^
+ -DProtobuf_INCLUDE_DIR="D:/WinSinga/dependencies/protobuf-2.6.1/src" ^
+ -DProtobuf_LIBRARIES="D:/WinSinga/dependencies/protobuf-2.6.1/vsprojects/x64/Release" ^
+ -DProtobuf_PROTOC_EXECUTABLE="D:/WinSinga/dependencies/protoc-2.6.1-win32/protoc.exe" ^
+ ..
+ ```
+
+- 在Visual Studio中打开生成的解决方案。
+- 将构建设置改为Release和x64。
+- 将src/api中的singa_wrap.cxx文件添加到singa_objects项目中。
+- 在 singa_objects 项目中，打开 Additional Include Directories。
+- 添加Python的include路径。
+- 添加numpy的include路径。
+- 添加protobuf的include路径。
+- 在 singa_objects 项目的预处理程序定义中， 添加 USE_GLOG。
+- 构建singa_objects项目。
+
+- 在singa项目中:
+ - 将singa_wrap.obj添加到对象库。
+ - 将目标名称改为"_singa_wrap"。
+ - 将目标扩展名为.pyd。
+ - 将配置类型改为动态库(.dll)。
+ - 进入Additional Library Directories，添加路径到python、openblas、protobuf和glog库。
+ - 在Additional Dependencies中添加libopenblas.lib、libglog.lib和libprotobuf.lib。
+
+- 构建singa项目
+
+## 安装python模块
+
+
+- 将build/python/setup.py中的`_singa_wrap.so`改为`_singa_wrap.pyd`。
+- 将`src/proto/python_out`中的文件复制到`build/python/singa/proto`中。
+
+- （可选）创建并激活一个虚拟环境：
+ ```shell
+ mkdir SingaEnv
+ virtualenv SingaEnv
+ SingaEnv\Scripts\activate
+ ```
+
+- 进入build/python文件夹，运行:
+
+ ```shell
+ python setup.py install
+ ```
+
+- 将 _singa_wrap.pyd、libglog.dll 和 libopenblas.dll 添加到路径中，或者将它们复制到 python site-packages 中的 singa package 文件夹中。
+
+
+- 通过运行如下命令，来验证SINGA是否安装成功：
+
+ ```shell
+ python -c "from singa import tensor"
+ ```
+
+构建过程的视频教程可以在这里找到：
+
+[](https://www.youtube.com/watch?v=cteER7WeiGk)
+
+## 运行单元测试
+
+- 在测试文件夹中，生成Visual Studio解决方案：
+
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64"
+ ```
+
+- 在Visual Studio中打开生成的解决方案。
+
+- 更改build settings为Release和x64。
+
+- 构建glog项目。
+
+- 在test_singa项目中:
+ - 将 USE_GLOG 添加到Preprocessor Definitions中。
+ - 在 Additional Include Directories 中， 添加上面第 2 步中使用的 GLOG_INCLUDE_DIR、 CBLAS_INCLUDE_DIR 和 Protobuf_INCLUDE_DIR 的路径。同时添加build和build/include文件夹。
+ - 转到Additional Library Directories，添加openblas、protobuf和glog库的路径。同时添加 build/src/singa_objects.dir/Release。
+ - 转到 Additional Dependencies 并添加 libopenblas.lib、libglog.lib 和 libprotobuf.lib。修改两个库的名字：gtest.lib和singa_objects.lib。
+
+- 构建test_singa项目。
+
+- 将libglog.dll和libopenblas.dll添加到路径中，或者将它们复制到test/release文件夹中，使其可用。
+
+- 单元测试可以通过如下方式执行：
+
+ - 从命令行:
+
+ ```shell
+ test_singa.exe
+ ```
+
+ - 从Visual Studio:
+ - 右键点击test_singa项目，选择 "Set as StartUp Project"。
+ - 在Debug菜单中，选择'Start Without Debugging'。
+
+单元测试的视频教程可以在这里找到:
+
+[](https://www.youtube.com/watch?v=393gPtzMN1k)
+
+## 构建包含cuda的GPU支持
+
+在本节中，我们将扩展前面的步骤来启用GPU。
+
+### 安装依赖项
+
+除了上面第1节的依赖关系外，我们还需要以下内容：
+
+- CUDA
+
+ 从 https://developer.nvidia.com/cuda-downloads 下载一个合适的版本，如9.1。确保已经安装了Visual Studio集成模块。
+
+
+- cuDNN
+
+ 从 https://developer.nvidia.com/cudnn 下载一个合适的版本，如7.1。
+
+- cnmem:
+
+ - 从 https://github.com/NVIDIA/cnmem 下载最新版本。
+ - 构建Visual Studio解决方案：
+
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64"
+ ```
+
+ - 在Visual Studio中打开生成的解决方案。
+ - 将build settings改为Release和x64。
+ - 构建cnmem项目。
+
+### 构建SINGA源代码
+
+- 调用 cmake 并添加系统路径，类似于下面的例子：
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64" ^
+ -DGLOG_INCLUDE_DIR="D:/WinSinga/dependencies/glog-0.3.5/src/windows" ^
+ -DGLOG_LIBRARIES="D:/WinSinga/dependencies/glog-0.3.5/x64/Release" ^
+ -DCBLAS_INCLUDE_DIR="D:/WinSinga/dependencies/openblas-0.2.20/lapack-netlib/CBLAS/include" ^
+ -DCBLAS_LIBRARIES="D:/WinSinga/dependencies/openblas-0.2.20/lib/RELEASE" ^
+ -DProtobuf_INCLUDE_DIR="D:/WinSinga/dependencies/protobuf-2.6.1/src" ^
+ -DProtobuf_LIBRARIES="D:\WinSinga/dependencies/protobuf-2.6.1/vsprojects/x64/Release" ^
+ -DProtobuf_PROTOC_EXECUTABLE="D:/WinSinga/dependencies/protoc-2.6.1-win32/protoc.exe" ^
+ -DCUDNN_INCLUDE_DIR=D:\WinSinga\dependencies\cudnn-9.1-windows10-x64-v7.1\cuda\include ^
+ -DCUDNN_LIBRARIES=D:\WinSinga\dependencies\cudnn-9.1-windows10-x64-v7.1\cuda\lib\x64 ^
+ -DSWIG_DIR=D:\WinSinga\dependencies\swigwin-3.0.12 ^
+ -DSWIG_EXECUTABLE=D:\WinSinga\dependencies\swigwin-3.0.12\swig.exe ^
+ -DUSE_CUDA=YES ^
+ -DCUDNN_VERSION=7 ^
+ ..
+ ```
+
+* 为C++和Python生成swig接口。在src/api目录中：
+
+ ```shell
+ swig -python -c++ singa.i
+ ```
+
+* 在Visual Studio中打开生成的解决方案
+
+* 将build settings改为Release和x64
+
+#### 构建singa_objects
+
+- 将src/api中的singa_wrap.cxx文件添加到singa_objects项目中。
+- 在 singa_objects 项目中，打开 Additional Include Directories。
+- 添加Python的include路径
+- 添加numpy include路径
+- 添加protobuf包括路径
+- 增加CUDA、cuDNN和cnmem的包含路径。
+- 在 singa_objects 项目的预处理程序定义中， 加入 USE_GLOG、 USE_CUDA 和 USE_CUDNN。删除 DISABLE_WARNINGS。
+- 建立 singa_objects 项目
+
+#### 构建singa-kernel
+
+
+- 创建一个新的Visual Studio项目，类型为 "CUDA 9.1 Runtime"。给它起个名字，比如singa-kernel。
+- 该项目自带一个名为kernel.cu的初始文件，从项目中删除这个文件。
+- 添加这个文件：src/core/tensor/math_kernel.cu。
+- 在项目设置中。
+
+ - 将平台工具集设置为 "Visual Studio 2015 (v140)"
+ - 将 "配置类型 "设置为 "静态库(.lib)"
+ - 在include目录中，添加build/include。
+
+- 建立singa-kernel项目
+
+#### 构建singa
+
+- 在singa项目中：
+
+ - 将singa_wrap.obj添加到对象库中。
+ - 将目标名称改为"_singa_wrap"。
+ - 将目标扩展名为.pyd。
+ - 将配置类型改为动态库(.dll)。
+ - 到Additional Library Directories中添加python、openblas的路径。protobuf和glog库。
+ - 同时添加singa-kernel、cnmem、cuda和cudnn的library path。
+ - 到Additional Dependencies，并添加libopenblas.lib、libglog.lib和 libprotobuf.lib。
+ - 另外还要添加：singa-kernel.lib、cnmem.lib、cudnn.lib、cuda.lib、cublas.lib。curand.lib和cudart.lib。
+
+- 构建singa项目。
+
+### Install Python module
+
+- 将 build/python/setup.py 中的 _singa_wrap.so 改为 _singa_wrap.pyd。
+
+- 将 src/proto/python_out 中的文件复制到 build/python/singa/proto 中。
+
+- （可选） 创建并激活虚拟环境:
+
+ ```shell
+ mkdir SingaEnv
+ virtualenv SingaEnv
+ SingaEnv\Scripts\activate
+ ```
+
+- 进入build/python文件夹，运行:
+
+ ```shell
+ python setup.py install
+ ```
+
+- 将 _singa_wrap.pyd, libglog.dll, libopenblas.dll, cnmem.dll, CUDA Runtime (例如 cudart64_91.dll) 和 cuDNN (例如 cudnn64_7.dll) 添加到路径中，或者将它们复制到 python site-packages 中的 singa package 文件夹中。
+
+- 通过运行如下命令来验证SINGA是否已经安装：
+
+ ```shell
+ python -c "from singa import device; dev = device.create_cuda_gpu()"
+ ```
+
+这个部分的视频教程可以在这里找到：
+
+[](https://www.youtube.com/watch?v=YasKVjRtuDs)
+
+### 运行单元测试
+
+- 在测试文件夹中，生成Visual Studio解决方案：
+
+ ```shell
+ cmake -G "Visual Studio 15 2017 Win64"
+ ```
+
+- 在Visual Studio中打开生成的解决方案，或者将项目添加到步骤5.2中创建的singa解决方案中。
+
+- 将build settings改为Release和x64。
+
+- 构建 glog 项目。
+
+- 在test_singa项目中:
+
+ - 将 USE_GLOG; USE_CUDA; USE_CUDNN 添加到Preprocessor Definitions中。
+ - 在 Additional Include Directories 中， 添加上面 5.2 中使用的 GLOG_INCLUDE_DIR、 CBLAS_INCLUDE_DIR 和 Protobuf_INCLUDE_DIR 的路径。同时添加build、build/include、CUDA和cuDNN的include文件夹。
+ - 转到Additional Library Directories，添加openblas、protobuf和glog库的路径。同时添加 build/src/singa_objects.dir/Release、singa-kernel、cnmem、CUDA 和 cuDNN 库的路径。
+ - 在Additional Dependencies中添加libopenblas.lib; libglog.lib; libprotobuf.lib; cnmem.lib; cudnn.lib; cuda.lib; cublas.lib; curand.lib; cudart.lib; singa-kernel.lib。修正两个库的名字：gtest.lib和singa_objects.lib。
+
+* 构建.
+
+* 将libglog.dll、libopenblas.dll、cnmem.dll、cudart64_91.dll和cudnn64_7.dll添加到路径中，或将它们复制到test/release文件夹中，使其可用。
+
+* 单元测试可以通过如下方式执行：
+
+ - 从命令行:
+
+ ```shell
+ test_singa.exe
+ ```
+
+ - 从 Visual Studio:
+ - 右键点击test_singa项目，选择 'Set as StartUp Project'.
+ - 从Debug菜单，选择 'Start Without Debugging'
+
+运行单元测试的视频教程可以在这里找到：
+
+[](https://www.youtube.com/watch?v=YOjwtrvTPn4)
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/installation.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/installation.md
new file mode 100644
index 0000000..009566c
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/installation.md
@@ -0,0 +1,151 @@
+---
+id: version-3.1.0_Chinese-installation
+title: Installation
+original_id: installation
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 使用Conda
+
+Conda是一个Python、CPP等包的包管理器。
+
+目前，SINGA有Linux和MacOSX的conda包。推荐使用[Miniconda3](https://conda.io/miniconda.html)来配合SINGA使用，安装miniconda后，执行以下命令之一安装SINGA。
+
+1. 只使用CPU
+ [](https://colab.research.google.com/drive/1Ntkhi-Z6XTR8WYPXiLwujHd2dOm0772V?usp=sharing)
+
+```shell
+$ conda install -c nusdbsystem -c conda-forge singa-cpu
+```
+
+2. 使用带CUDA和cuDNN的GPU（需要CUDA驱动>=384.81）
+ [](https://colab.research.google.com/drive/1do_TLJe18IthLOnBOsHCEe-FFPGk1sPJ?usp=sharing)
+
+```shell
+$ conda install -c nusdbsystem -c conda-forge singa-gpu
+```
+
+3. 安装特定版本的SINGA，下面的命令列出了所有可用的SINGA软件包：
+
+```shell
+$ conda search -c nusdbsystem singa
+
+Loading channels: done
+# Name Version Build Channel
+singa 3.1.0.rc2 cpu_py36 nusdbsystem
+singa 3.1.0.rc2 cudnn7.6.5_cuda10.2_py36 nusdbsystem
+singa 3.1.0.rc2 cudnn7.6.5_cuda10.2_py37 nusdbsystem
+```
+
+<!--- > Please note that using the nightly built images is not recommended except for SINGA development and testing. Using stable releases is recommended. -->
+
+下面的命令将安装SINGA的特定版本：
+
+```shell
+$ conda install -c nusdbsystem -c conda-forge singa=X.Y.Z=cpu_py36
+```
+
+如果运行以下命令没有报错：
+```shell
+$ python -c "from singa import tensor"
+```
+
+那么SINGA就安装成功了。
+
+## 使用pip
+
+1. 只使用CPU
+ [](https://colab.research.google.com/drive/17RA056Brwk0vBQTFaZ-l9EbqwADO0NA9?usp=sharing)
+
+```bash
+pip install singa -f http://singa.apache.org/docs/next/wheel-cpu.html --trusted-host singa.apache.org
+```
+
+您可以通过`singa==<version>`安装特定版本的SINGA，其中`<version>`字段应被替换，例如`3.1.0`。可用的SINGA版本在链接中列出。
+
+要安装最新的开发版本，请将链接替换为
+http://singa.apache.org/docs/next/wheel-cpu-dev.html
+
+2. 使用CUDA和cuDNN的GPU
+ [](https://colab.research.google.com/drive/1W30IPCqj5fG8ADAQsFqclaCLyIclVcJL?usp=sharing)
+
+```bash
+pip install singa -f http://singa.apache.org/docs/next/wheel-gpu.html --trusted-host singa.apache.org
+```
+
+您也可以配置SINGA版本和CUDA版本，比如`singa==3.1.0+cuda10.2`，SINGA版本和CUDA版本的可用组合在链接中列出。
+
+要安装最新的开发版本，请将链接替换为
+http://singa.apache.org/docs/next/wheel-gpu-dev.html
+
+注意：你本地Python环境的Python版本将被用来寻找相应的wheel包。例如，如果你本地的Python是3.6，那么就会通过pip选择在Python 3.6上编译的wheel包并安装。事实上，wheel文件的名称包括SINGA版本、CUDA版本和Python版本。因此，`pip`知道要下载和安装哪个wheel文件。
+
+参考setup.py文件顶部的注释，了解如何构建wheel包。
+
+## 使用Docker
+
+按照[说明](https://docs.docker.com/install/)在你的本地主机上安装Docker。将您的用户添加到[docker组](https://docs.docker.com/install/linux/linux-postinstall/)中，以便在没有`sudo`的情况下运行docker命令。
+
+1. 仅使用CPU
+
+```shell
+$ docker run -it apache/singa:X.Y.Z-cpu-ubuntu16.04 /bin/bash
+```
+
+2. 要使用GPU，在安装Docker后安装
+ [Nvidia-Docker](https://github.com/NVIDIA/nvidia-docker)
+
+```shell
+$ nvidia-docker run -it apache/singa:X.Y.Z-cuda9.0-cudnn7.4.2-ubuntu16.04 /bin/bash
+```
+
+3. 关于SINGA Docker镜像（标签）的完整列表，请访问[docker hub site](https://hub.docker.com/r/apache/singa/)。对于每个docker镜像，标签的命名为：
+
+```shell
+version-(cpu|gpu)[-devel]
+```
+
+| Tag | Description | Example value |
+| --------- | -------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `version` | SINGA version | '2.0.0-rc0', '2.0.0', '1.2.0' |
+| `cpu` | the image cannot run on GPUs | 'cpu' |
+| `gpu` | the image can run on Nvidia GPUs | 'gpu', or 'cudax.x-cudnnx.x' e.g., 'cuda10.0-cudnn7.3' |
+| `devel` | indicator for development | if absent, SINGA Python package is installed for runtime only; if present, the building environment is also created, you can recompile SINGA from source at '/root/singa' |
+| `OS` | indicate OS version number | 'ubuntu16.04', 'ubuntu18.04' |
+
+## 从源码编译
+您可以使用本地构建工具或conda-build在本地主机操作系统上或在Docker容器中从源代码[构建和安装SINGA](build.md)。
+
+## FAQ
+
+- Q: `from singa import tensor`错误
+
+ A: 执行下面的命令，检查详细的错误：
+
+ ```shell
+ python -c "from singa import _singa_wrap"
+ # go to the folder of _singa_wrap.so
+ ldd path to _singa_wrap.so
+ python
+ >> import importlib
+ >> importlib.import_module('_singa_wrap')
+ ```
+ `_singa_wrap.so` 的文件夹是 `~/miniconda3/lib/python3.7/site-packages/singa`。通常情况下，这个错误是由于依赖的库不匹配或缺失造成的，例如 cuDNN 或 protobuf。解决方法是创建一个新的虚拟环境，并在该环境中安装SINGA，例如：
+
+
+ ```shell
+ conda create -n singa
+ conda activate singa
+ conda install -c nusdbsystem -c conda-forge singa-cpu
+ ```
+
+- Q: 使用虚拟环境时，每次安装SINGA时，都会重新安装numpy。但是，当我运行`import numpy`时，numpy没有被使用。
+
+ A:
+ 这可能是由`PYTHONPATH`环境变量引起的，在使用虚拟环境时，应将其设置为空，以避免与虚拟环境的路径冲突。
+
+- Q: 当我在Mac OS X中运行SINGA时，得到如下错误 "Fatal Python error:
+ PyThreadState_Get: no current thread Abort trap: 6"
+
+ A: 这个错误通常发生在系统中有多个 Python 版本的时候，例如，操作系统自带的版本和 Homebrew 安装的版本。SINGA链接的Python必须与Python解释器相同。您可以通过`which python`来检查解释器python版本并通过`otool -L <path to _singa_wrap.so>` 检查 SINGA 链接的 Python，如果通过conda安装SINGA，这个问题应该可以解决。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/issue-tracking.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/issue-tracking.md
new file mode 100644
index 0000000..af94f1f
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/issue-tracking.md
@@ -0,0 +1,11 @@
+---
+id: version-3.1.0_Chinese-issue-tracking
+title: Issue Tracking
+original_id: issue-tracking
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA使用[JIRA](https://issues.apache.org/jira/browse/singa)来管理问题，包括bug、新功能和讨论。
+
+我们现在正在转移到[Github Issues](https://github.com/apache/singa/issues)。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/mail-lists.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/mail-lists.md
new file mode 100644
index 0000000..8205c2f
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/mail-lists.md
@@ -0,0 +1,15 @@
+---
+id: version-3.1.0_Chinese-mail-lists
+title: Project Mailing Lists
+original_id: mail-lists
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+这些是为这个项目建立的邮件列表。每个列表都有一个订阅、退订和归档链接：
+
+| Name | Post | Subscribe | Unsubscribe | Archive |
+| ----------- | ------------------------------------ | ---------------------------------------------------------------- | -------------------------------------------------------------------- | ----------------------------------------------------------------------------------- |
+| Development | <dev@singa.incubator.apache.org> | [Subscribe](mailto:dev-subscribe@singa.incubator.apache.org) | [Unsubscribe](mailto:dev-unsubscribe@singa.incubator.apache.org.) | [mail-archives.apache.org](http://mail-archives.apache.org/mod_mbox/singa-dev/) |
+| Commits | <commits@singa.incubator.apache.org> | [Subscribe](mailto:commits-subscribe@singa.incubator.apache.org) | [Unsubscribe](mailto:commits-unsubscribe@singa.incubator.apache.org) | [mail-archives.apache.org](http://mail-archives.apache.org/mod_mbox/singa-commits/) |
+| Security | <security@singa.apache.org> | private | private | private |
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/onnx.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/onnx.md
new file mode 100644
index 0000000..ade6d96
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/onnx.md
@@ -0,0 +1,674 @@
+---
+id: version-3.1.0_Chinese-onnx
+title: ONNX
+original_id: onnx
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+[ONNX](https://onnx.ai/) 是机器学习模型的开放表示格式，它使AI开发人员能够在不同的库和工具中使用模型。SINGA支持加载ONNX格式模型用于训练和inference，并将使用SINGA API（如[Module](./module)）定义的模型保存为ONNX格式。
+
+SINGA在以下[版本](https://github.com/onnx/onnx/blob/master/docs/Versioning.md)中的ONNX中测试过。
+
+| ONNX version | File format version | Opset version ai.onnx | Opset version ai.onnx.ml | Opset version ai.onnx.training |
+| ------------ | ------------------- | --------------------- | ------------------------ | ------------------------------ |
+| 1.6.0 | 6 | 11 | 2 | - |
+
+## 通常用法
+
+### 从ONNX中读取一个Model到SINGA
+
+在通过 `onnx.load` 从磁盘加载 ONNX 模型后，您需要更新模型的batch_size，因为对于大多数模型来说，它们使用一个占位符来表示其批处理量。我们在这里举一个例子，若要 `update_batch_size`，你只需要更新输入和输出的 batch_size，内部的 tensors 的形状会自动推断出来。
+
+
+然后，您可以使用 `sonnx.prepare` 来准备 SINGA 模型。该函数将 ONNX 模型图中的所有节点迭代并翻译成 SINGA 运算符，加载所有存储的权重并推断每个中间张量的形状。
+
+```python3
+import onnx
+from singa import device
+from singa import sonnx
+
+# if the input has multiple tensors? can put this function inside prepare()?
+def update_batch_size(onnx_model, batch_size):
+ model_input = onnx_model.graph.input[0]
+ model_input.type.tensor_type.shape.dim[0].dim_value = batch_size
+ model_output = onnx_model.graph.output[0]
+ model_output.type.tensor_type.shape.dim[0].dim_value = batch_size
+ return onnx_model
+
+
+model_path = "PATH/To/ONNX/MODEL"
+onnx_model = onnx.load(model_path)
+
+# set batch size
+onnx_model = update_batch_size(onnx_model, 1)
+
+# convert onnx graph nodes into SINGA operators
+dev = device.create_cuda_gpu()
+sg_ir = sonnx.prepare(onnx_model, device=dev)
+```
+
+### Inference SINGA模型
+
+一旦创建了模型，就可以通过调用`sg_ir.run`进行inference。输入和输出必须是SINGA Tensor实例，由于SINGA模型以列表形式返回输出，如果只有一个输出，你只需要从输出中取第一个元素即可。
+
+```python3
+# can warp the following code in prepare()
+# and provide a flag training=True/False?
+
+class Infer:
+
+
+ def __init__(self, sg_ir):
+ self.sg_ir = sg_ir
+
+ def forward(self, x):
+ return sg_ir.run([x])[0]
+
+
+data = get_dataset()
+x = tensor.Tensor(device=dev, data=data)
+
+model = Infer(sg_ir)
+y = model.forward(x)
+```
+
+### 将SINGA模型保存成ONNX格式
+
+给定输入时序和输出时序，由运算符产生的模型，你可以追溯所有内部操作。因此，一个SINGA模型是由输入和输出张量定义的，要将 SINGA 模型导出为 ONNX 格式，您只需提供输入和输出张量列表。
+
+```python3
+# x is the input tensor, y is the output tensor
+sonnx.to_onnx([x], [y])
+```
+
+### 在ONNX模型上重新训练
+
+要使用 SINGA 训练（或改进）ONNX 模型，您需要设置内部的张量为可训练状态：
+
+```python3
+class Infer:
+
+ def __init__(self, sg_ir):
+ self.sg_ir = sg_ir
+ ## can wrap these codes in sonnx?
+ for idx, tens in sg_ir.tensor_map.items():
+ # allow the tensors to be updated
+ tens.requires_grad = True
+ tens.stores_grad = True
+
+ def forward(self, x):
+ return sg_ir.run([x])[0]
+
+autograd.training = False
+model = Infer(sg_ir)
+
+autograd.training = True
+# then you training the model like normal
+# give more details??
+```
+
+### 在ONNX模型上做迁移学习
+
+您也可以在ONNX模型的最后附加一些图层来进行转移学习。`last_layers` 意味着您从 [0, last_layers] 切断 ONNX 层。然后您可以通过普通的SINGA模型附加更多的层。
+
+```python3
+class Trans:
+
+ def __init__(self, sg_ir, last_layers):
+ self.sg_ir = sg_ir
+ self.last_layers = last_layers
+ self.append_linear1 = autograd.Linear(500, 128, bias=False)
+ self.append_linear2 = autograd.Linear(128, 32, bias=False)
+ self.append_linear3 = autograd.Linear(32, 10, bias=False)
+
+ def forward(self, x):
+ y = sg_ir.run([x], last_layers=self.last_layers)[0]
+ y = self.append_linear1(y)
+ y = autograd.relu(y)
+ y = self.append_linear2(y)
+ y = autograd.relu(y)
+ y = self.append_linear3(y)
+ y = autograd.relu(y)
+ return y
+
+autograd.training = False
+model = Trans(sg_ir, -1)
+
+# then you training the model like normal
+```
+
+## 一个完整示例
+
+本部分以mnist为例，介绍SINGA ONNX的使用方法。在这部分，将展示如何导出、加载、inference、再训练和迁移学习 mnist 模型的例子。您可以在[这里](https://colab.research.google.com/drive/1-YOfQqqw3HNhS8WpB8xjDQYutRdUdmCq)试用这部分内容。
+
+### 读取数据集
+
+首先，你需要导入一些必要的库，并定义一些辅助函数来下载和预处理数据集：
+
+```python
+import os
+import urllib.request
+import gzip
+import numpy as np
+import codecs
+
+from singa import device
+from singa import tensor
+from singa import opt
+from singa import autograd
+from singa import sonnx
+import onnx
+
+
+def load_dataset():
+ train_x_url = 'http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz'
+ train_y_url = 'http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz'
+ valid_x_url = 'http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz'
+ valid_y_url = 'http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz'
+ train_x = read_image_file(check_exist_or_download(train_x_url)).astype(
+ np.float32)
+ train_y = read_label_file(check_exist_or_download(train_y_url)).astype(
+ np.float32)
+ valid_x = read_image_file(check_exist_or_download(valid_x_url)).astype(
+ np.float32)
+ valid_y = read_label_file(check_exist_or_download(valid_y_url)).astype(
+ np.float32)
+ return train_x, train_y, valid_x, valid_y
+
+
+def check_exist_or_download(url):
+
+ download_dir = '/tmp/'
+
+ name = url.rsplit('/', 1)[-1]
+ filename = os.path.join(download_dir, name)
+ if not os.path.isfile(filename):
+ print("Downloading %s" % url)
+ urllib.request.urlretrieve(url, filename)
+ return filename
+
+
+def read_label_file(path):
+ with gzip.open(path, 'rb') as f:
+ data = f.read()
+ assert get_int(data[:4]) == 2049
+ length = get_int(data[4:8])
+ parsed = np.frombuffer(data, dtype=np.uint8, offset=8).reshape(
+ (length))
+ return parsed
+
+
+def get_int(b):
+ return int(codecs.encode(b, 'hex'), 16)
+
+
+def read_image_file(path):
+ with gzip.open(path, 'rb') as f:
+ data = f.read()
+ assert get_int(data[:4]) == 2051
+ length = get_int(data[4:8])
+ num_rows = get_int(data[8:12])
+ num_cols = get_int(data[12:16])
+ parsed = np.frombuffer(data, dtype=np.uint8, offset=16).reshape(
+ (length, 1, num_rows, num_cols))
+ return parsed
+
+
+def to_categorical(y, num_classes):
+ y = np.array(y, dtype="int")
+ n = y.shape[0]
+ categorical = np.zeros((n, num_classes))
+ categorical[np.arange(n), y] = 1
+ categorical = categorical.astype(np.float32)
+ return categorical
+```
+
+### MNIST模型
+
+然后你可以定义一个叫做**CNN**的类来构造mnist模型，这个模型由几个卷积层、池化层、全连接层和relu层组成。你也可以定义一个函数来计算我们结果的**准确性**。最后，你可以定义一个**训练函数**和一个**测试函数**来处理训练和预测的过程。
+
+```python
+class CNN:
+ def __init__(self):
+ self.conv1 = autograd.Conv2d(1, 20, 5, padding=0)
+ self.conv2 = autograd.Conv2d(20, 50, 5, padding=0)
+ self.linear1 = autograd.Linear(4 * 4 * 50, 500, bias=False)
+ self.linear2 = autograd.Linear(500, 10, bias=False)
+ self.pooling1 = autograd.MaxPool2d(2, 2, padding=0)
+ self.pooling2 = autograd.MaxPool2d(2, 2, padding=0)
+
+ def forward(self, x):
+ y = self.conv1(x)
+ y = autograd.relu(y)
+ y = self.pooling1(y)
+ y = self.conv2(y)
+ y = autograd.relu(y)
+ y = self.pooling2(y)
+ y = autograd.flatten(y)
+ y = self.linear1(y)
+ y = autograd.relu(y)
+ y = self.linear2(y)
+ return y
+
+
+def accuracy(pred, target):
+ y = np.argmax(pred, axis=1)
+ t = np.argmax(target, axis=1)
+ a = y == t
+ return np.array(a, "int").sum() / float(len(t))
+
+
+def train(model,
+ x,
+ y,
+ epochs=1,
+ batch_size=64,
+ dev=device.get_default_device()):
+ batch_number = x.shape[0] // batch_size
+
+ for i in range(epochs):
+ for b in range(batch_number):
+ l_idx = b * batch_size
+ r_idx = (b + 1) * batch_size
+
+ x_batch = tensor.Tensor(device=dev, data=x[l_idx:r_idx])
+ target_batch = tensor.Tensor(device=dev, data=y[l_idx:r_idx])
+
+ output_batch = model.forward(x_batch)
+ # onnx_model = sonnx.to_onnx([x_batch], [y])
+ # print('The model is:\n{}'.format(onnx_model))
+
+ loss = autograd.softmax_cross_entropy(output_batch, target_batch)
+ accuracy_rate = accuracy(tensor.to_numpy(output_batch),
+ tensor.to_numpy(target_batch))
+
+ sgd = opt.SGD(lr=0.001)
+ for p, gp in autograd.backward(loss):
+ sgd.update(p, gp)
+ sgd.step()
+
+ if b % 1e2 == 0:
+ print("acc %6.2f loss, %6.2f" %
+ (accuracy_rate, tensor.to_numpy(loss)[0]))
+ print("training completed")
+ return x_batch, output_batch
+
+def test(model, x, y, batch_size=64, dev=device.get_default_device()):
+ batch_number = x.shape[0] // batch_size
+
+ result = 0
+ for b in range(batch_number):
+ l_idx = b * batch_size
+ r_idx = (b + 1) * batch_size
+
+ x_batch = tensor.Tensor(device=dev, data=x[l_idx:r_idx])
+ target_batch = tensor.Tensor(device=dev, data=y[l_idx:r_idx])
+
+ output_batch = model.forward(x_batch)
+ result += accuracy(tensor.to_numpy(output_batch),
+ tensor.to_numpy(target_batch))
+
+ print("testing acc %6.2f" % (result / batch_number))
+```
+
+### 训练mnist模型并将其导出到onnx
+
+现在，你可以通过调用 **soonx.to_onnx** 函数来训练 mnist 模型并导出其 onnx 模型。
+
+```python
+def make_onnx(x, y):
+ return sonnx.to_onnx([x], [y])
+
+# create device
+dev = device.create_cuda_gpu()
+#dev = device.get_default_device()
+# create model
+model = CNN()
+# load data
+train_x, train_y, valid_x, valid_y = load_dataset()
+# normalization
+train_x = train_x / 255
+valid_x = valid_x / 255
+train_y = to_categorical(train_y, 10)
+valid_y = to_categorical(valid_y, 10)
+# do training
+autograd.training = True
+x, y = train(model, train_x, train_y, dev=dev)
+onnx_model = make_onnx(x, y)
+# print('The model is:\n{}'.format(onnx_model))
+
+# Save the ONNX model
+model_path = os.path.join('/', 'tmp', 'mnist.onnx')
+onnx.save(onnx_model, model_path)
+print('The model is saved.')
+```
+
+### Inference
+
+导出onnx模型后，可以在'/tmp'目录下找到一个名为**mnist.onnx**的文件，这个模型可以被其他库导入。现在，如果你想把这个onnx模型再次导入到singa中，并使用验证数据集进行推理，你可以定义一个叫做**Infer**的类，Infer的前向函数将被测试函数调用，对验证数据集进行推理。此外，你应该把训练的标签设置为**False**，以固定自变量算子的梯度。
+
+在导入onnx模型时，需要先调用**onnx.load**来加载onnx模型。然后将onnx模型输入到 **soonx.prepare**中进行解析，并启动到一个singa模型(代码中的**sg_ir**)。sg_ir里面包含了一个singa图，然后就可以通过输入到它的run函数中运行一步推理。
+
+```python
+class Infer:
+ def __init__(self, sg_ir):
+ self.sg_ir = sg_ir
+ for idx, tens in sg_ir.tensor_map.items():
+ # allow the tensors to be updated
+ tens.requires_grad = True
+ tens.stores_grad= True
+ sg_ir.tensor_map[idx] = tens
+
+ def forward(self, x):
+ return sg_ir.run([x])[0] # we can run one step of inference by feeding input
+
+# load the ONNX model
+onnx_model = onnx.load(model_path)
+sg_ir = sonnx.prepare(onnx_model, device=dev) # parse and initiate to a singa model
+
+# inference
+autograd.training = False
+print('The inference result is:')
+test(Infer(sg_ir), valid_x, valid_y, dev=dev)
+```
+
+### 重训练
+
+假设导入模型后，想再次对模型进行重新训练，我们可以定义一个名为**re_train**的函数。在调用这个re_train函数之前，我们应该将训练的标签设置为**True**，以使自变量运算符更新其梯度。而在完成训练后，我们再将其设置为**False**，以调用做推理的测试函数。
+
+```python
+def re_train(sg_ir,
+ x,
+ y,
+ epochs=1,
+ batch_size=64,
+ dev=device.get_default_device()):
+ batch_number = x.shape[0] // batch_size
+
+ new_model = Infer(sg_ir)
+
+ for i in range(epochs):
+ for b in range(batch_number):
+ l_idx = b * batch_size
+ r_idx = (b + 1) * batch_size
+
+ x_batch = tensor.Tensor(device=dev, data=x[l_idx:r_idx])
+ target_batch = tensor.Tensor(device=dev, data=y[l_idx:r_idx])
+
+ output_batch = new_model.forward(x_batch)
+
+ loss = autograd.softmax_cross_entropy(output_batch, target_batch)
+ accuracy_rate = accuracy(tensor.to_numpy(output_batch),
+ tensor.to_numpy(target_batch))
+
+ sgd = opt.SGD(lr=0.01)
+ for p, gp in autograd.backward(loss):
+ sgd.update(p, gp)
+ sgd.step()
+
+ if b % 1e2 == 0:
+ print("acc %6.2f loss, %6.2f" %
+ (accuracy_rate, tensor.to_numpy(loss)[0]))
+ print("re-training completed")
+ return new_model
+
+# load the ONNX model
+onnx_model = onnx.load(model_path)
+sg_ir = sonnx.prepare(onnx_model, device=dev)
+
+# re-training
+autograd.training = True
+new_model = re_train(sg_ir, train_x, train_y, dev=dev)
+autograd.training = False
+test(new_model, valid_x, valid_y, dev=dev)
+```
+
+### 迁移学习
+
+最后，如果我们想做迁移学习，我们可以定义一个名为**Trans**的函数，在onnx模型后追加一些层。为了演示，代码中只在onnx模型后追加了几个线性（全连接）和relu。可以定义一个transfer_learning函数来处理transfer-learning模型的训练过程，而训练的标签和前面一个的一样。
+
+```python
+class Trans:
+ def __init__(self, sg_ir, last_layers):
+ self.sg_ir = sg_ir
+ self.last_layers = last_layers
+ self.append_linear1 = autograd.Linear(500, 128, bias=False)
+ self.append_linear2 = autograd.Linear(128, 32, bias=False)
+ self.append_linear3 = autograd.Linear(32, 10, bias=False)
+
+ def forward(self, x):
+ y = sg_ir.run([x], last_layers=self.last_layers)[0]
+ y = self.append_linear1(y)
+ y = autograd.relu(y)
+ y = self.append_linear2(y)
+ y = autograd.relu(y)
+ y = self.append_linear3(y)
+ y = autograd.relu(y)
+ return y
+
+def transfer_learning(sg_ir,
+ x,
+ y,
+ epochs=1,
+ batch_size=64,
+ dev=device.get_default_device()):
+ batch_number = x.shape[0] // batch_size
+
+ trans_model = Trans(sg_ir, -1)
+
+ for i in range(epochs):
+ for b in range(batch_number):
+ l_idx = b * batch_size
+ r_idx = (b + 1) * batch_size
+
+ x_batch = tensor.Tensor(device=dev, data=x[l_idx:r_idx])
+ target_batch = tensor.Tensor(device=dev, data=y[l_idx:r_idx])
+ output_batch = trans_model.forward(x_batch)
+
+ loss = autograd.softmax_cross_entropy(output_batch, target_batch)
+ accuracy_rate = accuracy(tensor.to_numpy(output_batch),
+ tensor.to_numpy(target_batch))
+
+ sgd = opt.SGD(lr=0.07)
+ for p, gp in autograd.backward(loss):
+ sgd.update(p, gp)
+ sgd.step()
+
+ if b % 1e2 == 0:
+ print("acc %6.2f loss, %6.2f" %
+ (accuracy_rate, tensor.to_numpy(loss)[0]))
+ print("transfer-learning completed")
+ return trans_mode
+
+# load the ONNX model
+onnx_model = onnx.load(model_path)
+sg_ir = sonnx.prepare(onnx_model, device=dev)
+
+# transfer-learning
+autograd.training = True
+new_model = transfer_learning(sg_ir, train_x, train_y, dev=dev)
+autograd.training = False
+test(new_model, valid_x, valid_y, dev=dev)
+```
+
+## ONNX模型库
+
+[ONNX 模型库](https://github.com/onnx/models)是由社区成员贡献的 ONNX 格式的预先训练的最先进模型的集合。SINGA 现在已经支持了几个 CV 和 NLP 模型。将来会支持更多模型。
+
+### 图像分类
+
+这套模型以图像作为输入，然后将图像中的主要物体分为1000个物体类别，如键盘、鼠标、铅笔和许多动物。
+
+| Model Class | Reference | Description | Link |
+| --------------------------------------------------------------------------------------------------- | ------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <b>[MobileNet](https://github.com/onnx/models/tree/master/vision/classification/mobilenet)</b> | [Sandler et al.](https://arxiv.org/abs/1801.04381) | 最适合移动和嵌入式视觉应用的轻量级深度神经网络。 <br>Top-5 error from paper - ~10% | [](https://colab.research.google.com/drive/1HsixqJMIpKyEPhkbB8jy7NwNEFEAUWAf) |
+| <b>[ResNet18](https://github.com/onnx/models/tree/master/vision/classification/resnet)</b> | [He et al.](https://arxiv.org/abs/1512.03385) | 一个CNN模型（多达152层），在对图像进行分类时，使用shortcut来实现更高的准确性。 <br> Top-5 error from paper - ~3.6% | [](https://colab.research.google.com/drive/1u1RYefSsVbiP4I-5wiBKHjsT9L0FxLm9) |
+| <b>[VGG16](https://github.com/onnx/models/tree/master/vision/classification/vgg)</b> | [Simonyan et al.](https://arxiv.org/abs/1409.1556) | 深度CNN模型（多达19层）。类似于AlexNet，但使用多个较小的内核大小的滤波器，在分类图像时提供更高的准确性。 <br>Top-5 error from paper - ~8% | [](https://colab.research.google.com/drive/14kxgRKtbjPCKKsDJVNi3AvTev81Gp_Ds) |
+| <b>[ShuffleNet_V2](https://github.com/onnx/models/tree/master/vision/classification/shufflenet)</b> | [Simonyan et al.](https://arxiv.org/pdf/1707.01083.pdf) | 专门为移动设备设计的计算效率极高的CNN模型。这种网络架构设计考虑了速度等直接指标，而不是FLOP等间接指标。 Top-1 error from paper - ~30.6% | [](https://colab.research.google.com/drive/19HfRu3YHP_H2z3BcZujVFRp23_J5XsuA?usp=sharing) |
+
+### 目标检测
+
+目标检测模型可以检测图像中是否存在多个对象，并将图像中检测到对象的区域分割出来。
+
+| Model Class | Reference | Description | Link |
+| ----------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <b>[Tiny YOLOv2](https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/tiny_yolov2)</b> | [Redmon et al.](https://arxiv.org/pdf/1612.08242.pdf) | 一个用于目标检测的实时CNN，可以检测20个不同的类。一个更复杂的完整YOLOv2网络的小版本。 | [](https://colab.research.google.com/drive/11V4I6cRjIJNUv5ZGsEGwqHuoQEie6b1T) |
+
+### 面部识别
+
+人脸检测模型可以识别和/或识别给定图像中的人脸和情绪。
+
+| Model Class | Reference | Description | Link |
+| --------------------------------------------------------------------------------------------------------- | -------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <b>[ArcFace](https://github.com/onnx/models/tree/master/vision/body_analysis/arcface)</b> | [Deng et al.](https://arxiv.org/abs/1801.07698) | 一种基于CNN的人脸识别模型，它可以学习人脸的判别特征，并对输入的人脸图像进行分析。 | [](https://colab.research.google.com/drive/1qanaqUKGIDtifdzEzJOHjEj4kYzA9uJC) |
+| <b>[Emotion FerPlus](https://github.com/onnx/models/tree/master/vision/body_analysis/emotion_ferplus)</b> | [Barsoum et al.](https://arxiv.org/abs/1608.01041) | 基于人脸图像训练的情感识别深度CNN。 | [](https://colab.research.google.com/drive/1XHtBQGRhe58PDi4LGYJzYueWBeWbO23r) |
+
+### 机器理解
+
+这个自然语言处理模型的子集，可以回答关于给定上下文段落的问题。
+
+| Model Class | Reference | Description | Link |
+| ----------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <b>[BERT-Squad](https://github.com/onnx/models/tree/master/text/machine_comprehension/bert-squad)</b> | [Devlin et al.](https://arxiv.org/pdf/1810.04805.pdf) | 该模型根据给定输入段落的上下文回答问题。 | [](https://colab.research.google.com/drive/1kud-lUPjS_u-TkDAzihBTw0Vqr0FjCE-) |
+| <b>[RoBERTa](https://github.com/onnx/models/tree/master/text/machine_comprehension/roberta)</b> | [Devlin et al.](https://arxiv.org/pdf/1907.11692.pdf) | 一个基于大型变换器的模型，根据给定的输入文本预测情感。 | [](https://colab.research.google.com/drive/1F-c4LJSx3Cb2jW6tP7f8nAZDigyLH6iN?usp=sharing) |
+| <b>[GPT-2](https://github.com/onnx/models/tree/master/text/machine_comprehension/gpt-2)</b> | [Devlin et al.](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) | 一个基于大型变换器的语言模型，给定一些文本中的单词序列，预测下一个单词。 | [](https://colab.research.google.com/drive/1ZlXLSIMppPch6HgzKRillJiUcWn3PiK7?usp=sharing) |
+
+## 支持的操作符
+
+onnx支持下列运算:
+
+- Acos
+- Acosh
+- Add
+- And
+- Asin
+- Asinh
+- Atan
+- Atanh
+- AveragePool
+- BatchNormalization
+- Cast
+- Ceil
+- Clip
+- Concat
+- ConstantOfShape
+- Conv
+- Cos
+- Cosh
+- Div
+- Dropout
+- Elu
+- Equal
+- Erf
+- Expand
+- Flatten
+- Gather
+- Gemm
+- GlobalAveragePool
+- Greater
+- HardSigmoid
+- Identity
+- LeakyRelu
+- Less
+- Log
+- MatMul
+- Max
+- MaxPool
+- Mean
+- Min
+- Mul
+- Neg
+- NonZero
+- Not
+- OneHot
+- Or
+- Pad
+- Pow
+- PRelu
+- Reciprocal
+- ReduceMean
+- ReduceSum
+- Relu
+- Reshape
+- ScatterElements
+- Selu
+- Shape
+- Sigmoid
+- Sign
+- Sin
+- Sinh
+- Slice
+- Softmax
+- Softplus
+- Softsign
+- Split
+- Sqrt
+- Squeeze
+- Sub
+- Sum
+- Tan
+- Tanh
+- Tile
+- Transpose
+- Unsqueeze
+- Upsample
+- Where
+- Xor
+
+### 对ONNX后端的特别说明
+
+- Conv, MaxPool 以及 AveragePool
+
+ 输入必须是1d`(N*C*H)`和2d`(N*C*H*W)`的形状，`dilation`必须是1。
+
+- BatchNormalization
+
+ `epsilon` 设定为1e-05，不能改变
+
+- Cast
+
+ 只支持float32和int32，其他类型都会转向这两种类型。
+
+- Squeeze and Unsqueeze
+
+ 如果你在`Tensor`和Scalar之间`Squeeze`或`Unsqueeze`时遇到错误，请向我们报告。
+
+- Empty tensor
+
+ 空张量在SINGA是非法的。
+
+## 实现
+
+SINGA ONNX的代码在`python/singa/soonx.py`中，主要有三个类，`SingaFrontend`、`SingaBackend`和`SingaRep`。`SingaFrontend`将SINGA模型翻译成ONNX模型；`SingaBackend`将ONNX模型翻译成`SingaRep`对象，其中存储了所有的SINGA运算符和张量（本文档中的张量指SINGA Tensor）；`SingaRep`可以像SINGA模型一样运行。
+
+### SingaFrontend
+
+`SingaFrontend`的入口函数是`singa_to_onnx_model`，它也被称为`to_onnx`，`singa_to_onnx_model`创建了ONNX模型，它还通过`singa_to_onnx_graph`创建了一个ONNX图。
+
+
+`singa_to_onnx_graph`接受模型的输出，并从输出中递归迭代SINGA模型的图，得到所有的运算符，形成一个队列。SINGA模型的输入和中间张量，即可训练的权重，同时被获取。输入存储在`onnx_model.graph.input`中；输出存储在`onnx_model.graph.output`中；可训练权重存储在`onnx_model.graph.initializer`中。
+
+然后将队列中的SINGA运算符逐一翻译成ONNX运算符。`_rename_operators` 定义了 SINGA 和 ONNX 之间的运算符名称映射。`_special_operators` 定义了翻译运算符时要使用的函数。
+
+此外，SINGA 中的某些运算符与 ONNX 的定义不同，即 ONNX 将 SINGA 运算符的某些属性视为输入，因此 `_unhandled_operators` 定义了处理特殊运算符的函数。
+
+由于SINGA中的布尔类型被视为int32，所以`_bool_operators`定义了要改变的操作符为布尔类型。
+
+### SingaBackend
+
+`SingaBackend`的入口函数是`prepare`，它检查ONNX模型的版本，然后调用`_onnx_model_to_singa_net`。
+
+`_onnx_model_to_singa_net`的目的是获取SINGA的时序和运算符。tensors在ONNX中以其名称存储在字典中，而操作符则以`namedtuple('SingaOps', ['name', 'op', 'handle', 'forward'])`的形式存储在队列中。对于每个运算符，`name`是它的ONNX节点名称；`op`是ONNX节点；`forward`是SINGA运算符的转发函数；`handle`是为一些特殊的运算符准备的，如Conv和Pooling，它们有`handle`对象。
+
+`_onnx_model_to_singa_net`的第一步是调用`_init_graph_parameter`来获取模型内的所有tensors。对于可训练的权重，可以从`onnx_model.graph.initializer`中初始化`SINGA Tensor`。请注意，权重也可能存储在图的输入或称为`Constant`的ONNX节点中，SINGA也可以处理这些。
+
+虽然所有的权重都存储在ONNX模型中，但模型的输入是未知的，只有它的形状和类型。所以SINGA支持两种方式来初始化输入，1、根据其形状和类型生成随机张量，2、允许用户分配输入。第一种方法对大多数模型都很好，但是对于一些模型，比如BERT，矩阵的指数不能随机生成，否则会产生错误。
+
+然后，`_onnx_model_to_singa_net`迭代ONNX图中的所有节点，将其翻译成SIGNA运算符。另外，`_rename_operators` 定义了 SINGA 和 ONNX 之间的运算符名称映射。`_special_operators` 定义翻译运算符时要使用的函数。`_run_node`通过输入时序来运行生成的 SINGA 模型，并存储其输出时序，供以后的运算符使用。
+
+该类最后返回一个`SingaRep`对象，并在其中存储所有SINGA时序和运算符。
+
+### SingaRep
+
+`SingaBackend`存储所有的SINGA tensors和运算符。`run`接受模型的输入，并按照运算符队列逐个运行SINGA运算符。用户可以使用`last_layers`来决定是否将模型运行到最后几层。将 `all_outputs` 设置为 `False` 表示只得到最后的输出，设置为 `True` 表示也得到所有的中间输出。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/optimizer.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/optimizer.md
new file mode 100644
index 0000000..bd897f7
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/optimizer.md
@@ -0,0 +1,123 @@
+---
+id: version-3.1.0_Chinese-optimizer
+title: Optimizer
+original_id: optimizer
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA支持各种流行的优化器，包括动量随机梯度下降、Adam、RMSProp和AdaGrad等。对于每一种优化器，它都支持使用衰减调度器来安排不同时间段的学习率。优化器和衰减调度器包含在`singa/opt.py`中。
+
+## 创建一个优化器
+
+1. 带动量的SGD
+
+```python
+# define hyperparameter learning rate
+lr = 0.001
+# define hyperparameter momentum
+momentum = 0.9
+# define hyperparameter weight decay
+weight_decay = 0.0001
+
+from singa import opt
+sgd = opt.SGD(lr=lr, momentum=momentum, weight_decay=weight_decay)
+```
+
+2. RMSProp
+
+```python
+# define hyperparameter learning rate
+lr = 0.001
+# define hyperparameter rho
+rho = 0.9
+# define hyperparameter epsilon
+epsilon = 1e-8
+# define hyperparameter weight decay
+weight_decay = 0.0001
+
+from singa import opt
+sgd = opt.RMSProp(lr=lr, rho=rho, epsilon=epsilon, weight_decay=weight_decay)
+```
+
+3. AdaGrad
+
+```python
+# define hyperparameter learning rate
+lr = 0.001
+# define hyperparameter epsilon
+epsilon = 1e-8
+# define hyperparameter weight decay
+weight_decay = 0.0001
+
+from singa import opt
+sgd = opt.AdaGrad(lr=lr, epsilon=epsilon, weight_decay=weight_decay)
+```
+
+4. Adam
+
+```python
+# define hyperparameter learning rate
+lr = 0.001
+# define hyperparameter beta 1
+beta_1= 0.9
+# define hyperparameter beta 2
+beta_1= 0.999
+# define hyperparameter epsilon
+epsilon = 1e-8
+# define hyperparameter weight decay
+weight_decay = 0.0001
+
+from singa import opt
+sgd = opt.Adam(lr=lr, beta_1=beta_1, beta_2=beta_2, epsilon=epsilon, weight_decay=weight_decay)
+```
+
+## 创建一个衰减调度器
+
+```python
+from singa import opt
+
+# define initial learning rate
+lr_init = 0.001
+# define the rate of decay in the decay scheduler
+decay_rate = 0.95
+# define whether the learning rate schedule is a staircase shape
+staircase=True
+# define the decay step of the decay scheduler (in this example the lr is decreased at every 2 steps)
+decay_steps = 2
+
+# create the decay scheduler, the schedule of lr becomes lr_init * (decay_rate ^ (step // decay_steps) )
+lr = opt.ExponentialDecay(0.1, 2, 0.5, True)
+# Use the lr to create an optimizer
+sgd = opt.SGD(lr=lr, momentum=0.9, weight_decay=0.0001)
+```
+
+## 使用模型API中的优化器
+
+当我们创建模型时，我们需要将优化器附加到模型上：
+
+```python
+# create a CNN using the Model API
+model = CNN()
+
+# initialize optimizer and attach it to the model
+sgd = opt.SGD(lr=0.005, momentum=0.9, weight_decay=1e-5)
+model.set_optimizer(sgd)
+```
+
+然后，当我们调用模型时，它会运行利用优化器的 `train_one_batch` 方法。
+
+因此，一个迭代循环优化模型的例子是：
+
+```python
+for b in range(num_train_batch):
+ # generate the next mini-batch
+ x, y = ...
+
+ # Copy the data into input tensors
+ tx.copy_from_numpy(x)
+ ty.copy_from_numpy(y)
+
+ # Training with one batch
+ out, loss = model(tx, ty)
+```
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/security.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/security.md
new file mode 100644
index 0000000..ab6ae50
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/security.md
@@ -0,0 +1,9 @@
+---
+id: version-3.1.0_Chinese-security
+title: Security
+original_id: security
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+用户可以向[SINGA安全团队邮件列表](mailto:security@singa.apache.org)报告安全漏洞。
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/software-stack.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/software-stack.md
new file mode 100644
index 0000000..fdd2cc3
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/software-stack.md
@@ -0,0 +1,85 @@
+---
+id: version-3.1.0_Chinese-software-stack
+title: Software Stack
+original_id: software-stack
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA的软件栈包括两个主要层次，底层后端和Python接口层。图1展示了它们与硬件的关系。后端组件为深度学习模型提供基本的数据结构，为调度和执行操作提供硬件抽象，为分布式训练提供通信组件。Python接口封装了部分CPP数据结构，并为神经网络训练提供了额外的高级类，方便实现复杂的神经网络模型。
+
+接下来，我们以自下而上的方式介绍软件栈。
+
+ <br/> **Figure 1 - SINGA V3
+软件栈结构**
+
+## 底层后端
+
+### Device
+
+每个`Device`实例，即一个设备，是针对一个硬件设备创建的，例如GPU或CPU。`Device`管理数据结构的内存，并调度操作的执行，例如，在CUDA流或CPU线程上。根据硬件及其编程语言，SINGA实现了以下特定的设备类。
+
+- **CudaGPU** 代表Nvidia GPU卡,执行单元是CUDA流。
+- **CppCPU** 代表一个普通的CPU,执行单位是CPU线程。
+- **OpenclGPU** 代表Nvidia和AMD的普通GPU卡,执行单位是CommandQueue,鉴于OpenCL与许多硬件设备兼容，如FPGA和ARM，OpenclGPU可以扩展到其他设备。
+
+### Tensor
+
+`Tensor`类表示一个多维数组，它存储模型变量，例如，来自卷积层的输入图像和特征图。每个`Tensor`实例（即张量）都分配在一个设备上，它管理张量的内存，并针对张量调度（计算）操作。大多数机器学习算法可以使用（密集或稀疏）张量抽象及其操作来表达。因此，SINGA将能够运行广泛的模型，包括深度学习模型和其他传统的机器学习模型。
+
+### Operator
+
+针对张量的运算符有两类，一类是线性代数运算符，如矩阵乘法，另一类是神经网络特有的运算符，如卷积和池化。线性代数运算符以`Tensor`函数的形式提供，并针对不同的硬件设备分别实现。
+
+- CppMath (tensor_math_cpp.h) 为CppCPU使用Cpp实现张量运算。
+- CudaMath (tensor_math_cuda.h) 为CudaGPU使用CUDA实现张量运算。
+- OpenclMath(tensor_math_opencl.h) 使用OpenCL实现了OpenclGPU的张量运算。
+
+神经网络的具体运算符也是单独实现的，例如：
+
+- GpuConvFoward (convolution.h) 在Nvidia GPU上通过CuDNN实现卷积的前向函数。
+- CpuConvForward (convolution.h) 在CPU上使用CPP实现卷积的前向函数。
+
+通常情况下，用户会创建一个`Device`实例，并使用它来创建多个`Tensor`实例。当用户调用Tensor函数或神经网络操作时，会调用常驻设备的相应实现。换句话说，操作符的实现对用户是透明的。
+
+Tensor和Device抽象是可扩展的，以支持使用不同编程语言的各种硬件设备。一个新的硬件设备将通过添加一个新的Device子类和相应的操作符的实现来支持。
+
+在速度和内存方面的优化是由`Device`的`Scheduler`和`MemPool`完成的。例如，`Scheduler`根据运算符的依赖性创建一个[计算图](./graph)，然后它可以优化运算符的执行顺序，以实现并行性和内存共享。
+
+### Communicator
+
+`Communicator`是为了支持[分布式训练](./dist-train)。它使用套接字、MPI和NCCL实现通信协议。通常情况下，用户只需要调用`put()`和`get()`等高级API来发送和接收tensors。针对拓扑结构、消息大小等的通信优化在内部完成。
+
+## Python 接口层
+
+所有的后端组件都通过SWIG作为Python模块暴露出来。此外，还添加了以下类来支持复杂神经网络的实现：
+
+### Opt
+
+`Opt`及其子类实现了使用参数梯度更新模型参数值的方法（如SGD）。子类[DistOpt](./dist-train)通过调用`Communicator`中的方法来同步分布式训练的worker之间的梯度。
+
+### Operator
+
+`Operator`包装了多个使用后端Tensor或神经网络运算符实现的函数。例如，前向函数和后向函数`ReLU`组成`ReLU`运算符。
+
+### Layer
+
+`Layer`和它的子类用参数来包装运算符。例如，卷积和线性运算符有权重和偏置参数。 这些参数由相应的`Layer`类维护。
+
+### Autograd
+
+
+`Autograd`通过记录运算符的前向函数的执行情况，自动按照相反的顺序调用后向函数，实现了[反向模式的自动微分](https://rufflewind.com/2016-12-30/reverse-mode-automatic-differentiation)。所有的函数都可以通过`Scheduler`进行缓冲，建立[计算图](./graph)，以提高效率和内存优化。
+
+### Model
+
+
+[Model](./graph)提供了一个简单的接口来实现新的网络模型。你只需要继承`Model`，通过创建和调用层或运算符来定义模型的前向传播。当训练数据输入到`Model`中时，`Model`会自动进行autograd并通过`Opt`自动更新参数。通过`Model` API，SINGA充分利用了命令式编程和声明式编程的优势。用户使用[Model](./graph) API按照PyTorch那样的命令式编程风格来实现一个网络，PyTorch在每次迭代中都会重新创建操作，而SINGA则不同，它在第一次迭代后就会缓冲操作，隐式地创建一个计算图（当该功能被启用时）。该图类似于使用声明式编程的库所创建的图，例如TensorFlow。因此，SINGA可以在计算图上应用内存和速度优化技术。
+
+### ONNX
+
+
+为了支持ONNX，SINGA实现了一个[sonnx](./onnx)模块，其中包括。
+
+- SingaFrontend用于将SINGA模型保存为onnx格式。
+- SingaBackend 用于将 onnx 格式的模型加载到 SINGA 中进行训练和inference。
\ No newline at end of file
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/source-repository.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/source-repository.md
new file mode 100644
index 0000000..434e867
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/source-repository.md
@@ -0,0 +1,24 @@
+---
+id: version-3.1.0_Chinese-source-repository
+title: Source Repository
+original_id: source-repository
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+
+本项目使用[Git](http://git-scm.com/)来管理其源代码。关于Git的使用说明可以在 http://git-scm.com/documentation 找到。
+
+## Repository
+
+以下是在线资源库的链接：
+
+- https://gitbox.apache.org/repos/asf?p=singa.git
+
+Github镜像在：
+
+- https://github.com/apache/singa
+
+代码可以从任何一个repo中克隆，例如：
+
+ git clone https://github.com/apache/singa.git
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/team-list.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/team-list.md
new file mode 100644
index 0000000..294aa51
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/team-list.md
@@ -0,0 +1,57 @@
+---
+id: version-3.1.0_Chinese-team-list
+title: The SINGA Team
+original_id: team-list
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+一个成功的项目需要许多人扮演许多角色，有些成员编写代码或文档，而其他成员则作为测试人员，提交补丁和建议。
+
+SINGA社区的开发者主要来自新加坡国立大学、浙江大学、网易、大阪大学、yzBigData等。
+
+## PMC
+
+| Name | Email | Organization |
+| ------------- | ----------------------- | --------------------------------------------- |
+| Gang Chen | cg@apache.org | Zhejiang University |
+| Anh Dinh | dinhtta@apache.org | Singapore University of Technology and Design |
+| Ted Dunning | tdunning@apache.org | Apache Software Foundation |
+| Jinyang Gao | jinyang@apache.org | DAMO Academy, Alibaba Group |
+| Alan Gates | gates@apache.org | Apache Software Foundation |
+| Zhaojing Luo | zhaojing@apache.org | National University of Singapore |
+| Thejas Nair | thejas@apache.org | Apache Software Foundation |
+| Beng Chin Ooi | ooibc@apache.org | National University of Singapore |
+| Moaz Reyad | moaz@apache.org | Université Grenoble Alpes |
+| Kian-Lee Tan | tankianlee@apache.org | National University of Singapore |
+| Sheng Wang | wangsh@apache.org | DAMO Academy, Alibaba Group |
+| Wei Wang | wangwei@apache.org | National University of Singapore |
+| Zhongle Xie | zhongle@apache.org | National University of Singapore |
+| Sai Ho Yeung | chrishkchris@apache.org | National University of Singapore |
+| Meihui Zhang | meihuizhang@apache.org | Beijing Institute of Technology |
+| Kaiping Zheng | kaiping@apache.org | National University of Singapore |
+
+## Committers
+
+| Name | Email | Organization |
+| ------------ | ---------------------- | --------------------------------------------- |
+| Xiangrui Cai | caixr@apache.org | Nankai University |
+| Chonho Lee | chonho@apache.org | Osaka University |
+| Shicong Lin | shicong@apache.org | National University of Singapore |
+| Rulin Xing | rulin@apache.org | Huazhong University of Science and Technology |
+| Wanqi Xue | xuewanqi@apache.org | Nanyang Technological University |
+| Joddiy Zhang | joddiyzhang@apache.org | National University of Singapore |
+
+## Contributors
+
+| Name | Email | Organization |
+| ------------------ | ---------------------------- | -------------------------------- |
+| Haibo Chen | hzchenhaibo@corp.netease.com | NetEase |
+| Shicheng Chen | chengsc@comp.nus.edu.sg | National University of Singapore |
+| Xin Ji | vincent.j.xin@gmail.com | Visenze, Singapore |
+| Anthony K. H. Tung | atung@comp.nus.edu.sg | National University of Singapore |
+| Ji Wang | wangji@mzhtechnologies.com | Hangzhou MZH Technologies |
+| Yuan Wang | wangyuan@corp.netease.com | NetEase |
+| Wenfeng Wu | dcswuw@gmail.com | Freelancer, China |
+| Kaiyuan Yang | yangky@comp.nus.edu.sg | National University of Singapore |
+| Chang Yao | yaochang2009@gmail.com | Hangzhou MZH Technologies |
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/tensor.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/tensor.md
new file mode 100644
index 0000000..00e22a5
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/tensor.md
@@ -0,0 +1,245 @@
+---
+id: version-3.1.0_Chinese-tensor
+title: Tensor
+original_id: tensor
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+每个Tensor实例都是分配在特定设备实例上的多维数组。Tensor实例可以存储变量，并在不同类型的硬件设备上提供线性代数运算，而无需用户察觉。需要注意的是，除了复制函数外，用户需要确保张量操作数分配在同一个设备上。
+
+## Tensor用法
+
+### 创建Tensor
+
+```python
+>>> import numpy as np
+>>> from singa import tensor
+>>> tensor.from_numpy( np.asarray([[1, 0, 0], [0, 1, 0]], dtype=np.float32) )
+[[1. 0. 0.]
+ [0. 1. 0.]]
+```
+
+### 转换到numpy
+
+```python
+>>> a = np.asarray([[1, 0, 0], [0, 1, 0]], dtype=np.float32)
+>>> tensor.from_numpy(a)
+[[1. 0. 0.]
+ [0. 1. 0.]]
+>>> tensor.to_numpy(tensor.from_numpy(a))
+array([[1., 0., 0.],
+ [0., 1., 0.]], dtype=float32)
+```
+
+### Tensor方法
+
+```python
+>>> t = tensor.from_numpy(a)
+>>> t.transpose([1,0])
+[[1. 0.]
+ [0. 1.]
+ [0. 0.]]
+```
+
+`Tensor`变换，最多支持6维。
+
+```python
+>>> a = tensor.random((2,3,4,5,6,7))
+>>> a.shape
+(2, 3, 4, 5, 6, 7)
+>>> a.reshape((2,3,4,5,7,6)).transpose((3,2,1,0,4,5)).shape
+(5, 4, 3, 2, 7, 6)
+```
+
+### Tensor算术方法
+
+`tensor`是实时计算的：
+
+```python
+>>> t + 1
+[[2. 1. 1.]
+ [1. 2. 1.]]
+>>> t / 5
+[[0.2 0. 0. ]
+ [0. 0.2 0. ]]
+```
+
+`tensor` broadcasting运算:
+
+```python
+>>> a
+[[1. 2. 3.]
+ [4. 5. 6.]]
+>>> b
+[[1. 2. 3.]]
+>>> a + b
+[[2. 4. 6.]
+ [5. 7. 9.]]
+>>> a * b
+[[ 1. 4. 9.]
+ [ 4. 10. 18.]]
+>>> a / b
+[[1. 1. 1. ]
+ [4. 2.5 2. ]]
+>>> a/=b # inplace operation
+>>> a
+[[1. 1. 1. ]
+ [4. 2.5 2. ]]
+```
+
+`tensor` broadcasting矩阵乘法：
+
+```python
+>>> from singa import tensor
+>>> a = tensor.random((2,2,2,3))
+>>> b = tensor.random((2,3,4))
+>>> tensor.mult(a,b).shape
+(2, 2, 2, 4)
+```
+
+### Tensor函数
+
+`singa.tensor`模块中的函数在应用函数中定义的变换后返回新的`Tensor`对象。
+
+```python
+>>> tensor.log(t+1)
+[[0.6931472 0. 0. ]
+ [0. 0.6931472 0. ]]
+```
+
+### Tensor在不同Devices上
+
+`tensor`默认在主机(CPU)上创建；也可以通过指定设备在不同的硬件`device`上创建。`tensor`可以通过`to_device()`函数在`device`之间移动。
+
+```python
+>>> from singa import device
+>>> x = tensor.Tensor((2, 3), device.create_cuda_gpu())
+>>> x.gaussian(1,1)
+>>> x
+[[1.531889 1.0128608 0.12691343]
+ [2.1674204 3.083676 2.7421203 ]]
+>>> # move to host
+>>> x.to_device(device.get_default_device())
+```
+
+### 使用Tensor训练MLP
+
+```python
+
+"""
+ code snipet from examples/mlp/module.py
+"""
+
+label = get_label()
+data = get_data()
+
+dev = device.create_cuda_gpu_on(0)
+sgd = opt.SGD(0.05)
+
+# define tensor for input data and label
+tx = tensor.Tensor((400, 2), dev, tensor.float32)
+ty = tensor.Tensor((400,), dev, tensor.int32)
+model = MLP(data_size=2, perceptron_size=3, num_classes=2)
+
+# attached model to graph
+model.set_optimizer(sgd)
+model.compile([tx], is_train=True, use_graph=True, sequential=False)
+model.train()
+
+for i in range(1001):
+ tx.copy_from_numpy(data)
+ ty.copy_from_numpy(label)
+ out, loss = model(tx, ty, 'fp32', spars=None)
+
+ if i % 100 == 0:
+ print("training loss = ", tensor.to_numpy(loss)[0])
+```
+
+输出:
+
+```bash
+$ python3 examples/mlp/module.py
+training loss = 0.6158037
+training loss = 0.52852553
+training loss = 0.4571422
+training loss = 0.37274635
+training loss = 0.30146334
+training loss = 0.24906921
+training loss = 0.21128304
+training loss = 0.18390492
+training loss = 0.16362564
+training loss = 0.148164
+training loss = 0.13589878
+```
+
+## Tensor实现
+
+上一节介绍了`Tensor`的一般用法，下面将介绍其底层的实现。首先，将介绍Python和C++ tensors的设计。后面会讲到前端（Python）和后端（C++）如何连接，如何扩展。
+
+### Python Tensor
+
+Python类`Tensor`，定义在`python/singa/tensor.py`中，提供了高层的张量操作，用于实现深度学习操作（通过[autograd](./autograd)），以及终端用户的数据管理。
+
+它主要是通过简单地封装C++张量方法来工作，包括算术方法（如`sum`）和非算术方法（如`reshape`）。一些高级的算术运算以后会引入，并使用纯Python的张量API来实现，如`tensordot`。Python Tensor API可以利用灵活的方法轻松实现复杂的神经网络操作。
+
+### C++ Tensor
+
+C++类`Tensor`，定义在`include/singa/core/tensor.h`中，主要是管理存放数据的内存，并提供低级的API用于张量操作。同时，它还通过封装不同的后端（CUDA、BLAS、cuBLAS等）提供各种算术方法（如`matmul`）。
+
+#### 执行背景和内存块
+
+Tensor的两个重要概念或者说数据结构是执行背景`device`，和内存块`Block`。
+
+每个`Tensor`物理上存储在一个硬件设备上，并由硬件设备管理，代表执行背景（CPU、GPU），Tensor的数学计算是在设备上执行的。
+
+Tensor数据在`Block`实例中，定义在`include/singa/core/common.h`中。`Block`拥有底层数据，而tensors则在描述tensor的元数据上拥有所有权，比如`shape`、`stride`。
+
+#### Tensor数学后端
+
+为了利用不同后端硬件设备提供的高效数学库，SINGA为每个支持的后端设备提供了一套Tensor函数的实现。
+
+- 'tensor_math_cpp.h'为CppCPU设备使用Cpp（带CBLAS）实现操作。
+- 'tensor_math_cuda.h'为CudaGPU设备实现了使用Cuda(带cuBLAS)的操作。
+- 'tensor_math_opencl.h'为OpenclGPU设备实现了使用OpenCL的操作。
+
+### 将C++ APIs暴露给Python
+
+
+SWIG(http://www.swig.org/)是一个可以自动将C++ API转换为Python API的工具。SINGA使用SWIG将C++ APIs公开到Python中。SWIG会生成几个文件，包括`python/singa/singa_wrap.py`。Python模块(如`tensor`、`device`和`autograd`)导入这个模块来调用C++ API来实现Python类和函数。
+
+```python
+import tensor
+
+t = tensor.Tensor(shape=(2, 3))
+```
+
+例如，当按上面的方法创建Python `Tensor`实例时，`Tensor`类的实现会创建一个在`singa_wrap.py`中定义的Tensor类的实例，它对应于C++ `Tensor`类。为了清楚起见，`singa_wrap.py`中的`Tensor`类在`tensor.py`中被称为`CTensor`。
+
+```python
+# in tensor.py
+from . import singa_wrap as singa
+
+CTensor = singa.Tensor
+```
+
+### 创建新的Tensor函数
+
+
+有了前面的描述所奠定的基础，扩展张量函数可以用自下而上的方式轻松完成。对于数学运算，其步骤是：
+
+- 在`tensor.h`中声明新的API。
+- 使用 `tensor.cc` 中预定义的宏生成代码，参考 `GenUnaryTensorFn(Abs);` 。
+- 在`tensor_math.h`中声明template 方法/函数。
+- 至少在CPU(`tensor_math_cpp.h`)和GPU(`tensor_math_cuda.h`)中进行真正的实现。
+- 将 API 加入到 `src/api/core_tensor.i` 中，通过 SWIG 公开 API。
+- 通过调用 `singa_wrap.py` 中自动生成的函数，在 `tensor.py` 中定义 Python Tensor API。
+- 在适当的地方编写单元测试
+
+## Python API
+
+_进行中_
+
+## CPP API
+
+_进行中_
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/time-profiling.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/time-profiling.md
new file mode 100644
index 0000000..497a9da
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/time-profiling.md
@@ -0,0 +1,154 @@
+---
+id: version-3.1.0_Chinese-time-profiling
+title: Time Profiling
+original_id: time-profiling
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+SINGA支持对图中缓冲的每个运算符进行时间分析。为了利用时间分析功能，我们首先调用`device.SetVerbosity`方法来设置时间分析器的verbosity，然后调用`device.PrintTimeProfiling`来打印出时间分析的结果。
+
+## 设置时间分析Verbosity
+
+要使用时间分析功能，我们需要设置verbosity。有三个级别的verbosity。在默认值`verbosity == 0`的情况下，它不会进行任何时间分析。当我们设置`verbosity == 1`时，它将对前向和后向传播时间进行分析。当`verbosity == 2`时，它将对图中每个缓冲操作所花费的时间进行分析。
+
+以下是设置时间分析功能的示例代码：
+
+```python
+# create a device
+from singa import device
+dev = device.create_cuda_gpu()
+# set the verbosity
+verbosity = 2
+dev.SetVerbosity(verbosity)
+# optional: skip the first 5 iterations when profiling the time
+dev.SetSkipIteration(5)
+```
+
+那么，当我们在程序的最后完成训练后，我们就可以通过调用`device.PrintTimeProfiling`方法来打印时间分析结果。
+
+```python
+dev.PrintTimeProfiling()
+```
+
+## 不同Verbosity的输出示例
+
+我们可以运行ResNet的[示例](https://github.com/apache/singa/blob/master/examples/cnn/benchmark.py)，看看不同的verbosity设置的输出。
+
+1. `verbosity == 1`
+
+```
+Time Profiling:
+Forward Propagation Time : 0.0409127 sec
+Backward Propagation Time : 0.114813 sec
+```
+
+2. `verbosity == 2`
+
+```
+Time Profiling:
+OP_ID0. SetValue : 1.73722e-05 sec
+OP_ID1. cudnnConvForward : 0.000612724 sec
+OP_ID2. GpuBatchNormForwardTraining : 0.000559449 sec
+OP_ID3. ReLU : 0.000375004 sec
+OP_ID4. GpuPoolingForward : 0.000240041 sec
+OP_ID5. SetValue : 3.4176e-06 sec
+OP_ID6. cudnnConvForward : 0.000115619 sec
+OP_ID7. GpuBatchNormForwardTraining : 0.000150415 sec
+OP_ID8. ReLU : 9.95494e-05 sec
+OP_ID9. SetValue : 3.22432e-06 sec
+OP_ID10. cudnnConvForward : 0.000648668 sec
+OP_ID11. GpuBatchNormForwardTraining : 0.000149793 sec
+OP_ID12. ReLU : 9.92118e-05 sec
+OP_ID13. SetValue : 3.37728e-06 sec
+OP_ID14. cudnnConvForward : 0.000400953 sec
+OP_ID15. GpuBatchNormForwardTraining : 0.000572181 sec
+OP_ID16. SetValue : 3.21312e-06 sec
+OP_ID17. cudnnConvForward : 0.000398698 sec
+OP_ID18. GpuBatchNormForwardTraining : 0.00056836 sec
+OP_ID19. Add : 0.000542246 sec
+OP_ID20. ReLU : 0.000372783 sec
+OP_ID21. SetValue : 3.25312e-06 sec
+OP_ID22. cudnnConvForward : 0.000260731 sec
+OP_ID23. GpuBatchNormForwardTraining : 0.000149041 sec
+OP_ID24. ReLU : 9.9072e-05 sec
+OP_ID25. SetValue : 3.10592e-06 sec
+OP_ID26. cudnnConvForward : 0.000637481 sec
+OP_ID27. GpuBatchNormForwardTraining : 0.000152577 sec
+OP_ID28. ReLU : 9.90518e-05 sec
+OP_ID29. SetValue : 3.28224e-06 sec
+OP_ID30. cudnnConvForward : 0.000404586 sec
+OP_ID31. GpuBatchNormForwardTraining : 0.000569679 sec
+OP_ID32. Add : 0.000542291 sec
+OP_ID33. ReLU : 0.00037211 sec
+OP_ID34. SetValue : 3.13696e-06 sec
+OP_ID35. cudnnConvForward : 0.000261219 sec
+OP_ID36. GpuBatchNormForwardTraining : 0.000148281 sec
+OP_ID37. ReLU : 9.89299e-05 sec
+OP_ID38. SetValue : 3.25216e-06 sec
+OP_ID39. cudnnConvForward : 0.000633644 sec
+OP_ID40. GpuBatchNormForwardTraining : 0.000150711 sec
+OP_ID41. ReLU : 9.84902e-05 sec
+OP_ID42. SetValue : 3.18176e-06 sec
+OP_ID43. cudnnConvForward : 0.000402752 sec
+OP_ID44. GpuBatchNormForwardTraining : 0.000571523 sec
+OP_ID45. Add : 0.000542435 sec
+OP_ID46. ReLU : 0.000372539 sec
+OP_ID47. SetValue : 3.24672e-06 sec
+OP_ID48. cudnnConvForward : 0.000493054 sec
+OP_ID49. GpuBatchNormForwardTraining : 0.000293142 sec
+OP_ID50. ReLU : 0.000190047 sec
+OP_ID51. SetValue : 3.14784e-06 sec
+OP_ID52. cudnnConvForward : 0.00148837 sec
+OP_ID53. GpuBatchNormForwardTraining : 8.34794e-05 sec
+OP_ID54. ReLU : 5.23254e-05 sec
+OP_ID55. SetValue : 3.40096e-06 sec
+OP_ID56. cudnnConvForward : 0.000292971 sec
+OP_ID57. GpuBatchNormForwardTraining : 0.00029174 sec
+OP_ID58. SetValue : 3.3248e-06 sec
+OP_ID59. cudnnConvForward : 0.000590154 sec
+OP_ID60. GpuBatchNormForwardTraining : 0.000294149 sec
+OP_ID61. Add : 0.000275119 sec
+OP_ID62. ReLU : 0.000189268 sec
+OP_ID63. SetValue : 3.2704e-06 sec
+OP_ID64. cudnnConvForward : 0.000341232 sec
+OP_ID65. GpuBatchNormForwardTraining : 8.3304e-05 sec
+OP_ID66. ReLU : 5.23667e-05 sec
+OP_ID67. SetValue : 3.19936e-06 sec
+OP_ID68. cudnnConvForward : 0.000542484 sec
+OP_ID69. GpuBatchNormForwardTraining : 8.60537e-05 sec
+OP_ID70. ReLU : 5.2479e-05 sec
+OP_ID71. SetValue : 3.41824e-06 sec
+OP_ID72. cudnnConvForward : 0.000291295 sec
+OP_ID73. GpuBatchNormForwardTraining : 0.000292795 sec
+OP_ID74. Add : 0.000274438 sec
+OP_ID75. ReLU : 0.000189689 sec
+OP_ID76. SetValue : 3.21984e-06 sec
+OP_ID77. cudnnConvForward : 0.000338776 sec
+OP_ID78. GpuBatchNormForwardTraining : 8.484e-05 sec
+OP_ID79. ReLU : 5.29408e-05 sec
+OP_ID80. SetValue : 3.18208e-06 sec
+OP_ID81. cudnnConvForward : 0.000545542 sec
+OP_ID82. GpuBatchNormForwardTraining : 8.40976e-05 sec
+OP_ID83. ReLU : 5.2256e-05 sec
+OP_ID84. SetValue : 3.36256e-06 sec
+OP_ID85. cudnnConvForward : 0.000293003 sec
+OP_ID86. GpuBatchNormForwardTraining : 0.0002989 sec
+OP_ID87. Add : 0.000275041 sec
+OP_ID88. ReLU : 0.000189867 sec
+OP_ID89. SetValue : 3.1184e-06 sec
+OP_ID90. cudnnConvForward : 0.000340417 sec
+OP_ID91. GpuBatchNormForwardTraining : 8.39395e-05 sec
+OP_ID92. ReLU : 5.26544e-05 sec
+OP_ID93. SetValue : 3.2336e-06 sec
+OP_ID94. cudnnConvForward : 0.000539787 sec
+OP_ID95. GpuBatchNormForwardTraining : 8.2753e-05 sec
+OP_ID96. ReLU : 4.86758e-05 sec
+OP_ID97. SetValue : 3.24384e-06 sec
+OP_ID98. cudnnConvForward : 0.000287108 sec
+OP_ID99. GpuBatchNormForwardTraining : 0.000293127 sec
+OP_ID100. Add : 0.000269478 sec
+.
+.
+.
+```
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-cpu-dev.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-cpu-dev.md
new file mode 100644
index 0000000..cd9227d
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-cpu-dev.md
@@ -0,0 +1,13 @@
+---
+id: version-3.1.0_Chinese-wheel-cpu-dev
+title: CPU only (dev version)
+original_id: wheel-cpu-dev
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 3.0.0.dev200720
+
+- [Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720-cp36-cp36m-manylinux2014_x86_64.whl)
+- [Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720-cp37-cp37m-manylinux2014_x86_64.whl)
+- [Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720-cp38-cp38-manylinux2014_x86_64.whl)
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-cpu.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-cpu.md
new file mode 100644
index 0000000..46ced6d
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-cpu.md
@@ -0,0 +1,19 @@
+---
+id: version-3.1.0_Chinese-wheel-cpu
+title: CPU only
+original_id: wheel-cpu
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 3.1.0
+
+- [Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0-cp36-cp36m-manylinux2014_x86_64.whl)
+- [Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0-cp37-cp37m-manylinux2014_x86_64.whl)
+- [Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0-cp38-cp38-manylinux2014_x86_64.whl)
+
+## 3.0.0
+
+- [Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0-cp36-cp36m-manylinux2014_x86_64.whl)
+- [Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0-cp37-cp37m-manylinux2014_x86_64.whl)
+- [Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0-cp38-cp38-manylinux2014_x86_64.whl)
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-gpu-dev.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-gpu-dev.md
new file mode 100644
index 0000000..ab0f37e
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-gpu-dev.md
@@ -0,0 +1,13 @@
+---
+id: version-3.1.0_Chinese-wheel-gpu-dev
+title: CUDA enabled (dev version)
+original_id: wheel-gpu-dev
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 3.0.0.dev200720
+
+- [CUDA10.2, cuDNN 7.6.5, Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720%2Bcuda10.2-cp36-cp36m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720%2Bcuda10.2-cp37-cp37m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0.dev200720%2Bcuda10.2-cp38-cp38-manylinux2014_x86_64.whl)
diff --git a/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-gpu.md b/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-gpu.md
new file mode 100644
index 0000000..08bb4f9
--- /dev/null
+++ b/docs-site/website/versioned_docs/version-3.1.0_Chinese/wheel-gpu.md
@@ -0,0 +1,22 @@
+---
+id: version-3.1.0_Chinese-wheel-gpu
+title: CUDA enabled
+original_id: wheel-gpu
+---
+
+<!--- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->
+
+## 3.1.0
+
+- [CUDA10.2, cuDNN 7.6.5, Python
+ 3.6]https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0%2Bcuda10.2-cp36-cp36m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python
+ 3.7]https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0%2Bcuda10.2-cp37-cp37m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python
+ 3.8]https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.1.0%2Bcuda10.2-cp38-cp38-manylinux2014_x86_64.whl)
+
+## 3.0.0
+
+- [CUDA10.2, cuDNN 7.6.5, Python 3.6](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0%2Bcuda10.2-cp36-cp36m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python 3.7](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0%2Bcuda10.2-cp37-cp37m-manylinux2014_x86_64.whl)
+- [CUDA10.2, cuDNN 7.6.5, Python 3.8](https://singa-wheel.s3-ap-southeast-1.amazonaws.com/singa-3.0.0%2Bcuda10.2-cp38-cp38-manylinux2014_x86_64.whl)
diff --git a/docs-site/website/versioned_sidebars/version-2.0.0-sidebars.json b/docs-site/website/versioned_sidebars/version-2.0.0-sidebars.json
index 116f5d6..9359a63 100644
--- a/docs-site/website/versioned_sidebars/version-2.0.0-sidebars.json
+++ b/docs-site/website/versioned_sidebars/version-2.0.0-sidebars.json
@@ -22,7 +22,7 @@
"version-2.0.0-model-zoo-imagenet-vgg"
],
"Development": [
- "version-2.0.0-download-singa",
+ "version-2.0.0-downloads",
"version-2.0.0-build",
"version-2.0.0-contribute-code",
"version-2.0.0-contribute-docs",
diff --git a/docs-site/website/versioned_sidebars/version-3.0.0-sidebars.json b/docs-site/website/versioned_sidebars/version-3.0.0-sidebars.json
index 06aa553..54fa55b 100644
--- a/docs-site/website/versioned_sidebars/version-3.0.0-sidebars.json
+++ b/docs-site/website/versioned_sidebars/version-3.0.0-sidebars.json
@@ -14,7 +14,7 @@
"version-3.0.0-dist-train"
],
"Development": [
- "version-3.0.0-download-singa",
+ "version-3.0.0-downloads",
"version-3.0.0-build",
"version-3.0.0-contribute-code",
"version-3.0.0-contribute-docs",
diff --git a/docs-site/website/versioned_sidebars/version-3.0.0.rc1-sidebars.json b/docs-site/website/versioned_sidebars/version-3.0.0.rc1-sidebars.json
index 982c16a..1705bfa 100644
--- a/docs-site/website/versioned_sidebars/version-3.0.0.rc1-sidebars.json
+++ b/docs-site/website/versioned_sidebars/version-3.0.0.rc1-sidebars.json
@@ -13,7 +13,7 @@
"version-3.0.0.rc1-dist-train"
],
"Development": [
- "version-3.0.0.rc1-download-singa",
+ "version-3.0.0.rc1-downloads",
"version-3.0.0.rc1-build",
"version-3.0.0.rc1-contribute-code",
"version-3.0.0.rc1-contribute-docs",
diff --git a/docs-site/website/versioned_sidebars/version-3.1.0-sidebars.json b/docs-site/website/versioned_sidebars/version-3.1.0-sidebars.json
index dcd83c5..3331fc4 100644
--- a/docs-site/website/versioned_sidebars/version-3.1.0-sidebars.json
+++ b/docs-site/website/versioned_sidebars/version-3.1.0-sidebars.json
@@ -16,7 +16,7 @@
"version-3.1.0-time-profiling"
],
"Development": [
- "version-3.1.0-download-singa",
+ "version-3.1.0-downloads",
"version-3.1.0-build",
"version-3.1.0-contribute-code",
"version-3.1.0-contribute-docs",
@@ -32,6 +32,12 @@
"version-3.1.0-security",
"version-3.1.0-team-list",
"version-3.1.0-history-singa"
+ ],
+ "Wheel Package": [
+ "version-3.1.0-wheel-cpu",
+ "version-3.1.0-wheel-cpu-dev",
+ "version-3.1.0-wheel-gpu",
+ "version-3.1.0-wheel-gpu-dev"
]
}
}
diff --git a/docs-site/website/versioned_sidebars/version-3.1.0_Chinese-sidebars.json b/docs-site/website/versioned_sidebars/version-3.1.0_Chinese-sidebars.json
new file mode 100644
index 0000000..d07516a
--- /dev/null
+++ b/docs-site/website/versioned_sidebars/version-3.1.0_Chinese-sidebars.json
@@ -0,0 +1,43 @@
+{
+ "version-3.1.0_Chinese-docs": {
+ "Getting Started": [
+ "version-3.1.0_Chinese-installation",
+ "version-3.1.0_Chinese-software-stack",
+ "version-3.1.0_Chinese-examples"
+ ],
+ "Guides": [
+ "version-3.1.0_Chinese-device",
+ "version-3.1.0_Chinese-tensor",
+ "version-3.1.0_Chinese-autograd",
+ "version-3.1.0_Chinese-optimizer",
+ "version-3.1.0_Chinese-graph",
+ "version-3.1.0_Chinese-onnx",
+ "version-3.1.0_Chinese-dist-train",
+ "version-3.1.0_Chinese-time-profiling"
+ ],
+ "Development": [
+ "version-3.1.0_Chinese-downloads",
+ "version-3.1.0_Chinese-build",
+ "version-3.1.0_Chinese-contribute-code",
+ "version-3.1.0_Chinese-contribute-docs",
+ "version-3.1.0_Chinese-how-to-release",
+ "version-3.1.0_Chinese-git-workflow"
+ ]
+ },
+ "version-3.1.0_Chinese-community": {
+ "Community": [
+ "version-3.1.0_Chinese-source-repository",
+ "version-3.1.0_Chinese-mail-lists",
+ "version-3.1.0_Chinese-issue-tracking",
+ "version-3.1.0_Chinese-security",
+ "version-3.1.0_Chinese-team-list",
+ "version-3.1.0_Chinese-history-singa"
+ ],
+ "Wheel Package": [
+ "version-3.1.0_Chinese-wheel-cpu",
+ "version-3.1.0_Chinese-wheel-cpu-dev",
+ "version-3.1.0_Chinese-wheel-gpu",
+ "version-3.1.0_Chinese-wheel-gpu-dev"
+ ]
+ }
+}
diff --git a/docs-site/website/versions.json b/docs-site/website/versions.json
index 59b2d18..502197b 100644
--- a/docs-site/website/versions.json
+++ b/docs-site/website/versions.json
@@ -2,5 +2,6 @@
"3.1.0",
"3.0.0",
"3.0.0.rc1",
- "2.0.0"
+ "2.0.0",
+ "3.1.0_Chinese"
]
diff --git a/docs-site/website/versions_otherlang.json b/docs-site/website/versions_otherlang.json
new file mode 100644
index 0000000..b1ad820
--- /dev/null
+++ b/docs-site/website/versions_otherlang.json
@@ -0,0 +1,3 @@
+[
+ "3.1.0_Chinese"
+]
