Each Tensor instance is a multi-dimensional array allocated on a specific Device instance. Tensor instances store variables and provide linear algebra operations over different types of hardware devices without user awareness. Note that users need to make sure the tensor operands are allocated on the same device except copy functions.
>>> 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.]]
>>> 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)
>>> t = tensor.from_numpy(a) >>> t.transpose([1,0]) [[1. 0.] [0. 1.] [0. 0.]]
tensor
transformation up to 6 dims
>>> 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
is evaluated in real time.
>>> t + 1 [[2. 1. 1.] [1. 2. 1.]] >>> t / 5 [[0.2 0. 0. ] [0. 0.2 0. ]]
tensor
broadcasting arithmetic:
>>> 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 on matrix multiplication (GEMM)
>>> 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)
Functions in module singa.tensor
return new tensor
object after applying the transformation defined in the function.
>>> tensor.log(t+1) [[0.6931472 0. 0. ] [0. 0.6931472 0. ]]
tensor
is created on host (CPU) by default; it can also be created on different hardware devices by specifying the device
. A tensor
could be moved between device
s via to_device()
function.
>>> 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())
""" 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])
Output:
$ 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
The previous section shows the general usage of Tensor
, the implementation under the hood will be covered below. First, the design of Python and C++ tensors will be introduced. Later part will talk about how the frontend (Python) and backend (C++) are connected and how to extend them.
Python class Tensor
, defined in python/singa/tensor.py
, provides high level tensor manipulations for implementing deep learning operations (via autograd), as well as data management by end users.
It primarily works by simply wrapping around C++ tensor methods, both arithmetic (e.g. sum
) and non arithmetic methods (e.g. reshape
). Some advanced arithmetic operations are later introduced and implemented using pure Python tensor API, e.g. tensordot
. Python Tensor APIs could be used to implement complex neural network operations easily with the flexible methods available.
C++ class Tensor
, defined in include/singa/core/tensor.h
, primarily manages the memory that holds the data, and provides low level APIs for tensor manipulation. Also, it provides various arithmetic methods (e.g. matmul
) by wrapping different backends (CUDA, BLAS, cuBLAS, etc.).
Two important concepts or data structures for Tensor
are the execution context device
, and the memory block Block
.
Each Tensor
is physically stored on and managed by a hardware device, representing the execution context (CPU, GPU). Tensor math calculations are executed on the device.
Tensor data in a Block
instance, defined in include/singa/core/common.h
. Block
owns the underlying data, while tensors take ownership on the metadata describing the tensor, like shape
, strides
.
To leverage on the efficient math libraries provided by different backend hardware devices, SINGA has one set of implementations of Tensor functions for each supported backend.
SWIG(http://www.swig.org/) is a tool that can automatically convert C++ APIs into Python APIs. SINGA uses SWIG to expose the C++ APIs to Python. Several files are generated by SWIG, including python/singa/singa_wrap.py
. The Python modules (e.g., tensor
, device
and autograd
) imports this module to call the C++ APIs for implementing the Python classes and functions.
import tensor t = tensor.Tensor(shape=(2, 3))
For example, when a Python Tensor
instance is created as above, the Tensor
class implementation creates an instance of the Tensor
class defined in singa_wrap.py
, which corresponds to the C++ Tensor
class. For clarity, the Tensor
class in singa_wrap.py
is referred as CTensor
in tensor.py
.
# in tensor.py from . import singa_wrap as singa CTensor = singa.Tensor
With the groundwork set by the previous description, extending tensor functions could be done easily in a bottom up manner. For math operations, the steps are:
tensor.h
tensor.cc
, refer to GenUnaryTensorFn(Abs);
as an example.tensor_math.h
tensor_math_cpp.h
) and GPU(tensor_math_cuda.h
)src/api/core_tensor.i
tensor.py
by calling the automatically generated function in singa_wrap.py
work in progress
work in progress