Google Git
Sign in
apache / singa-site / 880502395b7c5f9727ef6c00e72bead0663f1fa3 / . / content / docs / 3.1.0 / onnx.html
blob: f1753ac515ed7eb430cd4b3360f18473d1f47bb5 [file] [log] [blame]
<!DOCTYPE html><html lang="en"><head><meta charSet="utf-8"/><meta http-equiv="X-UA-Compatible" content="IE=edge"/><title>ONNX · Apache SINGA</title><meta name="viewport" content="width=device-width"/><meta name="generator" content="Docusaurus"/><meta name="description" content="&lt;!--- 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 &quot;License&quot;); 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 &quot;AS IS&quot; 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. --&gt;"/><meta name="docsearch:version" content="3.1.0"/><meta name="docsearch:language" content="en"/><meta property="og:title" content="ONNX · Apache SINGA"/><meta property="og:type" content="website"/><meta property="og:url" content="https://singa.apache.org/"/><meta property="og:description" content="&lt;!--- 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 &quot;License&quot;); 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 &quot;AS IS&quot; 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. --&gt;"/><meta property="og:image" content="https://singa.apache.org/img/singa_twitter_banner.jpeg"/><meta name="twitter:card" content="summary"/><meta name="twitter:image" content="https://singa.apache.org/img/singa_twitter_banner.jpeg"/><link rel="shortcut icon" href="/img/favicon.ico"/><link rel="stylesheet" href="https://cdn.jsdelivr.net/docsearch.js/1/docsearch.min.css"/><link rel="stylesheet" href="//cdnjs.cloudflare.com/ajax/libs/highlight.js/9.12.0/styles/atom-one-dark.min.css"/><link rel="alternate" type="application/atom+xml" href="https://singa.apache.org/blog/atom.xml" title="Apache SINGA Blog ATOM Feed"/><link rel="alternate" type="application/rss+xml" href="https://singa.apache.org/blog/feed.xml" title="Apache SINGA Blog RSS Feed"/><link rel="stylesheet" href="https://fonts.googleapis.com/css?family=Source+Sans+Pro:400,400i,700"/><link rel="stylesheet" href="https://fonts.googleapis.com/css2?family=Baloo+Paaji+2&amp;family=Source+Sans+Pro:wght@200;300&amp;display=swap"/><script type="text/javascript" src="https://buttons.github.io/buttons.js"></script><script src="https://unpkg.com/vanilla-back-to-top@7.1.14/dist/vanilla-back-to-top.min.js"></script><script>
document.addEventListener('DOMContentLoaded', function() {
addBackToTop(
{"zIndex":100}
)
});
</script><script src="/js/scrollSpy.js"></script><link rel="stylesheet" href="/css/main.css"/><script src="/js/codetabs.js"></script></head><body class="sideNavVisible separateOnPageNav"><div class="fixedHeaderContainer"><div class="headerWrapper wrapper"><header><a href="/"><img class="logo" src="/img/singa.png" alt="Apache SINGA"/></a><a href="/versions"><h3>3.1.0</h3></a><div class="navigationWrapper navigationSlider"><nav class="slidingNav"><ul class="nav-site nav-site-internal"><li class="siteNavGroupActive"><a href="/docs/3.1.0/installation" target="_self">Docs</a></li><li class=""><a href="/docs/3.1.0/source-repository" target="_self">Community</a></li><li class=""><a href="/blog/" target="_self">News</a></li><li class=""><a href="https://apache-singa.readthedocs.io/en/latest/" target="_self">API</a></li><li class="navSearchWrapper reactNavSearchWrapper"><input type="text" id="search_input_react" placeholder="Search" title="Search"/></li><li class=""><a href="https://github.com/apache/singa" target="_self">GitHub</a></li></ul></nav></div></header></div></div><div class="navPusher"><div class="docMainWrapper wrapper"><div class="docsNavContainer" id="docsNav"><nav class="toc"><div class="toggleNav"><section class="navWrapper wrapper"><div class="navBreadcrumb wrapper"><div class="navToggle" id="navToggler"><div class="hamburger-menu"><div class="line1"></div><div class="line2"></div><div class="line3"></div></div></div><h2><i>›</i><span>Guides</span></h2><div class="tocToggler" id="tocToggler"><i class="icon-toc"></i></div></div><div class="navGroups"><div class="navGroup"><h3 class="navGroupCategoryTitle">Getting Started</h3><ul class=""><li class="navListItem"><a class="navItem" href="/docs/3.1.0/installation">Installation</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/software-stack">Software Stack</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/examples">Examples</a></li></ul></div><div class="navGroup"><h3 class="navGroupCategoryTitle">Guides</h3><ul class=""><li class="navListItem"><a class="navItem" href="/docs/3.1.0/device">Device</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/tensor">Tensor</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/autograd">Autograd</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/optimizer">Optimizer</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/graph">Model</a></li><li class="navListItem navListItemActive"><a class="navItem" href="/docs/3.1.0/onnx">ONNX</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/dist-train">Distributed Training</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/time-profiling">Time Profiling</a></li></ul></div><div class="navGroup"><h3 class="navGroupCategoryTitle">Development</h3><ul class=""><li class="navListItem"><a class="navItem" href="/docs/3.1.0/downloads">Download SINGA</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/build">Build SINGA from Source</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/contribute-code">How to Contribute Code</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/contribute-docs">How to Contribute to Documentation</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/how-to-release">How to Prepare a Release</a></li><li class="navListItem"><a class="navItem" href="/docs/3.1.0/git-workflow">Git Workflow</a></li></ul></div></div></section></div><script>
var coll = document.getElementsByClassName('collapsible');
var checkActiveCategory = true;
for (var i = 0; i < coll.length; i++) {
var links = coll[i].nextElementSibling.getElementsByTagName('*');
if (checkActiveCategory){
for (var j = 0; j < links.length; j++) {
if (links[j].classList.contains('navListItemActive')){
coll[i].nextElementSibling.classList.toggle('hide');
coll[i].childNodes[1].classList.toggle('rotate');
checkActiveCategory = false;
break;
}
}
}
coll[i].addEventListener('click', function() {
var arrow = this.childNodes[1];
arrow.classList.toggle('rotate');
var content = this.nextElementSibling;
content.classList.toggle('hide');
});
}
document.addEventListener('DOMContentLoaded', function() {
createToggler('#navToggler', '#docsNav', 'docsSliderActive');
createToggler('#tocToggler', 'body', 'tocActive');
var headings = document.querySelector('.toc-headings');
headings && headings.addEventListener('click', function(event) {
var el = event.target;
while(el !== headings){
if (el.tagName === 'A') {
document.body.classList.remove('tocActive');
break;
} else{
el = el.parentNode;
}
}
}, false);
function createToggler(togglerSelector, targetSelector, className) {
var toggler = document.querySelector(togglerSelector);
var target = document.querySelector(targetSelector);
if (!toggler) {
return;
}
toggler.onclick = function(event) {
event.preventDefault();
target.classList.toggle(className);
};
}
});
</script></nav></div><div class="container mainContainer docsContainer"><div class="wrapper"><div class="post"><header class="postHeader"><a class="edit-page-link button" href="https://github.com/apache/singa-doc/blob/master/docs-site/docs/onnx.md" target="_blank" rel="noreferrer noopener">Edit</a><h1 id="__docusaurus" class="postHeaderTitle">ONNX</h1></header><article><div><span><!--- 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. -->
<p><a href="https://onnx.ai/">ONNX</a> is an open representation format for machine learning
models, which enables AI developers to use models across different libraries and
tools. SINGA supports loading ONNX format models for training and inference, and
saving models defined using SINGA APIs (e.g., <a href="./module">Module</a>) into ONNX
format.</p>
<p>SINGA has been tested with the following
<a href="https://github.com/onnx/onnx/blob/master/docs/Versioning.md">version</a> of ONNX.</p>
<table>
<thead>
<tr><th>ONNX version</th><th>File format version</th><th>Opset version ai.onnx</th><th>Opset version ai.onnx.ml</th><th>Opset version ai.onnx.training</th></tr>
</thead>
<tbody>
<tr><td>1.6.0</td><td>6</td><td>11</td><td>2</td><td>-</td></tr>
</tbody>
</table>
<h2><a class="anchor" aria-hidden="true" id="general-usage"></a><a href="#general-usage" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>General usage</h2>
<h3><a class="anchor" aria-hidden="true" id="loading-an-onnx-model-into-singa"></a><a href="#loading-an-onnx-model-into-singa" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Loading an ONNX Model into SINGA</h3>
<p>After loading an ONNX model from disk by <code>onnx.load</code>, You only need to update
the batch-size of input using <code>tensor.PlaceHolder</code> after SINGA v3.0, the shape
of internal tensors will be inferred automatically.</p>
<p>Then, you should define a class inheriting from <code>sonnx.SONNXModel</code> and implement
two methods <code>forward</code> for forward work and <code>train_one_batch</code> for training work.
After you call <code>model.compile</code>, the SONNX iterates and translates all the nodes
within the ONNX model's graph into SINGA operators, loads all stored weights and
infers each intermediate tensor's shape.</p>
<pre><code class="hljs css language-python3"><span class="hljs-keyword">import</span> onnx
<span class="hljs-keyword">from</span> singa <span class="hljs-keyword">import</span> device
<span class="hljs-keyword">from</span> singa <span class="hljs-keyword">import</span> sonnx
<span class="hljs-class"><span class="hljs-keyword">class</span> <span class="hljs-title">MyModel</span><span class="hljs-params">(sonnx.SONNXModel)</span>:</span>
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">__init__</span><span class="hljs-params">(self, onnx_model)</span>:</span>
super(MyModel, self).__init__(onnx_model)
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">forward</span><span class="hljs-params">(self, *x)</span>:</span>
y = super(MyModel, self).forward(*x)
<span class="hljs-comment"># Since SINGA model returns the output as a list,</span>
<span class="hljs-comment"># if there is only one output,</span>
<span class="hljs-comment"># you just need to take the first element.</span>
<span class="hljs-keyword">return</span> y[<span class="hljs-number">0</span>]
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">train_one_batch</span><span class="hljs-params">(self, x, y)</span>:</span>
<span class="hljs-keyword">pass</span>
model_path = <span class="hljs-string">"PATH/To/ONNX/MODEL"</span>
onnx_model = onnx.load(model_path)
<span class="hljs-comment"># convert onnx model into SINGA model</span>
dev = device.create_cuda_gpu()
x = tensor.PlaceHolder(INPUT.shape, device=dev)
model = MyModel(onnx_model)
model.compile([x], is_train=<span class="hljs-literal">False</span>, use_graph=<span class="hljs-literal">True</span>, sequential=<span class="hljs-literal">True</span>)
</code></pre>
<h3><a class="anchor" aria-hidden="true" id="inference-singa-model"></a><a href="#inference-singa-model" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Inference SINGA model</h3>
<p>Once the model is created, you can do inference by calling <code>model.forward</code>. The
input and output must be SINGA <code>Tensor</code> instances.</p>
<pre><code class="hljs css language-python3"><span class="hljs-attr">x</span> = tensor.Tensor(device=dev, data=INPUT)
<span class="hljs-attr">y</span> = model.forward(x)
</code></pre>
<h3><a class="anchor" aria-hidden="true" id="saving-singa-model-into-onnx-format"></a><a href="#saving-singa-model-into-onnx-format" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Saving SINGA model into ONNX Format</h3>
<p>Given the input tensors and the output tensors generated by the operators the
model, you can trace back all internal operations. Therefore, a SINGA model is
defined by the input and outputs tensors. To export a SINGA model into ONNX
format, you just need to provide the input and output tensor list.</p>
<pre><code class="hljs css language-python3"># <span class="hljs-symbol">x</span> is the input tensor, <span class="hljs-symbol">y</span> is the output tensor
sonnx.to_onnx([<span class="hljs-symbol">x</span>], [<span class="hljs-symbol">y</span>])
</code></pre>
<h3><a class="anchor" aria-hidden="true" id="re-training-an-onnx-model"></a><a href="#re-training-an-onnx-model" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Re-training an ONNX model</h3>
<p>To train (or refine) an ONNX model using SINGA, you need to implement the
<code>train_one_batch</code> from <code>sonnx.SONNXModel</code> and mark the <code>is_train=True</code> when
calling <code>model.compile</code>.</p>
<pre><code class="hljs css language-python3">from singa import opt
from singa import autograd
<span class="hljs-class"><span class="hljs-keyword">class</span> <span class="hljs-title">MyModel</span>(<span class="hljs-title">sonnx</span>.<span class="hljs-title">SONNXModel</span>):</span>
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">__init__</span><span class="hljs-params">(<span class="hljs-keyword">self</span>, onnx_model)</span></span>:
<span class="hljs-keyword">super</span>(MyModel, <span class="hljs-keyword">self</span>).__init_<span class="hljs-number">_</span>(onnx_model)
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">forward</span><span class="hljs-params">(<span class="hljs-keyword">self</span>, *x)</span></span>:
y = <span class="hljs-keyword">super</span>(MyModel, <span class="hljs-keyword">self</span>).forward(*x)
<span class="hljs-keyword">return</span> y[<span class="hljs-number">0</span>]
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">train_one_batch</span><span class="hljs-params">(<span class="hljs-keyword">self</span>, x, y, dist_option, spars)</span></span>:
out = <span class="hljs-keyword">self</span>.forward(x)
loss = autograd.softmax_cross_entropy(out, y)
<span class="hljs-keyword">if</span> dist_option == <span class="hljs-string">'fp32'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_update(loss)
elif dist_option == <span class="hljs-string">'fp16'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_update_half(loss)
elif dist_option == <span class="hljs-string">'partialUpdate'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_partial_update(loss)
elif dist_option == <span class="hljs-string">'sparseTopK'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_sparse_update(loss,
topK=True,
spars=spars)
elif dist_option == <span class="hljs-string">'sparseThreshold'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_sparse_update(loss,
topK=False,
spars=spars)
<span class="hljs-keyword">return</span> out, loss
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">set_optimizer</span><span class="hljs-params">(<span class="hljs-keyword">self</span>, optimizer)</span></span>:
<span class="hljs-keyword">self</span>.optimizer = optimizer
sgd = opt.SGD(lr=<span class="hljs-number">0</span>.<span class="hljs-number">005</span>, momentum=<span class="hljs-number">0</span>.<span class="hljs-number">9</span>, weight_decay=<span class="hljs-number">1</span>e-<span class="hljs-number">5</span>)
model.set_optimizer(sgd)
model.compile([tx], is_train=True, use_graph=graph, sequential=True)
</code></pre>
<h3><a class="anchor" aria-hidden="true" id="transfer-learning-an-onnx-model"></a><a href="#transfer-learning-an-onnx-model" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Transfer-learning an ONNX model</h3>
<p>You also can append some layers to the end of the ONNX model to do
transfer-learning. The <code>last_layers</code> accept a negative integer indicating the
layer to cut off from. For example, <code>-1</code> means cut off after the final output(do
not cut off any layer), <code>-2</code> means you cut off after the last second layer.</p>
<pre><code class="hljs css language-python3">from singa import opt
from singa import autograd
<span class="hljs-class"><span class="hljs-keyword">class</span> <span class="hljs-title">MyModel</span>(<span class="hljs-title">sonnx</span>.<span class="hljs-title">SONNXModel</span>):</span>
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">__init__</span><span class="hljs-params">(<span class="hljs-keyword">self</span>, onnx_model)</span></span>:
<span class="hljs-keyword">super</span>(MyModel, <span class="hljs-keyword">self</span>).__init_<span class="hljs-number">_</span>(onnx_model)
<span class="hljs-keyword">self</span>.linear = layer.Linear(<span class="hljs-number">1000</span>, <span class="hljs-number">3</span>)
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">forward</span><span class="hljs-params">(<span class="hljs-keyword">self</span>, *x)</span></span>:
<span class="hljs-comment"># cut off after the last third layer</span>
<span class="hljs-comment"># and append a linear layer</span>
y = <span class="hljs-keyword">super</span>(MyModel, <span class="hljs-keyword">self</span>).forward(*x, last_layers=-<span class="hljs-number">3</span>)[<span class="hljs-number">0</span>]
y = <span class="hljs-keyword">self</span>.linear(y)
<span class="hljs-keyword">return</span> y
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">train_one_batch</span><span class="hljs-params">(<span class="hljs-keyword">self</span>, x, y, dist_option, spars)</span></span>:
out = <span class="hljs-keyword">self</span>.forward(x)
loss = autograd.softmax_cross_entropy(out, y)
<span class="hljs-keyword">if</span> dist_option == <span class="hljs-string">'fp32'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_update(loss)
elif dist_option == <span class="hljs-string">'fp16'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_update_half(loss)
elif dist_option == <span class="hljs-string">'partialUpdate'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_partial_update(loss)
elif dist_option == <span class="hljs-string">'sparseTopK'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_sparse_update(loss,
topK=True,
spars=spars)
elif dist_option == <span class="hljs-string">'sparseThreshold'</span>:
<span class="hljs-keyword">self</span>.optimizer.backward_and_sparse_update(loss,
topK=False,
spars=spars)
<span class="hljs-keyword">return</span> out, loss
<span class="hljs-function"><span class="hljs-keyword">def</span> <span class="hljs-title">set_optimizer</span><span class="hljs-params">(<span class="hljs-keyword">self</span>, optimizer)</span></span>:
<span class="hljs-keyword">self</span>.optimizer = optimizer
sgd = opt.SGD(lr=<span class="hljs-number">0</span>.<span class="hljs-number">005</span>, momentum=<span class="hljs-number">0</span>.<span class="hljs-number">9</span>, weight_decay=<span class="hljs-number">1</span>e-<span class="hljs-number">5</span>)
model.set_optimizer(sgd)
model.compile([tx], is_train=True, use_graph=graph, sequential=True)
</code></pre>
<h2><a class="anchor" aria-hidden="true" id="onnx-model-zoo"></a><a href="#onnx-model-zoo" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>ONNX model zoo</h2>
<p>The <a href="https://github.com/onnx/models">ONNX Model Zoo</a> is a collection of
pre-trained, state-of-the-art models in the ONNX format contributed by community
members. SINGA has supported several CV and NLP models now. More models are
going to be supported soon.</p>
<h3><a class="anchor" aria-hidden="true" id="image-classification"></a><a href="#image-classification" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Image Classification</h3>
<p>This collection of models take images as input, then classifies the major
objects in the images into 1000 object categories such as keyboard, mouse,
pencil, and many animals.</p>
<table>
<thead>
<tr><th>Model Class</th><th>Reference</th><th>Description</th><th>Link</th></tr>
</thead>
<tbody>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/vision/classification/mobilenet">MobileNet</a></b></td><td><a href="https://arxiv.org/abs/1801.04381">Sandler et al.</a></td><td>Light-weight deep neural network best suited for mobile and embedded vision applications. <br>Top-5 error from paper - ~10%</td><td><a href="https://colab.research.google.com/drive/1HsixqJMIpKyEPhkbB8jy7NwNEFEAUWAf"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a></td></tr>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/vision/classification/resnet">ResNet18</a></b></td><td><a href="https://arxiv.org/abs/1512.03385">He et al.</a></td><td>A CNN model (up to 152 layers). Uses shortcut connections to achieve higher accuracy when classifying images. <br> Top-5 error from paper - ~3.6%</td><td><a href="https://colab.research.google.com/drive/1u1RYefSsVbiP4I-5wiBKHjsT9L0FxLm9"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a></td></tr>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/vision/classification/vgg">VGG16</a></b></td><td><a href="https://arxiv.org/abs/1409.1556">Simonyan et al.</a></td><td>Deep CNN model(up to 19 layers). Similar to AlexNet but uses multiple smaller kernel-sized filters that provides more accuracy when classifying images. <br>Top-5 error from paper - ~8%</td><td><a href="https://colab.research.google.com/drive/14kxgRKtbjPCKKsDJVNi3AvTev81Gp_Ds"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a></td></tr>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/vision/classification/shufflenet">ShuffleNet_V2</a></b></td><td><a href="https://arxiv.org/pdf/1707.01083.pdf">Simonyan et al.</a></td><td>Extremely computation efficient CNN model that is designed specifically for mobile devices. This network architecture design considers direct metric such as speed, instead of indirect metric like FLOP. Top-1 error from paper - ~30.6%</td><td>[<img src="https://colab.research.google.com/drive/19HfRu3YHP_H2z3BcZujVFRp23_J5XsuA?usp=sharing" alt="Open In Colab"></td></tr>
</tbody>
</table>
<p>We also give some re-training examples by using VGG and ResNet, please check
<code>examples/onnx/training</code>.</p>
<h3><a class="anchor" aria-hidden="true" id="object-detection"></a><a href="#object-detection" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Object Detection</h3>
<p>Object detection models detect the presence of multiple objects in an image and
segment out areas of the image where the objects are detected.</p>
<table>
<thead>
<tr><th>Model Class</th><th>Reference</th><th>Description</th><th>Link</th></tr>
</thead>
<tbody>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/tiny_yolov2">Tiny YOLOv2</a></b></td><td><a href="https://arxiv.org/pdf/1612.08242.pdf">Redmon et al.</a></td><td>A real-time CNN for object detection that detects 20 different classes. A smaller version of the more complex full YOLOv2 network.</td><td><a href="https://colab.research.google.com/drive/11V4I6cRjIJNUv5ZGsEGwqHuoQEie6b1T"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a></td></tr>
</tbody>
</table>
<h3><a class="anchor" aria-hidden="true" id="face-analysis"></a><a href="#face-analysis" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Face Analysis</h3>
<p>Face detection models identify and/or recognize human faces and emotions in
given images.</p>
<table>
<thead>
<tr><th>Model Class</th><th>Reference</th><th>Description</th><th>Link</th></tr>
</thead>
<tbody>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/vision/body_analysis/arcface">ArcFace</a></b></td><td><a href="https://arxiv.org/abs/1801.07698">Deng et al.</a></td><td>A CNN based model for face recognition which learns discriminative features of faces and produces embeddings for input face images.</td><td><a href="https://colab.research.google.com/drive/1qanaqUKGIDtifdzEzJOHjEj4kYzA9uJC"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a></td></tr>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/vision/body_analysis/emotion_ferplus">Emotion FerPlus</a></b></td><td><a href="https://arxiv.org/abs/1608.01041">Barsoum et al.</a></td><td>Deep CNN for emotion recognition trained on images of faces.</td><td><a href="https://colab.research.google.com/drive/1XHtBQGRhe58PDi4LGYJzYueWBeWbO23r"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a></td></tr>
</tbody>
</table>
<h3><a class="anchor" aria-hidden="true" id="machine-comprehension"></a><a href="#machine-comprehension" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Machine Comprehension</h3>
<p>This subset of natural language processing models that answer questions about a
given context paragraph.</p>
<table>
<thead>
<tr><th>Model Class</th><th>Reference</th><th>Description</th><th>Link</th></tr>
</thead>
<tbody>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/text/machine_comprehension/bert-squad">BERT-Squad</a></b></td><td><a href="https://arxiv.org/pdf/1810.04805.pdf">Devlin et al.</a></td><td>This model answers questions based on the context of the given input paragraph.</td><td><a href="https://colab.research.google.com/drive/1kud-lUPjS_u-TkDAzihBTw0Vqr0FjCE-"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a></td></tr>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/text/machine_comprehension/roberta">RoBERTa</a></b></td><td><a href="https://arxiv.org/pdf/1907.11692.pdf">Devlin et al.</a></td><td>A large transformer-based model that predicts sentiment based on given input text.</td><td><a href="https://colab.research.google.com/drive/1F-c4LJSx3Cb2jW6tP7f8nAZDigyLH6iN?usp=sharing"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a></td></tr>
<tr><td><b><a href="https://github.com/onnx/models/tree/master/text/machine_comprehension/gpt-2">GPT-2</a></b></td><td><a href="https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf">Devlin et al.</a></td><td>A large transformer-based language model that given a sequence of words within some text, predicts the next word.</td><td>[<img src="https://colab.research.google.com/drive/1ZlXLSIMppPch6HgzKRillJiUcWn3PiK7?usp=sharing" alt="Open In Colab"></td></tr>
</tbody>
</table>
<h2><a class="anchor" aria-hidden="true" id="supported-operators"></a><a href="#supported-operators" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Supported operators</h2>
<p>The following operators are supported:</p>
<ul>
<li>Acos</li>
<li>Acosh</li>
<li>Add</li>
<li>And</li>
<li>Asin</li>
<li>Asinh</li>
<li>Atan</li>
<li>Atanh</li>
<li>AveragePool</li>
<li>BatchNormalization</li>
<li>Cast</li>
<li>Ceil</li>
<li>Clip</li>
<li>Concat</li>
<li>ConstantOfShape</li>
<li>Conv</li>
<li>Cos</li>
<li>Cosh</li>
<li>Div</li>
<li>Dropout</li>
<li>Elu</li>
<li>Equal</li>
<li>Erf</li>
<li>Expand</li>
<li>Flatten</li>
<li>Gather</li>
<li>Gemm</li>
<li>GlobalAveragePool</li>
<li>Greater</li>
<li>HardSigmoid</li>
<li>Identity</li>
<li>LeakyRelu</li>
<li>Less</li>
<li>Log</li>
<li>MatMul</li>
<li>Max</li>
<li>MaxPool</li>
<li>Mean</li>
<li>Min</li>
<li>Mul</li>
<li>Neg</li>
<li>NonZero</li>
<li>Not</li>
<li>OneHot</li>
<li>Or</li>
<li>Pad</li>
<li>Pow</li>
<li>PRelu</li>
<li>Reciprocal</li>
<li>ReduceMean</li>
<li>ReduceSum</li>
<li>Relu</li>
<li>Reshape</li>
<li>ScatterElements</li>
<li>Selu</li>
<li>Shape</li>
<li>Sigmoid</li>
<li>Sign</li>
<li>Sin</li>
<li>Sinh</li>
<li>Slice</li>
<li>Softmax</li>
<li>Softplus</li>
<li>Softsign</li>
<li>Split</li>
<li>Sqrt</li>
<li>Squeeze</li>
<li>Sub</li>
<li>Sum</li>
<li>Tan</li>
<li>Tanh</li>
<li>Tile</li>
<li>Transpose</li>
<li>Unsqueeze</li>
<li>Upsample</li>
<li>Where</li>
<li>Xor</li>
</ul>
<h3><a class="anchor" aria-hidden="true" id="special-comments-for-onnx-backend"></a><a href="#special-comments-for-onnx-backend" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Special comments for ONNX backend</h3>
<ul>
<li><p>Conv, MaxPool and AveragePool</p>
<p>Input must be 1d<code>(N*C*H)</code> and 2d(<code>N*C*H*W</code>) shape and <code>dilation</code> must be 1.</p></li>
<li><p>BatchNormalization</p>
<p><code>epsilon</code> is 1e-05 and cannot be changed.</p></li>
<li><p>Cast</p>
<p>Only support float32 and int32, other types are casted to these two types.</p></li>
<li><p>Squeeze and Unsqueeze</p>
<p>If you encounter errors when you <code>Squeeze</code> or <code>Unsqueeze</code> between <code>Tensor</code> and
Scalar, please report to us.</p></li>
<li><p>Empty tensor Empty tensor is illegal in SINGA.</p></li>
</ul>
<h2><a class="anchor" aria-hidden="true" id="implementation"></a><a href="#implementation" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Implementation</h2>
<p>The code of SINGA ONNX locates at <code>python/singa/soonx.py</code>. There are four main
class, <code>SingaFrontend</code>, <code>SingaBackend</code>, <code>SingaRep</code> and <code>SONNXModel</code>.
<code>SingaFrontend</code> translates a SINGA model to an ONNX model; <code>SingaBackend</code>
translates an ONNX model to <code>SingaRep</code> object which stores all SINGA operators
and tensors(the tensor in this doc means SINGA <code>Tensor</code>); <code>SingaRep</code> can be run
like a SINGA model. <code>SONNXModel</code> inherits from <code>model.Model</code> which defines a
unified API for SINGA.</p>
<h3><a class="anchor" aria-hidden="true" id="singafrontend"></a><a href="#singafrontend" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>SingaFrontend</h3>
<p>The entry function of <code>SingaFrontend</code> is <code>singa_to_onnx_model</code> which also is
called <code>to_onnx</code>. <code>singa_to_onnx_model</code> creates the ONNX model, and it also
create a ONNX graph by using <code>singa_to_onnx_graph</code>.</p>
<p><code>singa_to_onnx_graph</code> accepts the output of the model, and recursively iterate
the SINGA model's graph from the output to get all operators to form a queue.
The input and intermediate tensors, i.e, trainable weights, of the SINGA model
is picked up at the same time. The input is stored in <code>onnx_model.graph.input</code>;
the output is stored in <code>onnx_model.graph.output</code>; and the trainable weights are
stored in <code>onnx_model.graph.initializer</code>.</p>
<p>Then the SINGA operator in the queue is translated to ONNX operators one by one.
<code>_rename_operators</code> defines the operators name mapping between SINGA and ONNX.
<code>_special_operators</code> defines which function to be used to translate the
operator.</p>
<p>In addition, some operators in SINGA has different definition with ONNX, that
is, ONNX regards some attributes of SINGA operators as input, so
<code>_unhandled_operators</code> defines which function to handle the special operator.</p>
<p>Since the bool type is regarded as int32 in SINGA, <code>_bool_operators</code> defines the
operators to be changed as bool type.</p>
<h3><a class="anchor" aria-hidden="true" id="singabackend"></a><a href="#singabackend" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>SingaBackend</h3>
<p>The entry function of <code>SingaBackend</code> is <code>prepare</code> which checks the version of
ONNX model and call <code>_onnx_model_to_singa_ops</code> then.</p>
<p>The purpose of <code>_onnx_model_to_singa_ops</code> is to get SINGA tensors and operators.
The tensors are stored in a dictionary by their name in ONNX, and operators are
stored in queue by the form of <code>namedtuple('SingaOps', ['node', 'operator'])</code>.
For each operator, <code>node</code> is an instance from OnnxNode which is defined to store
some basic information for an ONNX node; <code>operator</code> is the SINGA operator's
forward function;</p>
<p>The first step of <code>_onnx_model_to_singa_ops</code> has four steps, the first one is to
call <code>_parse_graph_params</code> to get all tensors stored as <code>params</code>. Then call
<code>_parse_graph_inputs_outputs</code> to get all input and output information stores as
<code>inputs</code> and <code>outputs</code>. Finally, it iterators all nodes within the ONNX graph
and parses it by <code>_onnx_node_to_singa_op</code> as SIGNA operators or layers and store
them as <code>outputs</code>. Some weights are stored within an ONNX node called
<code>Constant</code>, SONNX can handle them by <code>_onnx_constant_to_np</code> to store it into
<code>params</code>.</p>
<p>This class finally return a <code>SingaRep</code> object and stores above <code>params</code>,
<code>inputs</code>, <code>outputs</code>, <code>layers</code>.</p>
<h3><a class="anchor" aria-hidden="true" id="singarep"></a><a href="#singarep" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>SingaRep</h3>
<p><code>SingaBackend</code> stores all SINGA tensors and operators. <code>run</code> accepts the input
of the model and runs the SINGA operators one by one following the operators'
queue. The user can use <code>last_layers</code> to cut off the model after the last few
layers.</p>
<h3><a class="anchor" aria-hidden="true" id="sonnxmodel"></a><a href="#sonnxmodel" aria-hidden="true" class="hash-link"><svg class="hash-link-icon" aria-hidden="true" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>SONNXModel</h3>
<p><code>SONNXModel</code> inherits from <code>sonnx.SONNXModel</code> and implements the method
<code>forward</code> to provide a unified API with other SINGA models.</p>
</span></div></article></div><div class="docLastUpdate"><em>Last updated on 11/25/2020</em></div><div class="docs-prevnext"><a class="docs-prev button" href="/docs/3.1.0/graph"><span class="arrow-prev">← </span><span>Model</span></a><a class="docs-next button" href="/docs/3.1.0/dist-train"><span>Distributed Training</span><span class="arrow-next"> →</span></a></div></div></div><nav class="onPageNav"><ul class="toc-headings"><li><a href="#general-usage">General usage</a><ul class="toc-headings"><li><a href="#loading-an-onnx-model-into-singa">Loading an ONNX Model into SINGA</a></li><li><a href="#inference-singa-model">Inference SINGA model</a></li><li><a href="#saving-singa-model-into-onnx-format">Saving SINGA model into ONNX Format</a></li><li><a href="#re-training-an-onnx-model">Re-training an ONNX model</a></li><li><a href="#transfer-learning-an-onnx-model">Transfer-learning an ONNX model</a></li></ul></li><li><a href="#onnx-model-zoo">ONNX model zoo</a><ul class="toc-headings"><li><a href="#image-classification">Image Classification</a></li><li><a href="#object-detection">Object Detection</a></li><li><a href="#face-analysis">Face Analysis</a></li><li><a href="#machine-comprehension">Machine Comprehension</a></li></ul></li><li><a href="#supported-operators">Supported operators</a><ul class="toc-headings"><li><a href="#special-comments-for-onnx-backend">Special comments for ONNX backend</a></li></ul></li><li><a href="#implementation">Implementation</a><ul class="toc-headings"><li><a href="#singafrontend">SingaFrontend</a></li><li><a href="#singabackend">SingaBackend</a></li><li><a href="#singarep">SingaRep</a></li><li><a href="#sonnxmodel">SONNXModel</a></li></ul></li></ul></nav></div><footer class="nav-footer" id="footer"><section class="sitemap"><a href="/" class="nav-home"><img src="/img/singa-logo-square.png" alt="Apache SINGA" width="66" height="58"/></a><div><h5>Docs</h5><a href="/docs/installation">Getting Started</a><a href="/docs/device">Guides</a><a href="/en/https://apache-singa.readthedocs.io/en/latest/">API Reference</a><a href="/docs/examples">Examples</a><a href="/docs/download-singa">Development</a></div><div><h5>Community</h5><a href="/en/users.html">User Showcase</a><a href="/docs/history-singa">SINGA History</a><a href="/docs/team-list">SINGA Team</a><a href="/blog">SINGA News</a><a href="https://github.com/apache/singa">GitHub</a><div class="social"><a class="github-button" href="https://github.com/apache/singa" data-count-href="/apache/singa/stargazers" data-show-count="true" data-count-aria-label="# stargazers on GitHub" aria-label="Star this project on GitHub">apache/singa-doc</a></div><div class="social"><a href="https://twitter.com/ApacheSINGA" class="twitter-follow-button">Follow @ApacheSINGA</a></div></div><div><h5>Apache Software Foundation</h5><a href="https://apache.org/" target="_blank" rel="noreferrer noopener">Foundation</a><a href="http://www.apache.org/licenses/" target="_blank" rel="noreferrer noopener">License</a><a href="http://www.apache.org/foundation/sponsorship.html" target="_blank" rel="noreferrer noopener">Sponsorship</a><a href="http://www.apache.org/foundation/thanks.html" target="_blank" rel="noreferrer noopener">Thanks</a><a href="http://www.apache.org/events/current-event" target="_blank" rel="noreferrer noopener">Events</a><a href="http://www.apache.org/security/" target="_blank" rel="noreferrer noopener">Security</a></div></section><div style="width:100%;text-align:center"><a href="https://apache.org/" target="_blank" rel="noreferrer noopener" class="ApacheOpenSource"><img src="/img/asf_logo_wide.svg" alt="Apache Open Source"/></a><section class="copyright" style="max-width:60%;margin:0 auto">Copyright © 2023
The Apache Software Foundation. All rights reserved.
Apache SINGA, Apache, the Apache feather logo, and
the Apache SINGA project logos are trademarks of The
Apache Software Foundation. All other marks mentioned
may be trademarks or registered trademarks of their
respective owners.</section></div></footer></div><script type="text/javascript" src="https://cdn.jsdelivr.net/docsearch.js/1/docsearch.min.js"></script><script>window.twttr=(function(d,s, id){var js,fjs=d.getElementsByTagName(s)[0],t=window.twttr||{};if(d.getElementById(id))return t;js=d.createElement(s);js.id=id;js.src='https://platform.twitter.com/widgets.js';fjs.parentNode.insertBefore(js, fjs);t._e = [];t.ready = function(f) {t._e.push(f);};return t;}(document, 'script', 'twitter-wjs'));</script><script>
document.addEventListener('keyup', function(e) {
if (e.target !== document.body) {
return;
}
// keyCode for '/' (slash)
if (e.keyCode === 191) {
const search = document.getElementById('search_input_react');
search && search.focus();
}
});
</script><script>
var search = docsearch({
apiKey: '45202133606c0b5fa6d21cddc4725dd8',
indexName: 'apache_singa',
inputSelector: '#search_input_react',
algoliaOptions: {"facetFilters":["language:en","version:3.0.0"]}
});
</script></body></html>