-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathonnx_transfer.html
387 lines (265 loc) · 17.4 KB
/
onnx_transfer.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
<!DOCTYPE html>
<html class="writer-html5" lang="en" >
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Train from ONNX models — N2D2 documentation</title>
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="_static/pygments.css" type="text/css" />
<!--[if lt IE 9]>
<script src="_static/js/html5shiv.min.js"></script>
<![endif]-->
<script type="text/javascript" id="documentation_options" data-url_root="./" src="_static/documentation_options.js"></script>
<script src="_static/jquery.js"></script>
<script src="_static/underscore.js"></script>
<script src="_static/doctools.js"></script>
<script src="_static/language_data.js"></script>
<script async="async" src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.7/latest.js?config=TeX-AMS-MML_HTMLorMML"></script>
<script type="text/javascript" src="_static/js/theme.js"></script>
<link rel="author" title="About these documents" href="about.html" />
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="Post-training quantization" href="quant_post.html" />
<link rel="prev" title="Import ONNX models" href="onnx_import.html" />
</head>
<body class="wy-body-for-nav">
<div class="wy-grid-for-nav">
<nav data-toggle="wy-nav-shift" class="wy-nav-side">
<div class="wy-side-scroll">
<div class="wy-side-nav-search" >
<a href="index.html" class="icon icon-home" alt="Documentation Home"> N2D2
</a>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<p class="caption"><span class="caption-text">Introduction:</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="intro.html">Presentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="about.html">About N2D2-IP</a></li>
<li class="toctree-l1"><a class="reference internal" href="simus.html">Performing simulations</a></li>
<li class="toctree-l1"><a class="reference internal" href="perfs_tools.html">Performance evaluation tools</a></li>
<li class="toctree-l1"><a class="reference internal" href="tuto.html">Tutorials</a></li>
</ul>
<p class="caption"><span class="caption-text">ONNX Import:</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="onnx_convert.html">Obtain ONNX models</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx_import.html">Import ONNX models</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Train from ONNX models</a><ul>
<li class="toctree-l2"><a class="reference internal" href="#with-an-ini-file">With an INI file</a><ul>
<li class="toctree-l3"><a class="reference internal" href="#remove-the-original-classifier">1) Remove the original classifier</a></li>
<li class="toctree-l3"><a class="reference internal" href="#add-a-new-classifier-to-the-onnx-model">2) Add a new classifier to the ONNX model</a></li>
<li class="toctree-l3"><a class="reference internal" href="#fine-tuning-optional">3) Fine tuning (optional)</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="#with-the-python-api">With the Python API</a></li>
</ul>
</li>
</ul>
<p class="caption"><span class="caption-text">Quantization and Export:</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="quant_post.html">Post-training quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="quant_qat.html">[NEW] Quantization-Aware Training</a></li>
<li class="toctree-l1"><a class="reference internal" href="export_CPP.html">Export: C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="export_CPP_STM32.html">Export: C++/STM32</a></li>
<li class="toctree-l1"><a class="reference internal" href="export_TensorRT.html">Export: TensorRT</a></li>
<li class="toctree-l1"><a class="reference internal" href="export_DNeuro.html">Export: DNeuro</a></li>
<li class="toctree-l1"><a class="reference internal" href="export_ONNX.html">Export: ONNX</a></li>
<li class="toctree-l1"><a class="reference internal" href="export_legacy.html">Export: other / legacy</a></li>
</ul>
<p class="caption"><span class="caption-text">INI File Interface:</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="ini_intro.html">Introduction</a></li>
<li class="toctree-l1"><a class="reference internal" href="ini_databases.html">Databases</a></li>
<li class="toctree-l1"><a class="reference internal" href="ini_data_analysis.html">Stimuli data analysis</a></li>
<li class="toctree-l1"><a class="reference internal" href="ini_environment.html">Stimuli provider (Environment)</a></li>
<li class="toctree-l1"><a class="reference internal" href="ini_layers.html">Network Layers</a></li>
<li class="toctree-l1"><a class="reference internal" href="ini_target.html">Targets (outputs & losses)</a></li>
<li class="toctree-l1"><a class="reference internal" href="adversarial.html">Adversarial module</a></li>
</ul>
<p class="caption"><span class="caption-text">Python API:</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="containers.html">Containers</a></li>
<li class="toctree-l1"><a class="reference internal" href="cells.html">Cells</a></li>
<li class="toctree-l1"><a class="reference internal" href="databases.html">Databases</a></li>
<li class="toctree-l1"><a class="reference internal" href="stimuliprovider.html">StimuliProvider</a></li>
<li class="toctree-l1"><a class="reference internal" href="deepnet.html">DeepNet</a></li>
</ul>
<p class="caption"><span class="caption-text">C++ API / Developer:</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="dev_intro.html">Introduction</a></li>
</ul>
</div>
</div>
</nav>
<section data-toggle="wy-nav-shift" class="wy-nav-content-wrap">
<nav class="wy-nav-top" aria-label="top navigation">
<i data-toggle="wy-nav-top" class="fa fa-bars"></i>
<a href="index.html">N2D2</a>
</nav>
<div class="wy-nav-content">
<div class="rst-content">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="wy-breadcrumbs">
<li><a href="index.html" class="icon icon-home"></a> »</li>
<li>Train from ONNX models</li>
<li class="wy-breadcrumbs-aside">
<a href="_sources/onnx_transfer.rst.txt" rel="nofollow"> View page source</a>
</li>
</ul>
<hr/>
</div>
<div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
<div itemprop="articleBody">
<div class="section" id="train-from-onnx-models">
<h1>Train from ONNX models<a class="headerlink" href="#train-from-onnx-models" title="Permalink to this headline">¶</a></h1>
<p>The ONNX specification does not include any training parameter. To perform a
training on an imported ONNX model, it is possible to add the training elements
(solvers, learning rate scheduler…) on top of an ONNX model in N2D2, in the
INI file directly or using the Python API.</p>
<p>This is particularly useful to perform transfer learning from an existing ONNX
model trained on ImageNet for example.</p>
<div class="section" id="with-an-ini-file">
<h2>With an INI file<a class="headerlink" href="#with-an-ini-file" title="Permalink to this headline">¶</a></h2>
<p>We propose in this section to apply transfer learning to a MobileNet v1 ONNX
model. We assume that this model is obtained by converting the reference
pre-trained model from Google using the <code class="docutils literal notranslate"><span class="pre">tools/mobilenet_v1_to_onnx.sh</span></code> tool
provided in N2D2. The resulting model file name is therefore assumed to be
<code class="docutils literal notranslate"><span class="pre">mobilenet_v1_1.0_224.onnx</span></code>.</p>
<div class="section" id="remove-the-original-classifier">
<h3>1) Remove the original classifier<a class="headerlink" href="#remove-the-original-classifier" title="Permalink to this headline">¶</a></h3>
<p>The first step to perform transfer learning is to remove the existing classifier
from the ONNX model. To do so, one can simply use the <code class="docutils literal notranslate"><span class="pre">Ignore</span></code> parameter in
the ONNX INI section.</p>
<div class="highlight-ini notranslate"><div class="highlight"><pre><span></span><span class="k">[onnx]</span>
<span class="na">Input</span><span class="o">=</span><span class="s">sp</span>
<span class="na">Type</span><span class="o">=</span><span class="s">ONNX</span>
<span class="na">File</span><span class="o">=</span><span class="s">mobilenet_v1_1.0_224.onnx</span>
<span class="c1">; Remove the last layer and the softmax for transfer learning</span>
<span class="na">Ignore</span><span class="o">=</span><span class="s">Conv__252:0 MobilenetV1/Predictions/Softmax:0</span>
</pre></div>
</div>
</div>
<div class="section" id="add-a-new-classifier-to-the-onnx-model">
<h3>2) Add a new classifier to the ONNX model<a class="headerlink" href="#add-a-new-classifier-to-the-onnx-model" title="Permalink to this headline">¶</a></h3>
<p>The next step is to add a new classifier (fully connected layer with a softmax)
and connect it to the last layer in the ONNX model.</p>
<p>In order to properly handle graph dependencies, all the N2D2 layers connected
to a layer embedded in an ONNX model, must take the ONNX section name
(here <code class="docutils literal notranslate"><span class="pre">onnx</span></code>) as first input in the <code class="docutils literal notranslate"><span class="pre">Input</span></code> parameter. The actual inputs
are then added in the comma-separated list, which can mix ONNX and N2D2 layers.
In the example below, the average pooling layer from the ONNX model is connected
to the <code class="docutils literal notranslate"><span class="pre">Fc</span></code> cell:</p>
<div class="highlight-ini notranslate"><div class="highlight"><pre><span></span><span class="c1">; Here, we add our new layers for transfer learning</span>
<span class="k">[fc]</span>
<span class="c1">; first input MUST BE "onnx"</span>
<span class="c1">; for proper dependency handling</span>
<span class="na">Input</span><span class="o">=</span><span class="s">onnx,MobilenetV1/Logits/AvgPool_1a/AvgPool:0</span>
<span class="na">Type</span><span class="o">=</span><span class="s">Fc</span>
<span class="na">NbOutputs</span><span class="o">=</span><span class="s">100</span>
<span class="na">ActivationFunction</span><span class="o">=</span><span class="s">Linear</span>
<span class="na">WeightsFiller</span><span class="o">=</span><span class="s">XavierFiller</span>
<span class="na">ConfigSection</span><span class="o">=</span><span class="s">common.config</span>
<span class="k">[softmax]</span>
<span class="na">Input</span><span class="o">=</span><span class="s">fc</span>
<span class="na">Type</span><span class="o">=</span><span class="s">Softmax</span>
<span class="na">NbOutputs</span><span class="o">=</span><span class="s">[fc]NbOutputs</span>
<span class="na">WithLoss</span><span class="o">=</span><span class="s">1</span>
<span class="k">[softmax.Target]</span>
<span class="c1">; Common config for static model</span>
<span class="k">[common.config]</span>
<span class="na">WeightsSolver.LearningRate</span><span class="o">=</span><span class="s">0.01</span>
<span class="na">WeightsSolver.Momentum</span><span class="o">=</span><span class="s">0.9</span>
<span class="na">WeightsSolver.Decay</span><span class="o">=</span><span class="s">0.0005</span>
<span class="na">Solvers.LearningRatePolicy</span><span class="o">=</span><span class="s">StepDecay</span>
<span class="na">Solvers.LearningRateStepSize</span><span class="o">=</span><span class="s">[sp]_EpochSize</span>
<span class="na">Solvers.LearningRateDecay</span><span class="o">=</span><span class="s">0.993</span>
</pre></div>
</div>
<p>As this new classifier must be trained, all the training parameter must be
specified as usual for this layer.</p>
</div>
<div class="section" id="fine-tuning-optional">
<h3>3) Fine tuning (optional)<a class="headerlink" href="#fine-tuning-optional" title="Permalink to this headline">¶</a></h3>
<p>If one wants to also fine-tune the existing ONNX layers, one must set the
solver configuration for the ONNX layers, using default configuration sections.</p>
<p>Default configuration sections applies to all the layers of the same type in the
ONNX model. For example, to add default parameters to all convolution layers
in the ONNX model loaded in a section of type ONNX named <code class="docutils literal notranslate"><span class="pre">onnx</span></code>,
just add a section named <code class="docutils literal notranslate"><span class="pre">[onnx:Conv_def]</span></code> in the INI file. The name of the
default section follows the convention <code class="docutils literal notranslate"><span class="pre">[ONNXSection:N2D2CellType_def]</span></code>.</p>
<div class="highlight-ini notranslate"><div class="highlight"><pre><span></span><span class="c1">; Default section for ONNX Conv from section "onnx"</span>
<span class="c1">; "ConfigSection", solvers and fillers can be specified here...</span>
<span class="k">[onnx:Conv_def]</span>
<span class="na">ConfigSection</span><span class="o">=</span><span class="s">common.config</span>
<span class="c1">; Default section for ONNX Fc from section "onnx"</span>
<span class="k">[onnx:Fc_def]</span>
<span class="na">ConfigSection</span><span class="o">=</span><span class="s">common.config</span>
<span class="c1">; For BatchNorm, make sure the stats won't change if there is no fine-tuning</span>
<span class="k">[onnx:BatchNorm_def]</span>
<span class="na">ConfigSection</span><span class="o">=</span><span class="s">bn_notrain.config</span>
<span class="k">[bn_notrain.config]</span>
<span class="na">MovingAverageMomentum</span><span class="o">=</span><span class="s">0.0</span>
</pre></div>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Important: make sure that the BatchNorm stats does not change if the
BatchNorm layer are not fine-tuned! This can be done by setting the
parameter <code class="docutils literal notranslate"><span class="pre">MovingAverageMomentum</span></code> to 0.0 for the layer than must not be
fine-tuned.</p>
</div>
<p>It is possible to add parameters for a specific ONNX layer by adding a section
with the ONNX layer named.</p>
<p>You can fine-tune the whole network or only some of its layers, usually the last
ones. To stop the fine-tuning at a specific layer, one can simply prevent the
gradient from back-propagating further. This can be achieved with the
<code class="docutils literal notranslate"><span class="pre">BackPropagate=0</span></code> configuration parameter.</p>
<div class="highlight-ini notranslate"><div class="highlight"><pre><span></span><span class="k">[Conv__250]</span>
<span class="na">ConfigSection</span><span class="o">=</span><span class="s">common.config,notrain.config</span>
<span class="k">[notrain.config]</span>
<span class="na">BackPropagate</span><span class="o">=</span><span class="s">0</span>
</pre></div>
</div>
<p>For the full configuration related to this example and more information, have a
look in <code class="docutils literal notranslate"><span class="pre">models/MobileNet_v1_ONNX_transfer.ini</span></code>.</p>
</div>
</div>
<div class="section" id="with-the-python-api">
<h2>With the Python API<a class="headerlink" href="#with-the-python-api" title="Permalink to this headline">¶</a></h2>
<p>Coming soon.</p>
</div>
</div>
</div>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="quant_post.html" class="btn btn-neutral float-right" title="Post-training quantization" accesskey="n" rel="next">Next <span class="fa fa-arrow-circle-right"></span></a>
<a href="onnx_import.html" class="btn btn-neutral float-left" title="Import ONNX models" accesskey="p" rel="prev"><span class="fa fa-arrow-circle-left"></span> Previous</a>
</div>
<hr/>
<div role="contentinfo">
<p>
© Copyright 2019, CEA LIST
</p>
</div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a
<a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a>
provided by <a href="https://readthedocs.org">Read the Docs</a>.
</footer>
</div>
</div>
</section>
</div>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
</body>
</html>