Ryan OShea | de36e4a | 2020-03-13 16:26:19 +0000 | [diff] [blame] | 1 | <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"> |
| 2 | <html xmlns="http://www.w3.org/1999/xhtml"> |
| 3 | <head> |
| 4 | <meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/> |
| 5 | <meta http-equiv="X-UA-Compatible" content="IE=9"/> |
| 6 | <meta name="generator" content="Doxygen 1.8.13"/> |
| 7 | <meta name="viewport" content="width=device-width, initial-scale=1"/> |
| 8 | <title>ArmNN: TensorFlow operators that the Arm NN SDK supports</title> |
| 9 | <link href="tabs.css" rel="stylesheet" type="text/css"/> |
| 10 | <script type="text/javascript" src="jquery.js"></script> |
| 11 | <script type="text/javascript" src="dynsections.js"></script> |
| 12 | <link href="navtree.css" rel="stylesheet" type="text/css"/> |
| 13 | <script type="text/javascript" src="resize.js"></script> |
| 14 | <script type="text/javascript" src="navtreedata.js"></script> |
| 15 | <script type="text/javascript" src="navtree.js"></script> |
| 16 | <script type="text/javascript"> |
| 17 | $(document).ready(initResizable); |
| 18 | </script> |
| 19 | <link href="search/search.css" rel="stylesheet" type="text/css"/> |
| 20 | <script type="text/javascript" src="search/searchdata.js"></script> |
| 21 | <script type="text/javascript" src="search/search.js"></script> |
| 22 | <link href="doxygen.css" rel="stylesheet" type="text/css" /> |
| 23 | </head> |
| 24 | <body> |
| 25 | <div id="top"><!-- do not remove this div, it is closed by doxygen! --> |
| 26 | <div id="titlearea"> |
| 27 | <table cellspacing="0" cellpadding="0"> |
| 28 | <tbody> |
| 29 | <tr style="height: 56px;"> |
| 30 | <td id="projectalign" style="padding-left: 0.5em;"> |
| 31 | <div id="projectname">ArmNN |
| 32 |  <span id="projectnumber">NotReleased</span> |
| 33 | </div> |
| 34 | </td> |
| 35 | </tr> |
| 36 | </tbody> |
| 37 | </table> |
| 38 | </div> |
| 39 | <!-- end header part --> |
| 40 | <!-- Generated by Doxygen 1.8.13 --> |
| 41 | <script type="text/javascript"> |
| 42 | var searchBox = new SearchBox("searchBox", "search",false,'Search'); |
| 43 | </script> |
| 44 | <script type="text/javascript" src="menudata.js"></script> |
| 45 | <script type="text/javascript" src="menu.js"></script> |
| 46 | <script type="text/javascript"> |
| 47 | $(function() { |
| 48 | initMenu('',true,false,'search.php','Search'); |
| 49 | $(document).ready(function() { init_search(); }); |
| 50 | }); |
| 51 | </script> |
| 52 | <div id="main-nav"></div> |
| 53 | </div><!-- top --> |
| 54 | <div id="side-nav" class="ui-resizable side-nav-resizable"> |
| 55 | <div id="nav-tree"> |
| 56 | <div id="nav-tree-contents"> |
| 57 | <div id="nav-sync" class="sync"></div> |
| 58 | </div> |
| 59 | </div> |
| 60 | <div id="splitbar" style="-moz-user-select:none;" |
| 61 | class="ui-resizable-handle"> |
| 62 | </div> |
| 63 | </div> |
| 64 | <script type="text/javascript"> |
| 65 | $(document).ready(function(){initNavTree('md_src_armnn_tf_parser__tensor_flow_support.html','');}); |
| 66 | </script> |
| 67 | <div id="doc-content"> |
| 68 | <!-- window showing the filter options --> |
| 69 | <div id="MSearchSelectWindow" |
| 70 | onmouseover="return searchBox.OnSearchSelectShow()" |
| 71 | onmouseout="return searchBox.OnSearchSelectHide()" |
| 72 | onkeydown="return searchBox.OnSearchSelectKey(event)"> |
| 73 | </div> |
| 74 | |
| 75 | <!-- iframe showing the search results (closed by default) --> |
| 76 | <div id="MSearchResultsWindow"> |
| 77 | <iframe src="javascript:void(0)" frameborder="0" |
| 78 | name="MSearchResults" id="MSearchResults"> |
| 79 | </iframe> |
| 80 | </div> |
| 81 | |
| 82 | <div class="header"> |
| 83 | <div class="headertitle"> |
| 84 | <div class="title">TensorFlow operators that the Arm NN SDK supports </div> </div> |
| 85 | </div><!--header--> |
| 86 | <div class="contents"> |
| 87 | <div class="textblock"><p>This reference guide provides a list of TensorFlow operators the Arm NN SDK currently supports.</p> |
| 88 | <p>The Arm NN SDK TensorFlow parser currently only supports fp32 operators.</p> |
| 89 | <h2>Fully supported</h2> |
| 90 | <p><b>avg_pool</b></p> |
| 91 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/avg_pool">avg_pool documentation</a> for more information.</p> |
| 92 | <p><b>bias_add</b></p> |
| 93 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/bias_add">bias_add documentation</a> for more information.</p> |
| 94 | <p><b>conv2d</b></p> |
| 95 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/conv2d">conv2d documentation</a> for more information.</p> |
| 96 | <p><b>expand_dims</b></p> |
| 97 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/expand_dims">expand_dims documentation</a> for more information.</p> |
| 98 | <p><b>gather</b></p> |
| 99 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/gather">gather documentation</a> for more information.</p> |
| 100 | <p><b>identity</b></p> |
| 101 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/identity">identity documentation</a> for more information.</p> |
| 102 | <p><b>local_response_normalization</b></p> |
| 103 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/local_response_normalization">local_response_normalization documentation</a> for more information.</p> |
| 104 | <p><b>max_pool</b></p> |
| 105 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/max_pool">max_pool documentation</a> for more information.</p> |
| 106 | <p><b>placeholder</b></p> |
| 107 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/placeholder">placeholder documentation</a> for more information.</p> |
| 108 | <p><b>reduce_mean</b></p> |
| 109 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/reduce_mean">reduce_mean documentation</a> for more information.</p> |
| 110 | <p><b>relu</b></p> |
| 111 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/relu">relu documentation</a> for more information.</p> |
| 112 | <p><b>relu6</b></p> |
| 113 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/relu6">relu6 documentation</a> for more information.</p> |
| 114 | <p><b>rsqrt</b></p> |
| 115 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/math/rsqrt">rsqrt documentation</a> for more information.</p> |
| 116 | <p><b>shape</b></p> |
| 117 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/shape">shape documentation</a> for more information.</p> |
| 118 | <p><b>sigmoid</b></p> |
| 119 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/sigmoid">sigmoid documentation</a> for more information.</p> |
| 120 | <p><b>softplus</b></p> |
| 121 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/softplus">softplus documentation</a> for more information.</p> |
| 122 | <p><b>squeeze</b></p> |
| 123 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/squeeze">squeeze documentation</a> for more information.</p> |
| 124 | <p><b>tanh</b></p> |
| 125 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/tanh">tanh documentation</a> for more information.</p> |
| 126 | <h2>Partially supported</h2> |
| 127 | <p><b>add</b></p> |
| 128 | <p>The parser does not support all forms of <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>, only broadcasting of scalars and 1D tensors. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/add">add operator documentation</a> for more information.</p> |
| 129 | <p><b>add_n</b></p> |
| 130 | <p>The parser does not support all forms of <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>, only broadcasting of scalars and 1D tensors. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/add_n">add operator documentation</a> for more information.</p> |
| 131 | <p><b>concat</b></p> |
| 132 | <p>Arm NN supports concatenation along the channel dimension for data formats NHWC and NCHW.</p> |
| 133 | <p><b>constant</b></p> |
| 134 | <p>The parser does not support the optional <code>shape</code> argument. It always infers the shape of the output tensor from <code>value</code>. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/constant">constant documentation</a> for further information.</p> |
| 135 | <p><b>depthwise_conv2d_native</b></p> |
| 136 | <p>The parser only supports a dilation rate of (1,1,1,1). See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d_native">depthwise_conv2d_native documentation</a> for more information.</p> |
| 137 | <p><b>equal</b></p> |
| 138 | <p>The parser does not support all forms of <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>, only broadcasting of 4D and 1D tensors. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/math/equal">equal operator documentation</a> for more information.</p> |
| 139 | <p><b>fused_batch_norm</b></p> |
| 140 | <p>The parser does not support training outputs. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/fused_batch_norm">fused_batch_norm documentation</a> for more information.</p> |
| 141 | <p><b>greater</b></p> |
| 142 | <p>The parser does not support all forms of <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>, only broadcasting of 4D and 1D tensors. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/math/greater">greater operator documentation</a> for more information.</p> |
| 143 | <p><b>matmul</b></p> |
| 144 | <p>The parser only supports constant weights in a fully connected layer. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/matmul">matmul documentation</a> for more information.</p> |
| 145 | <p><b>maximum</b></p> |
| 146 | <p>where maximum is used in one of the following ways</p> |
| 147 | <ul> |
| 148 | <li>max(mul(a, x), x)</li> |
| 149 | <li>max(mul(x, a), x)</li> |
| 150 | <li>max(x, mul(a, x))</li> |
| 151 | <li>max(x, mul(x, a)</li> |
| 152 | </ul> |
| 153 | <p>This is interpreted as a ActivationLayer with a LeakyRelu activation function. Any other usage of max will result in the insertion of a simple maximum layer. The parser does not support all forms of <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/maximum">maximum documentation</a> for more information.</p> |
| 154 | <p><b>minimum</b></p> |
| 155 | <p>The parser does not support all forms of <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>, only broadcasting of 4D and 1D tensors. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/math/minimum">minimum operator documentation</a> for more information.</p> |
| 156 | <p><b>multiply</b></p> |
| 157 | <p>The parser does not support all forms of <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>, only broadcasting of scalars and 1D tensors. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/multiply">multiply documentation</a> for more information.</p> |
| 158 | <p><b>pad</b></p> |
| 159 | <p>Only supports tf.pad function with mode = 'CONSTANT' and constant_values = 0. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/pad">pad documentation</a> for more information.</p> |
| 160 | <p><b>realdiv</b></p> |
| 161 | <p>The parser does not support all forms of <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>, only broadcasting of scalars and 1D tensors. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/realdiv">realdiv documentation</a> for more information.</p> |
| 162 | <p><b>reshape</b></p> |
| 163 | <p>The parser does not support reshaping to or from 4D. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/reshape">reshape documentation</a> for more information.</p> |
| 164 | <p><b>resize_images</b></p> |
| 165 | <p>The parser only supports <code>ResizeMethod.BILINEAR</code> with <code>align_corners=False</code>. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/image/resize_images">resize_images documentation</a> for more information.</p> |
| 166 | <p><b>softmax</b></p> |
| 167 | <p>The parser only supports 2D inputs and does not support selecting the <code>softmax</code> dimension. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/nn/softmax">softmax documentation</a> for more information.</p> |
| 168 | <p><b>split</b></p> |
| 169 | <p>Arm NN supports split along the channel dimension for data formats NHWC and NCHW.</p> |
| 170 | <p><b>subtract</b></p> |
| 171 | <p>The parser does not support all forms of broadcasting <a href="https://www.tensorflow.org/performance/xla/broadcasting">broadcast composition</a>, only broadcasting of scalars and 1D tensors. See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/math/subtract">subtract documentation</a> for more information.</p> |
| 172 | <p><b>pack/stack</b></p> |
| 173 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/stack">stack documentation</a> for more information.</p> |
| 174 | <p><b>strided_slice</b></p> |
| 175 | <p>See the TensorFlow <a href="https://www.tensorflow.org/api_docs/python/tf/strided_slice">strided_slice documentation</a> for more information.</p> |
| 176 | <h2>Tested networks</h2> |
| 177 | <p>Arm tests these operators with the following TensorFlow fp32 neural networks:</p> |
| 178 | <ul> |
| 179 | <li>Cifar10</li> |
| 180 | <li>Lenet</li> |
| 181 | <li>Simple MNIST. For more information check out the <a href="https://developer.arm.com/technologies/machine-learning-on-arm/developer-material/how-to-guides/deploying-a-tensorflow-mnist-model-on-arm-nn">tutorial</a> on the Arm Developer portal.</li> |
| 182 | <li>mobilenet_v1_1.0_224. The Arm NN SDK only supports the non-quantized version of the network. See the https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md "MobileNet_v1 documentation" for more information on quantized networks.</li> |
| 183 | <li>inception_v3. The Arm NN SDK only supports the official inception_v3 transformed model. See the TensorFlow documentation on <a href="https://www.tensorflow.org/mobile/prepare_models">preparing models for mobile deployment</a> for more information on how to transform the inception_v3 network.</li> |
| 184 | </ul> |
| 185 | <p>More machine learning operators will be supported in future releases. </p> |
| 186 | </div></div><!-- contents --> |
| 187 | </div><!-- doc-content --> |
| 188 | <!-- start footer part --> |
| 189 | <div id="nav-path" class="navpath"><!-- id is needed for treeview function! --> |
| 190 | <ul> |
| 191 | <li class="footer">Generated on Fri Mar 13 2020 16:07:01 for ArmNN by |
| 192 | <a href="http://www.doxygen.org/index.html"> |
| 193 | <img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.13 </li> |
| 194 | </ul> |
| 195 | </div> |
| 196 | </body> |
| 197 | </html> |