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
|
<!-- Copyright (c) 2020 ARM Limited. -->
<!-- -->
<!-- SPDX-License-Identifier: MIT -->
<!-- -->
<!-- HTML header for doxygen 1.8.13-->
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
<meta http-equiv="X-UA-Compatible" content="IE=9"/>
<meta name="generator" content="Doxygen 1.8.13"/>
<meta name="robots" content="NOINDEX, NOFOLLOW" />
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<title>ArmNN: delegate/IntegrateDelegateIntoPython.md Source File</title>
<link href="tabs.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="jquery.js"></script>
<script type="text/javascript" src="dynsections.js"></script>
<link href="navtree.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="resize.js"></script>
<script type="text/javascript" src="navtreedata.js"></script>
<script type="text/javascript" src="navtree.js"></script>
<script type="text/javascript">
$(document).ready(initResizable);
</script>
<link href="search/search.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="search/searchdata.js"></script>
<script type="text/javascript" src="search/search.js"></script>
<script type="text/x-mathjax-config">
MathJax.Hub.Config({
extensions: ["tex2jax.js"],
jax: ["input/TeX","output/HTML-CSS"],
});
</script><script type="text/javascript" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js"></script>
<link href="doxygen.css" rel="stylesheet" type="text/css" />
<link href="stylesheet.css" rel="stylesheet" type="text/css"/>
</head>
<body>
<div id="top"><!-- do not remove this div, it is closed by doxygen! -->
<div id="titlearea">
<table cellspacing="0" cellpadding="0">
<tbody>
<tr style="height: 56px;">
<img alt="ArmNN" src="Arm_NN_horizontal_blue.png" style="max-width: 10rem; margin-top: .5rem; margin-left 10px"/>
<td style="padding-left: 0.5em;">
<div id="projectname">
 <span id="projectnumber">21.02</span>
</div>
</td>
</tr>
</tbody>
</table>
</div>
<!-- end header part -->
<!-- Generated by Doxygen 1.8.13 -->
<script type="text/javascript">
var searchBox = new SearchBox("searchBox", "search",false,'Search');
</script>
<script type="text/javascript" src="menudata.js"></script>
<script type="text/javascript" src="menu.js"></script>
<script type="text/javascript">
$(function() {
initMenu('',true,false,'search.php','Search');
$(document).ready(function() { init_search(); });
});
</script>
<div id="main-nav"></div>
</div><!-- top -->
<div id="side-nav" class="ui-resizable side-nav-resizable">
<div id="nav-tree">
<div id="nav-tree-contents">
<div id="nav-sync" class="sync"></div>
</div>
</div>
<div id="splitbar" style="-moz-user-select:none;"
class="ui-resizable-handle">
</div>
</div>
<script type="text/javascript">
$(document).ready(function(){initNavTree('_integrate_delegate_into_python_8md.xhtml','');});
</script>
<div id="doc-content">
<!-- window showing the filter options -->
<div id="MSearchSelectWindow"
onmouseover="return searchBox.OnSearchSelectShow()"
onmouseout="return searchBox.OnSearchSelectHide()"
onkeydown="return searchBox.OnSearchSelectKey(event)">
</div>
<!-- iframe showing the search results (closed by default) -->
<div id="MSearchResultsWindow">
<iframe src="javascript:void(0)" frameborder="0"
name="MSearchResults" id="MSearchResults">
</iframe>
</div>
<div class="header">
<div class="headertitle">
<div class="title">delegate/IntegrateDelegateIntoPython.md</div> </div>
</div><!--header-->
<div class="contents">
<a href="_integrate_delegate_into_python_8md.xhtml">Go to the documentation of this file.</a><div class="fragment"><div class="line"><a name="l00001"></a><span class="lineno"> 1</span> # Integrate the TfLite delegate into TfLite using Python</div><div class="line"><a name="l00002"></a><span class="lineno"> 2</span> If you have built the TfLite delegate as a separate dynamic library then this tutorial will show you how you can</div><div class="line"><a name="l00003"></a><span class="lineno"> 3</span> integrate it in TfLite to run models using python.</div><div class="line"><a name="l00004"></a><span class="lineno"> 4</span> </div><div class="line"><a name="l00005"></a><span class="lineno"> 5</span> Here is an example python script showing how to do this. In this script we are making use of the </div><div class="line"><a name="l00006"></a><span class="lineno"> 6</span> [external adaptor](https://www.tensorflow.org/lite/performance/implementing_delegate#option_2_leverage_external_delegate) </div><div class="line"><a name="l00007"></a><span class="lineno"> 7</span> tool of TfLite that allows you to load delegates at runtime.</div><div class="line"><a name="l00008"></a><span class="lineno"> 8</span> ```python</div><div class="line"><a name="l00009"></a><span class="lineno"> 9</span> import numpy as np</div><div class="line"><a name="l00010"></a><span class="lineno"> 10</span> import tflite_runtime.interpreter as tflite</div><div class="line"><a name="l00011"></a><span class="lineno"> 11</span> </div><div class="line"><a name="l00012"></a><span class="lineno"> 12</span> # Load TFLite model and allocate tensors.</div><div class="line"><a name="l00013"></a><span class="lineno"> 13</span> # (if you are using the complete tensorflow package you can find load_delegate in tf.experimental.load_delegate)</div><div class="line"><a name="l00014"></a><span class="lineno"> 14</span> armnn_delegate = tflite.load_delegate( library="<your-armnn-build-dir>/delegate/libarmnnDelegate.so",</div><div class="line"><a name="l00015"></a><span class="lineno"> 15</span>  options={"backends": "CpuAcc,GpuAcc,CpuRef", "logging-severity":"info"})</div><div class="line"><a name="l00016"></a><span class="lineno"> 16</span> # Delegates/Executes all operations supported by ArmNN to/with ArmNN</div><div class="line"><a name="l00017"></a><span class="lineno"> 17</span> interpreter = tflite.Interpreter(model_path="<your-armnn-repo-dir>/delegate/python/test/test_data/mock_model.tflite", </div><div class="line"><a name="l00018"></a><span class="lineno"> 18</span>  experimental_delegates=[armnn_delegate])</div><div class="line"><a name="l00019"></a><span class="lineno"> 19</span> interpreter.allocate_tensors()</div><div class="line"><a name="l00020"></a><span class="lineno"> 20</span> </div><div class="line"><a name="l00021"></a><span class="lineno"> 21</span> # Get input and output tensors.</div><div class="line"><a name="l00022"></a><span class="lineno"> 22</span> input_details = interpreter.get_input_details()</div><div class="line"><a name="l00023"></a><span class="lineno"> 23</span> output_details = interpreter.get_output_details()</div><div class="line"><a name="l00024"></a><span class="lineno"> 24</span> </div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span> # Test model on random input data.</div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span> input_shape = input_details[0]['shape']</div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> input_data = np.array(np.random.random_sample(input_shape), dtype=np.uint8)</div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> interpreter.set_tensor(input_details[0]['index'], input_data)</div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> </div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> interpreter.invoke()</div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> </div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span> # Print out result</div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> output_data = interpreter.get_tensor(output_details[0]['index'])</div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span> print(output_data)</div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span> ```</div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> </div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span> # Prepare the environment</div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span> Pre-requisites:</div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span>  * Dynamically build Arm NN Delegate library</div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span>  * python3 (Depends on TfLite version)</div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span>  * virtualenv</div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span>  * numpy (Depends on TfLite version)</div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span>  * tflite_runtime (>=2.0, depends on Arm NN Delegate)</div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span> </div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span> If you haven't built the delegate yet then take a look at the [build guide](./BuildGuideNative.md).</div><div class="line"><a name="l00046"></a><span class="lineno"> 46</span> </div><div class="line"><a name="l00047"></a><span class="lineno"> 47</span> We recommend creating a virtual environment for this tutorial. For the following code to work python3 is needed. Please</div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span> also check the documentation of the TfLite version you want to use. There might be additional prerequisites for the python</div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> version.</div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span> ```bash</div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span> # Install python3 (We ended up with python3.5.3) and virtualenv</div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span> sudo apt-get install python3-pip</div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span> sudo pip3 install virtualenv</div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span> </div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span> # create a virtual environment</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span> cd your/tutorial/dir</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span> # creates a directory myenv at the current location</div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span> virtualenv -p python3 myenv </div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span> # activate the environment</div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span> source myenv/bin/activate</div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span> ```</div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span> </div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span> Now that the environment is active we can install additional packages we need for our example script. As you can see </div><div class="line"><a name="l00064"></a><span class="lineno"> 64</span> in the python script at the start of this page, this tutorial uses the `tflite_runtime` rather than the whole tensorflow </div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span> package. The `tflite_runtime` is a package that wraps the TfLite Interpreter. Therefore it can only be used to run inferences of </div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span> TfLite models. But since Arm NN is only an inference engine itself this is a perfect match. The </div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span> `tflite_runtime` is also much smaller than the whole tensorflow package and better suited to run models on </div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span> mobile and embedded devices.</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span> </div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span> At the time of writing, there are no packages of either `tensorflow` or `tflite_runtime` available on `pypi` that </div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span> are built for an arm architecture. That means installing them using `pip` on your development board is currently not </div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span> possible. The TfLite [website](https://www.tensorflow.org/lite/guide/python) points you at prebuilt `tflite_runtime` </div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span> packages. However, that limits you to specific TfLite and Python versions. For this reason we will build the </div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span> `tflite_runtime` from source.</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span> </div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span> You will have downloaded the tensorflow repository in order to build the Arm NN delegate. In there you can find further </div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span> instructions on how to build the `tflite_runtime` under `tensorflow/lite/tools/pip_package/README.md`. This tutorial </div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span> uses bazel to build it natively but there are scripts for cross-compilation available as well.</div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span> ```bash</div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span> # Add the directory where bazel is built to your PATH so that the script can find it</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span> PATH=$PATH:your/build/dir/bazel/output</div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span> # Run the following script to build tflite_runtime natively.</div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span> tensorflow/lite/tools/pip_package/build_pip_package_with_bazel.sh</div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span> ```</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span> The execution of the script creates a `.whl` file which can be used by `pip` to install the TfLite Runtime package. </div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span> The build-script produces some output in which you can find the location where the `.whl` file was created. Then all that is </div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span> left to do is to install all necessary python packages with `pip`.</div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span> ```bash</div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span> pip install tensorflow/lite/tools/pip_package/gen/tflite_pip/python3/dist/tflite_runtime-2.3.1-py3-none-any.whl numpy</div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span> ```</div><div class="line"><a name="l00091"></a><span class="lineno"> 91</span> </div><div class="line"><a name="l00092"></a><span class="lineno"> 92</span> Your virtual environment is now all setup. Copy the final python script into a python file e.g. </div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span> `ExternalDelegatePythonTutorial.py`. Modify the python script above and replace `<your-armnn-build-dir>` and </div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span> `<your-armnn-repo-dir>` with the directories you have set up. If you've been using the [native build guide](./BuildGuideNative.md) </div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span> this will be `$BASEDIR/armnn/build` and `$BASEDIR/armnn`.</div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span> </div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span> Finally, execute the script:</div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span> ```bash</div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span> python ExternalDelegatePythonTutorial.py</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span> ```</div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span> The output should look similar to this:</div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span> ```bash</div><div class="line"><a name="l00103"></a><span class="lineno"> 103</span> Info: ArmNN v23.0.0</div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span> </div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span> Info: Initialization time: 0.56 ms</div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span> </div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span> INFO: TfLiteArmnnDelegate: Created TfLite ArmNN delegate.</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span> [[ 12 123 16 12 11 14 20 16 20 12]]</div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span> Info: Shutdown time: 0.28 ms</div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span> ```</div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span> </div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span> For more details on what kind of options you can pass to the Arm NN delegate please check </div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span> [armnn_delegate_adaptor.cpp](src/armnn_external_delegate.cpp).</div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span> </div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span> You can also test the functionality of the external delegate adaptor by running some unit tests:</div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span> ```bash</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span> pip install pytest</div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span> cd armnn/delegate/python/test</div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span> # You can deselect tests that require backends that your hardware doesn't support using markers e.g. -m "not GpuAccTest"</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span> pytest --delegate-dir="<your-armnn-build-dir>/armnn/delegate/libarmnnDelegate.so" -m "not GpuAccTest" </div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span> ```</div></div><!-- fragment --></div><!-- contents -->
</div><!-- doc-content -->
<!-- start footer part -->
<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
<ul>
<li class="navelem"><a class="el" href="_integrate_delegate_into_python_8md.xhtml">IntegrateDelegateIntoPython.md</a></li>
<li class="footer">Generated on Thu Feb 25 2021 17:27:28 for ArmNN by
<a href="http://www.doxygen.org/index.html">
<img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.13 </li>
</ul>
</div>
</body>
</html>
|
