Skip to content

Instantly share code, notes, and snippets.

@iwatake2222

iwatake2222/makefulluseofedgetpu.ipynb

Created May 9, 2020 11:51
Show Gist options
  • Save iwatake2222/5bb26123ebf891957109787d4f4dc685 to your computer and use it in GitHub Desktop.
Save iwatake2222/5bb26123ebf891957109787d4f4dc685 to your computer and use it in GitHub Desktop.
Download ZIP
MakeFullUseOfEdgeTPU.ipynb
Display the source blob
Display the rendered blob
Raw
makefulluseofedgetpu.ipynb
{
"nbformat": 4,
"nbformat_minor": 0,
"metadata": {
"colab": {
"name": "MakeFullUseOfEdgeTPU.ipynb",
"provenance": [],
"collapsed_sections": [
"XkL8FXvCl9BV"
],
"authorship_tag": "ABX9TyPJNGhxARgtJuhqBZQAxPdH",
"include_colab_link": true
},
"kernelspec": {
"name": "python3",
"display_name": "Python 3"
}
},
"cells": [
{
"cell_type": "markdown",
"metadata": {
"id": "view-in-github",
"colab_type": "text"
},
"source": [
"<a href=\"https://colab.research.google.com/gist/iwatake2222/5bb26123ebf891957109787d4f4dc685/makefulluseofedgetpu.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "lAMuurEvr6Af",
"colab_type": "text"
},
"source": [
"# Create very high TOPS tflite model for Edge TPU"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "-2hPvu7ar1SJ",
"colab_type": "text"
},
"source": [
"# Setup"
]
},
{
"cell_type": "code",
"metadata": {
"id": "XEmYPbuAy1Wc",
"colab_type": "code",
"colab": {}
},
"source": [
"# Install EdgeTPU compiler\n",
"!curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -\n",
"!echo \"deb https://packages.cloud.google.com/apt coral-edgetpu-stable main\" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list\n",
"!sudo apt-get update\n",
"!sudo apt-get install edgetpu-compiler"
],
"execution_count": 0,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "rfHlM9M2r4Xs",
"colab_type": "text"
},
"source": [
"# Create Keras model"
]
},
{
"cell_type": "code",
"metadata": {
"id": "Acemn-Iqt0tR",
"colab_type": "code",
"outputId": "77bd34bb-1df1-436f-8756-f311cb75bbbb",
"colab": {
"base_uri": "https://localhost:8080/",
"height": 1000
}
},
"source": [
"%tensorflow_version 1.x\n",
"from __future__ import absolute_import, division, print_function, unicode_literals\n",
"import tensorflow as tf\n",
"import numpy as np\n",
"\n",
"print(tf.__version__)\n",
"\n",
"# Model parameters\n",
"INTERNAL_MODEL_WIDTH = 128\n",
"INTERNAL_MODEL_HEIGHT = 128\n",
"FILTER_CHANNEL = 64\n",
"FILTER_KERNEL_SIZE = 3\n",
"FILTER_LAYERS_NUM = 100\n",
"\n",
"# Create model\n",
"input0 = tf.keras.layers.Input(shape=(2,2,3))\n",
"pad0 = tf.keras.layers.ZeroPadding2D(padding=((int)(INTERNAL_MODEL_WIDTH/2)-1, (int)(INTERNAL_MODEL_HEIGHT/2)-1))(input0)\n",
"conv = pad0\n",
"for i in range(FILTER_LAYERS_NUM):\n",
" conv = tf.keras.layers.Conv2D(\n",
" filters=FILTER_CHANNEL,\n",
" kernel_size=(FILTER_KERNEL_SIZE,FILTER_KERNEL_SIZE),\n",
" strides=(1,1),\n",
" padding=\"same\",\n",
" activation=tf.keras.activations.relu\n",
" )(conv)\n",
"convOut = tf.keras.layers.Conv2D(\n",
" filters=1,\n",
" kernel_size=(1,1),\n",
" strides=((int)(INTERNAL_MODEL_WIDTH/2), (int)(INTERNAL_MODEL_HEIGHT/2)),\n",
" padding=\"valid\",\n",
" activation=tf.keras.activations.relu\n",
")(conv)\n",
"model = tf.keras.models.Model(inputs=[input0], outputs=[convOut])\n",
"\n",
"# Save model\n",
"model.summary()\n",
"model.compile(\n",
" optimizer=tf.keras.optimizers.Adam(),\n",
" loss=tf.keras.losses.SparseCategoricalCrossentropy(),\n",
" metrics=[tf.keras.metrics.SparseCategoricalAccuracy()]\n",
")\n",
"model_name = \"model_conv_\" + str(INTERNAL_MODEL_WIDTH) + \"x\" + str(INTERNAL_MODEL_HEIGHT) + \"x\" + str(FILTER_CHANNEL) + \"x\" + str(FILTER_KERNEL_SIZE) + \"x\" + str(FILTER_LAYERS_NUM)\n",
"model.save(model_name + \".h5\")\n",
"\n",
"print(\"FLOPs = \" + (str)(INTERNAL_MODEL_WIDTH * INTERNAL_MODEL_HEIGHT * FILTER_KERNEL_SIZE * FILTER_KERNEL_SIZE * FILTER_CHANNEL * FILTER_CHANNEL * FILTER_LAYERS_NUM * 2))"
],
"execution_count": 91,
"outputs": [
{
"output_type": "stream",
"text": [
"1.15.2\n",
"Model: \"model\"\n",
"_________________________________________________________________\n",
"Layer (type) Output Shape Param # \n",
"=================================================================\n",
"input_1 (InputLayer) [(None, 2, 2, 3)] 0 \n",
"_________________________________________________________________\n",
"zero_padding2d (ZeroPadding2 (None, 128, 128, 3) 0 \n",
"_________________________________________________________________\n",
"conv2d (Conv2D) (None, 128, 128, 64) 1792 \n",
"_________________________________________________________________\n",
"conv2d_1 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_2 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_3 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_4 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_5 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_6 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_7 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_8 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_9 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_10 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_11 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_12 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_13 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_14 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_15 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_16 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_17 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_18 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_19 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_20 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_21 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_22 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_23 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_24 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_25 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_26 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_27 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_28 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_29 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_30 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_31 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_32 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_33 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_34 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_35 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_36 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_37 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_38 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_39 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_40 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_41 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_42 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_43 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_44 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_45 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_46 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_47 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_48 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_49 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_50 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_51 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_52 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_53 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_54 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_55 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_56 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_57 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_58 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_59 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_60 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_61 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_62 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_63 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_64 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_65 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_66 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_67 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_68 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_69 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_70 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_71 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_72 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_73 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_74 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_75 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_76 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_77 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_78 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_79 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_80 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_81 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_82 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_83 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_84 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_85 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_86 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_87 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_88 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_89 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_90 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_91 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_92 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_93 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_94 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_95 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_96 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_97 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_98 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_99 (Conv2D) (None, 128, 128, 64) 36928 \n",
"_________________________________________________________________\n",
"conv2d_100 (Conv2D) (None, 2, 2, 1) 65 \n",
"=================================================================\n",
"Total params: 3,657,729\n",
"Trainable params: 3,657,729\n",
"Non-trainable params: 0\n",
"_________________________________________________________________\n",
"FLOPs = 120795955200\n"
],
"name": "stdout"
}
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "SjOJGnCksQFR",
"colab_type": "text"
},
"source": [
"# Convert to tflite model (full integer quantization)"
]
},
{
"cell_type": "code",
"metadata": {
"id": "uaCM-iozuaYI",
"colab_type": "code",
"outputId": "f70e633f-1c82-45f3-cb4d-c2162219dc53",
"colab": {
"base_uri": "https://localhost:8080/",
"height": 69
}
},
"source": [
"%tensorflow_version 1.x\n",
"from __future__ import absolute_import, division, print_function, unicode_literals\n",
"import tensorflow as tf\n",
"import numpy as np\n",
"\n",
"## Prepara dataset generator for calibration\n",
"dummy_data_num = 2\n",
"dummy_data = []\n",
"for i in range(dummy_data_num):\n",
" dummy_data.append(np.random.rand(1, model.input_shape[1], model.input_shape[2], model.input_shape[3]))\n",
" dummy_data[i] = dummy_data[i].astype(np.float32)\n",
"\n",
"def representative_dataset_gen():\n",
" for i in range(dummy_data_num):\n",
" yield [dummy_data[i]]\n",
"\n",
"# Convert\n",
"# loaded_model = tf.keras.models.load_model(model_name + \".h5\")\n",
"# converter = tf.lite.TFLiteConverter.from_keras_model(loaded_model)\n",
"converter = tf.lite.TFLiteConverter.from_keras_model_file(model_name + \".h5\")\n",
"\n",
"converter.optimizations = [tf.lite.Optimize.DEFAULT]\n",
"converter.representative_dataset = representative_dataset_gen\n",
"\n",
"# For full integer quantization\n",
"converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]\n",
"converter.inference_input_type = tf.uint8\n",
"converter.inference_output_type = tf.uint8\n",
"converter.experimental_new_converter = True # will be no need in the future\n",
"\n",
"tflite_model = converter.convert()\n",
"open(model_name + \".tflite\", \"wb\").write(tflite_model)"
],
"execution_count": 92,
"outputs": [
{
"output_type": "stream",
"text": [
"INFO:tensorflow:Froze 202 variables.\n",
"INFO:tensorflow:Converted 202 variables to const ops.\n"
],
"name": "stdout"
},
{
"output_type": "execute_result",
"data": {
"text/plain": [
"3761240"
]
},
"metadata": {
"tags": []
},
"execution_count": 92
}
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "SjQHI2x7sUjz",
"colab_type": "text"
},
"source": [
"# Convert to edgetpu model"
]
},
{
"cell_type": "code",
"metadata": {
"id": "3meAgdciy8tE",
"colab_type": "code",
"outputId": "87b3067e-7723-4f6b-e374-41febc4560dd",
"colab": {
"base_uri": "https://localhost:8080/",
"height": 278
}
},
"source": [
"# Convert into edgetpu model\n",
"tflite_model_name = model_name + \".tflite\"\n",
"!edgetpu_compiler $tflite_model_name\n",
"\n",
"# Download to local\n",
"from google.colab import files\n",
"# files.download(model_name + \".tflite\")\n",
"files.download(model_name + \"_edgetpu.tflite\")"
],
"execution_count": 93,
"outputs": [
{
"output_type": "stream",
"text": [
"Edge TPU Compiler version 2.1.302470888\n",
"\n",
"Model compiled successfully in 667 ms.\n",
"\n",
"Input model: model_conv_128x128x64x3x100.tflite\n",
"Input size: 3.59MiB\n",
"Output model: model_conv_128x128x64x3x100_edgetpu.tflite\n",
"Output size: 3.77MiB\n",
"On-chip memory used for caching model parameters: 3.56MiB\n",
"On-chip memory remaining for caching model parameters: 3.35MiB\n",
"Off-chip memory used for streaming uncached model parameters: 0.00B\n",
"Number of Edge TPU subgraphs: 1\n",
"Total number of operations: 104\n",
"Operation log: model_conv_128x128x64x3x100_edgetpu.log\n",
"See the operation log file for individual operation details.\n"
],
"name": "stdout"
}
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "n6gugXkKl4wZ",
"colab_type": "text"
},
"source": [
"# Calculate FLOPS"
]
},
{
"cell_type": "code",
"metadata": {
"id": "Or4v0viVgxmn",
"colab_type": "code",
"colab": {}
},
"source": [
"!pip install flopco-keras "
],
"execution_count": 0,
"outputs": []
},
{
"cell_type": "code",
"metadata": {
"id": "95CYxLlwg0_9",
"colab_type": "code",
"outputId": "7e339226-1769-4c53-fde4-fa09c52783e8",
"colab": {
"base_uri": "https://localhost:8080/",
"height": 89
}
},
"source": [
"from flopco_keras import FlopCoKeras\n",
"stats = FlopCoKeras(model)\n",
"print(f\"FLOPs: {stats.total_flops / 1000 / 1000 / 1000 } [GFLOPs]\")\n",
"print(f\"MACs: {stats.total_macs}\")\n",
"print(f\"Relative FLOPs: {stats.relative_flops}\")"
],
"execution_count": 94,
"outputs": [
{
"output_type": "stream",
"text": [
"FLOPs: 119.749476868 [GFLOPs]\n",
"MACs: 59822309632\n",
"Relative FLOPs: [0.0, 0.0, 0.0004816027719567541, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 0.01009614538302068, 4.308995859487448e-09]\n"
],
"name": "stdout"
}
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "aWEKQ27wafKT",
"colab_type": "text"
},
"source": [
"# Code for inference (running on local PC with Edge TPU)\n",
"```\n",
"sudo pip3 install https://dl.google.com/coral/python/tflite_runtime-2.1.0.post1-cp37-cp37m-linux_armv7l.whl\n",
"echo \"deb https://packages.cloud.google.com/apt coral-edgetpu-stable main\" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list\n",
"curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -\n",
"sudo apt-get update\n",
"sudo apt-get install libedgetpu1-max\n",
"\n",
"python3 run.py\n",
"```\n",
"\n",
"Result\n",
"\n",
"```\n",
"pi@raspberrypi:~/top $ sudo python3 run.py\n",
"Inference time = 0.031113 [sec]\n",
"3848.906 [GFLOPS]\n",
"```\n"
]
},
{
"cell_type": "code",
"metadata": {
"id": "VWwDk9Hyaxky",
"colab_type": "code",
"colab": {}
},
"source": [
"import platform\n",
"import time\n",
"import tflite_runtime.interpreter as tflite\n",
"\n",
"EDGETPU_SHARED_LIB = {\n",
"\t'Linux': 'libedgetpu.so.1',\n",
"\t'Darwin': 'libedgetpu.1.dylib',\n",
"\t'Windows': 'edgetpu.dll'\n",
"}[platform.system()]\n",
"\n",
"def make_interpreter(model_file):\n",
"\tmodel_file, *device = model_file.split('@')\n",
"\treturn tflite.Interpreter(\n",
"\t\tmodel_path=model_file,\n",
"\t\texperimental_delegates=[\n",
"\t\t\ttflite.load_delegate(EDGETPU_SHARED_LIB,\n",
"\t\t\t\t\t\t\t\t{'device': device[0]} if device else {})\n",
"\t\t])\n",
"\n",
"def main():\n",
"\tinterpreter = make_interpreter(\"model_conv_128x128x64x3x100_edgetpu.tflite\")\n",
"\tinterpreter.allocate_tensors()\n",
"\tinterpreter.invoke()\n",
"\n",
"\tnum_measurement = 100\n",
"\tstart = time.perf_counter()\n",
"\tfor _ in range(num_measurement):\n",
"\t\tinterpreter.invoke()\n",
"\tinference_time = (time.perf_counter() - start) / num_measurement\n",
"\tprint(\"Inference time = %.6f [sec]\" % inference_time)\n",
"\tprint(\"%.3f [GFLOPS]\" % (119.749476868 / inference_time))\n",
"\n",
"if __name__ == \"__main__\":\n",
"\tmain()\n"
],
"execution_count": 0,
"outputs": []
},
{
"cell_type": "markdown",
"metadata": {
"id": "MSLD14DBeune",
"colab_type": "text"
},
"source": [
"================"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "XkL8FXvCl9BV",
"colab_type": "text"
},
"source": [
"# Calculate FLOPS (not working)\n",
"https://github.com/tensorflow/tensorflow/issues/32809"
]
},
{
"cell_type": "code",
"metadata": {
"id": "x6ra9jyl-zO3",
"colab_type": "code",
"colab": {}
},
"source": [
"# Calculate FLOPS\n",
"# https://stackoverflow.com/questions/49525776/how-to-calculate-a-mobilenet-flops-in-keras\n",
"# https://github.com/tensorflow/tensorflow/issues/32809\n",
"%tensorflow_version 2.x\n",
"def get_flops(model_h5_path):\n",
" session = tf.compat.v1.Session()\n",
" graph = tf.compat.v1.get_default_graph()\n",
" with graph.as_default():\n",
" with session.as_default():\n",
" # model = tf.keras.models.load_model(model_h5_path)\n",
" model = tf.keras.applications.mobilenet.MobileNet()\n",
" run_meta = tf.RunMetadata()\n",
" opts = tf.profiler.ProfileOptionBuilder.float_operation()\n",
"\n",
" # Optional: save printed results to file\n",
" # flops_log_path = os.path.join(tempfile.gettempdir(), 'tf_flops_log.txt')\n",
" # opts['output'] = 'file:outfile={}'.format(flops_log_path)\n",
"\n",
" # We use the Keras session graph in the call to the profiler.\n",
" flops = tf.profiler.profile(graph=graph,\n",
" run_meta=run_meta, cmd='op', options=opts)\n",
"\n",
" return flops.total_float_ops\n",
"\n",
"print((str)(get_flops(model_name + \".h5\") / 1000 / 1000 / 1000) + \" [GFLOPS]\")"
],
"execution_count": 0,
"outputs": []
},
{
"cell_type": "code",
"metadata": {
"id": "zR_nwg0NeuI-",
"colab_type": "code",
"colab": {}
},
"source": [
"def get_flops(model_h5_path):\n",
" session = tf.compat.v1.Session()\n",
" graph = tf.compat.v1.get_default_graph()\n",
" \n",
"\n",
" with graph.as_default():\n",
" with session.as_default():\n",
" model = tf.keras.models.load_model(model_h5_path)\n",
"\n",
" run_meta = tf.compat.v1.RunMetadata()\n",
" opts = tf.compat.v1.profiler.ProfileOptionBuilder.float_operation()\n",
" \n",
" # We use the Keras session graph in the call to the profiler.\n",
" flops = tf.compat.v1.profiler.profile(graph=graph,\n",
" run_meta=run_meta, cmd='op', options=opts)\n",
" \n",
" return flops.total_float_ops\n",
"print((str)(get_flops(model_name + \".h5\") / 1000 / 1000 / 1000) + \" [GFLOPS]\")"
],
"execution_count": 0,
"outputs": []
}
]
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment