{ "cells": [ { "cell_type": "markdown", "metadata": { "id": "wJcYs_ERTnnI" }, "source": [ "##### Copyright 2021 The TensorFlow Authors." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "cellView": "form", "id": "HMUDt0CiUJk9" }, "outputs": [], "source": [ "#@title Licensed under the Apache License, Version 2.0 (the \"License\");\n", "# you may not use this file except in compliance with the License.\n", "# You may obtain a copy of the License at\n", "#\n", "# https://www.apache.org/licenses/LICENSE-2.0\n", "#\n", "# Unless required by applicable law or agreed to in writing, software\n", "# distributed under the License is distributed on an \"AS IS\" BASIS,\n", "# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n", "# See the License for the specific language governing permissions and\n", "# limitations under the License." ] }, { "cell_type": "markdown", "metadata": { "id": "77z2OchJTk0l" }, "source": [ "# Migrating your TFLite code to TF2\n", "\n", "\n", " \n", " \n", " \n", " \n", "
\n", " \n", " \n", " View on TensorFlow.org\n", " \n", " \n", " \n", " Run in Google Colab\n", " \n", " \n", " \n", " View source on GitHub\n", " \n", " Download notebook\n", "
" ] }, { "cell_type": "markdown", "metadata": { "id": "meUTrR4I6m1C" }, "source": [ "[TensorFlow Lite](https://www.tensorflow.org/lite/guide) (TFLite) is a set of tools that helps developers run ML inference on-device (mobile, embedded, and IoT devices). The [TFLite converter](https://www.tensorflow.org/lite/convert) is one such tool that converts existing TF models into an optimized TFLite model format that can be efficiently run on-device.\n", "\n", "In this doc, you'll learn what changes you need to make to your TF to TFLite conversion code, followed by a few examples that do the same.\n", "\n", "\n", "## Changes to your TF to TFLite conversion code\n", "\n", "* If you're using a legacy TF1 model format (such as Keras file, frozen GraphDef, checkpoints, tf.Session), update it to TF1/TF2 SavedModel and use the TF2 converter API `tf.lite.TFLiteConverter.from_saved_model(...)` to convert it to a TFLite model (refer to Table 1).\n", "\n", "* Update the converter API flags (refer to Table 2).\n", "* Remove legacy APIs such as `tf.lite.constants`. (eg: Replace `tf.lite.constants.INT8` with `tf.int8`)\n", "\n", "// Table 1 // TFLite Python Converter API Update\n", "\n", "TF1 API | TF2 API |\n", "--- | --- |\n", "`tf.lite.TFLiteConverter.from_saved_model('saved_model/',..)` | *supported* |\n", "`tf.lite.TFLiteConverter.from_keras_model_file('model.h5',..)` | *removed (update to SavedModel format)* |\n", "`tf.lite.TFLiteConverter.from_frozen_graph('model.pb',..)` | *removed (update to SavedModel format)* |\n", "`tf.lite.TFLiteConverter.from_session(sess,...)` | *removed (update to SavedModel format)* |" ] }, { "cell_type": "markdown", "metadata": { "id": "Rf75rjeedigq" }, "source": [ "" ] }, { "cell_type": "markdown", "metadata": { "id": "XbVlZNizW1-Y" }, "source": [ "// Table 2 // TFLite Python Converter API Flags Update\n", "\n", "TF1 API | TF2 API |\n", "--- | --- |\n", "`allow_custom_ops`
`optimizations`
`representative_dataset`
`target_spec`
`inference_input_type`
`inference_output_type`
`experimental_new_converter`
`experimental_new_quantizer` | *supported*







|\n", "`input_tensors`
`output_tensors`
`input_arrays_with_shape`
`output_arrays`
`experimental_debug_info_func`| *removed (unsupported converter API arguments)*




|\n", "`change_concat_input_ranges`
`default_ranges_stats`
`get_input_arrays()`
`inference_type`
`quantized_input_stats`
`reorder_across_fake_quant` | *removed (unsupported quantization workflows)*





|\n", "`conversion_summary_dir`
`dump_graphviz_dir`
`dump_graphviz_video` | *removed (instead, visualize models using [Netron](https://lutzroeder.github.io/netron/) or [visualize.py](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tools/visualize.py))*


|\n", "`output_format`
`drop_control_dependency` | *removed (unsupported features in TF2)*

|" ] }, { "cell_type": "markdown", "metadata": { "id": "YdZSoIXEbhg-" }, "source": [ "## Examples\n", "\n", "You'll now walk through some examples to convert legacy TF1 models to TF1/TF2 SavedModels and then convert them to TF2 TFLite models.\n", "\n", "### Setup\n", "\n", "Start with the necessary TensorFlow imports." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "iE0vSfMXumKI" }, "outputs": [], "source": [ "import tensorflow as tf\n", "import tensorflow.compat.v1 as tf1\n", "import numpy as np\n", "\n", "import logging\n", "logger = tf.get_logger()\n", "logger.setLevel(logging.ERROR)\n", "\n", "import shutil\n", "def remove_dir(path):\n", " try:\n", " shutil.rmtree(path)\n", " except:\n", " pass" ] }, { "cell_type": "markdown", "metadata": { "id": "89VllCprnFto" }, "source": [ "Create all the necessary TF1 model formats." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "Bwq8EFiwjzjx" }, "outputs": [], "source": [ "# Create a TF1 SavedModel\n", "SAVED_MODEL_DIR = \"tf_saved_model/\"\n", "remove_dir(SAVED_MODEL_DIR)\n", "with tf1.Graph().as_default() as g:\n", " with tf1.Session() as sess:\n", " input = tf1.placeholder(tf.float32, shape=(3,), name='input')\n", " output = input + 2\n", " # print(\"result: \", sess.run(output, {input: [0., 2., 4.]}))\n", " tf1.saved_model.simple_save(\n", " sess, SAVED_MODEL_DIR,\n", " inputs={'input': input}, \n", " outputs={'output': output})\n", "print(\"TF1 SavedModel path: \", SAVED_MODEL_DIR)\n", "\n", "# Create a TF1 Keras model\n", "KERAS_MODEL_PATH = 'tf_keras_model.h5'\n", "model = tf1.keras.models.Sequential([\n", " tf1.keras.layers.InputLayer(input_shape=(128, 128, 3,), name='input'),\n", " tf1.keras.layers.Dense(units=16, input_shape=(128, 128, 3,), activation='relu'),\n", " tf1.keras.layers.Dense(units=1, name='output')\n", "])\n", "model.save(KERAS_MODEL_PATH, save_format='h5')\n", "print(\"TF1 Keras Model path: \", KERAS_MODEL_PATH)\n", "\n", "# Create a TF1 frozen GraphDef model\n", "GRAPH_DEF_MODEL_PATH = tf.keras.utils.get_file(\n", " 'mobilenet_v1_0.25_128',\n", " origin='https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_0.25_128_frozen.tgz',\n", " untar=True,\n", ") + '/frozen_graph.pb'\n", "\n", "print(\"TF1 frozen GraphDef path: \", GRAPH_DEF_MODEL_PATH)" ] }, { "cell_type": "markdown", "metadata": { "id": "EzMBpG5rdt-7" }, "source": [ "### 1. Convert a TF1 SavedModel to a TFLite model\n" ] }, { "cell_type": "markdown", "metadata": { "id": "GFWIlVridt_F" }, "source": [ "#### Before: Converting with TF1\n", "This is typical code for TF1-style TFlite conversion.\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "dzXHHBQRdt_F" }, "outputs": [], "source": [ "converter = tf1.lite.TFLiteConverter.from_saved_model(\n", " saved_model_dir=SAVED_MODEL_DIR,\n", " input_arrays=['input'],\n", " input_shapes={'input' : [3]}\n", ")\n", "converter.optimizations = {tf.lite.Optimize.DEFAULT}\n", "converter.change_concat_input_ranges = True\n", "tflite_model = converter.convert()\n", "# Ignore warning: \"Use '@tf.function' or '@defun' to decorate the function.\"" ] }, { "cell_type": "markdown", "metadata": { "id": "NUptsxK_MUy2" }, "source": [ "#### After: Converting with TF2\n", "\n", "Directly convert the TF1 SavedModel to a TFLite model, with a smaller v2 converter flags set." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "0OyBjZ6Kdt_F" }, "outputs": [], "source": [ "# Convert TF1 SavedModel to a TFLite model.\n", "converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir=SAVED_MODEL_DIR)\n", "converter.optimizations = {tf.lite.Optimize.DEFAULT}\n", "tflite_model = converter.convert()" ] }, { "cell_type": "markdown", "metadata": { "id": "yiwu3sso__fH" }, "source": [ "### 2. Convert a TF1 Keras model file to a TFLite model" ] }, { "cell_type": "markdown", "metadata": { "id": "9WTPvPih__fR" }, "source": [ "#### Before: Converting with TF1\n", "This is typical code for TF1-style TFlite conversion." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "9EXO0xYq__fR" }, "outputs": [], "source": [ "converter = tf1.lite.TFLiteConverter.from_keras_model_file(model_file=KERAS_MODEL_PATH)\n", "converter.optimizations = {tf.lite.Optimize.DEFAULT}\n", "converter.change_concat_input_ranges = True\n", "tflite_model = converter.convert()" ] }, { "cell_type": "markdown", "metadata": { "id": "9l6ppTtTZ5Bz" }, "source": [ "#### After: Converting with TF2\n", "\n", "First, convert the TF1 Keras model file to a TF2 SavedModel and then convert it to a TFLite model, with a smaller v2 converter flags set." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "IGB5ZMGl__fR" }, "outputs": [], "source": [ "# Convert TF1 Keras model file to TF2 SavedModel.\n", "model = tf.keras.models.load_model(KERAS_MODEL_PATH)\n", "model.save(filepath='saved_model_2/')\n", "\n", "# Convert TF2 SavedModel to a TFLite model.\n", "converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir='saved_model_2/')\n", "tflite_model = converter.convert()" ] }, { "cell_type": "markdown", "metadata": { "id": "v5Zf6G4M-sZz" }, "source": [ "### 3. Convert a TF1 frozen GraphDef to a TFLite model\n" ] }, { "cell_type": "markdown", "metadata": { "id": "DzCJOV7AUlGZ" }, "source": [ "#### Before: Converting with TF1\n", "\n", "This is typical code for TF1-style TFlite conversion." ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "r7RvcdRv6lll" }, "outputs": [], "source": [ "converter = tf1.lite.TFLiteConverter.from_frozen_graph(\n", " graph_def_file=GRAPH_DEF_MODEL_PATH,\n", " input_arrays=['input'],\n", " input_shapes={'input' : [1, 128, 128, 3]},\n", " output_arrays=['MobilenetV1/Predictions/Softmax'],\n", ")\n", "converter.optimizations = {tf.lite.Optimize.DEFAULT}\n", "converter.change_concat_input_ranges = True\n", "tflite_model = converter.convert()" ] }, { "cell_type": "markdown", "metadata": { "id": "ZdIogJsKaMNH" }, "source": [ "#### After: Converting with TF2\n", "\n", "First, convert the TF1 frozen GraphDef to a TF1 SavedModel and then convert it to a TFLite model, with a smaller v2 converter flags set.\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "Oigap0TZxjWG" }, "outputs": [], "source": [ "## Convert TF1 frozen Graph to TF1 SavedModel.\n", "\n", "# Load the graph as a v1.GraphDef\n", "import pathlib\n", "gdef = tf.compat.v1.GraphDef()\n", "gdef.ParseFromString(pathlib.Path(GRAPH_DEF_MODEL_PATH).read_bytes())\n", "\n", "# Convert the GraphDef to a tf.Graph\n", "with tf.Graph().as_default() as g:\n", " tf.graph_util.import_graph_def(gdef, name=\"\")\n", "\n", "# Look up the input and output tensors.\n", "input_tensor = g.get_tensor_by_name('input:0') \n", "output_tensor = g.get_tensor_by_name('MobilenetV1/Predictions/Softmax:0')\n", "\n", "# Save the graph as a TF1 Savedmodel\n", "remove_dir('saved_model_3/')\n", "with tf.compat.v1.Session(graph=g) as s:\n", " tf.compat.v1.saved_model.simple_save(\n", " session=s,\n", " export_dir='saved_model_3/',\n", " inputs={'input':input_tensor},\n", " outputs={'output':output_tensor})\n", "\n", "# Convert TF1 SavedModel to a TFLite model.\n", "converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir='saved_model_3/')\n", "converter.optimizations = {tf.lite.Optimize.DEFAULT}\n", "tflite_model = converter.convert()" ] }, { "cell_type": "markdown", "metadata": { "id": "MFbsddkOw4Wl" }, "source": [ "# Further reading\n", "\n", "* Refer to the [TFLite Guide](https://www.tensorflow.org/lite/guide) to learn more about the workflows and latest features.\n", "* If you're using TF1 code or legacy TF1 model formats (Keras `.h5` files, frozen GraphDef `.pb`, etc), please update your code and migrate your models to the [TF2 SavedModel format](https://www.tensorflow.org/guide/saved_model). \n" ] } ], "metadata": { "colab": { "collapsed_sections": [], "name": "tflite.ipynb", "toc_visible": true }, "kernelspec": { "display_name": "Python 3", "name": "python3" } }, "nbformat": 4, "nbformat_minor": 0 }