{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_DDaAex5Q7u-"
      },
      "source": [
        "##### Copyright 2019 The TensorFlow Authors."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "W1dWWdNHQ9L0"
      },
      "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": "6Y8E0lw5eYWm"
      },
      "source": [
        "# 训练后整数量化"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "CIGrZZPTZVeO"
      },
      "source": [
        "<table class=\"tfo-notebook-buttons\" align=\"left\">\n",
        "  <td><a target=\"_blank\" href=\"https://tensorflow.google.cn/lite/performance/post_training_integer_quant\"><img src=\"https://tensorflow.google.cn/images/tf_logo_32px.png\">在 TensorFlow.org 查看</a></td>\n",
        "  <td><a target=\"_blank\" href=\"https://colab.research.google.com/github/tensorflow/docs-l10n/blob/master/site/zh-cn/lite/performance/post_training_integer_quant.ipynb\"><img src=\"https://tensorflow.google.cn/images/colab_logo_32px.png\">在 Google Colab 中运行 </a></td>\n",
        "  <td><a target=\"_blank\" href=\"https://github.com/tensorflow/docs-l10n/blob/master/site/zh-cn/lite/performance/post_training_integer_quant.ipynb\">     <img src=\"https://tensorflow.google.cn/images/GitHub-Mark-32px.png\">     在 GitHub 上查看源代码</a></td>\n",
        "  <td><a href=\"https://storage.googleapis.com/tensorflow_docs/docs-l10n/site/zh-cn/lite/performance/post_training_integer_quant.ipynb\"><img src=\"https://tensorflow.google.cn/images/download_logo_32px.png\">下载笔记本</a></td>\n",
        "</table>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "BTC1rDAuei_1"
      },
      "source": [
        "## 概述\n",
        "\n",
        "整数量化是一种优化策略，可将 32 位浮点数（如权重和激活输出）转换为 8 位定点数。这样可以缩减模型大小并加快推理速度，这对低功耗设备（如[微控制器](https://tensorflow.google.cn/lite/microcontrollers)）很有价值。仅支持整数的加速器（如 [Edge TPU](https://coral.ai/)）也需要使用此数据格式。\n",
        "\n",
        "在本教程中，您将从头开始训练一个 MNIST 模型、将其转换为 TensorFlow Lite 文件，并使用[训练后量化](https://tensorflow.google.cn/lite/performance/post_training_quantization)对其进行量化。最后，您将检查转换后模型的准确率并将其与原始浮点模型进行比较。\n",
        "\n",
        "实际上，对模型进行量化的程度有几种选项。在本教程中，您将执行“全整数量化”，它会将所有权重和激活输出转换为 8 位整数数据，而其他策略可能会将部分数据保留为浮点。\n",
        "\n",
        "要详细了解各种量化策略，请阅读 [TensorFlow Lite 模型优化](https://tensorflow.google.cn/lite/performance/model_optimization)。\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "dDqqUIZjZjac"
      },
      "source": [
        "## 设置"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "I0nR5AMEWq0H"
      },
      "source": [
        "为了量化输入和输出张量，我们需要使用 TensorFlow 2.3 中新添加的 API："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "WsN6s5L1ieNl"
      },
      "outputs": [],
      "source": [
        "import logging\n",
        "logging.getLogger(\"tensorflow\").setLevel(logging.DEBUG)\n",
        "\n",
        "import tensorflow as tf\n",
        "import numpy as np\n",
        "print(\"TensorFlow version: \", tf.__version__)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "2XsEP17Zelz9"
      },
      "source": [
        "## 生成 TensorFlow 模型"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5NMaNZQCkW9X"
      },
      "source": [
        "我们将构建一个简单的模型来对 [MNIST 数据集](https://tensorflow.google.cn/datasets/catalog/mnist)中的数字进行分类。\n",
        "\n",
        "此训练不会花很长时间，因为只对模型进行 5 个周期的训练，训练到约 98% 的准确率。"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "eMsw_6HujaqM"
      },
      "outputs": [],
      "source": [
        "# Load MNIST dataset\n",
        "mnist = tf.keras.datasets.mnist\n",
        "(train_images, train_labels), (test_images, test_labels) = mnist.load_data()\n",
        "\n",
        "# Normalize the input image so that each pixel value is between 0 to 1.\n",
        "train_images = train_images.astype(np.float32) / 255.0\n",
        "test_images = test_images.astype(np.float32) / 255.0\n",
        "\n",
        "# Define the model architecture\n",
        "model = tf.keras.Sequential([\n",
        "  tf.keras.layers.InputLayer(input_shape=(28, 28)),\n",
        "  tf.keras.layers.Reshape(target_shape=(28, 28, 1)),\n",
        "  tf.keras.layers.Conv2D(filters=12, kernel_size=(3, 3), activation='relu'),\n",
        "  tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),\n",
        "  tf.keras.layers.Flatten(),\n",
        "  tf.keras.layers.Dense(10)\n",
        "])\n",
        "\n",
        "# Train the digit classification model\n",
        "model.compile(optimizer='adam',\n",
        "              loss=tf.keras.losses.SparseCategoricalCrossentropy(\n",
        "                  from_logits=True),\n",
        "              metrics=['accuracy'])\n",
        "model.fit(\n",
        "  train_images,\n",
        "  train_labels,\n",
        "  epochs=5,\n",
        "  validation_data=(test_images, test_labels)\n",
        ")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "KuTEoGFYd8aM"
      },
      "source": [
        "## 转换为 TensorFlow Lite 模型"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "xl8_fzVAZwOh"
      },
      "source": [
        "现在，您可以使用TensorFlow Lite [Converter](https://tensorflow.google.cn/lite/models/convert) 将训练后的模型转换为 TensorFlow Lite 格式，并应用不同程度的量化。\n",
        "\n",
        "请注意，某些版本的量化会将部分数据保留为浮点格式。因此，以下各个部分将以量化程度不断增加的顺序展示每个选项，直到获得完全由 int8 或 uint8 数据组成的模型。（请注意，我们在每个部分中重复了一些代码，使您能够看到每个选项的全部量化步骤。）\n",
        "\n",
        "首先，下面是一个没有量化的转换后模型："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "_i8B2nDZmAgQ"
      },
      "outputs": [],
      "source": [
        "converter = tf.lite.TFLiteConverter.from_keras_model(model)\n",
        "\n",
        "tflite_model = converter.convert()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "7BONhYtYocQY"
      },
      "source": [
        "它现在是一个 TensorFlow Lite 模型，但所有参数数据仍使用 32 位浮点值。"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "jPYZwgZTwJMT"
      },
      "source": [
        "### 使用动态范围量化进行转换\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Hjvq1vpJd4U_"
      },
      "source": [
        "现在，我们启用默认的 `optimizations` 标记来量化所有固定参数（例如权重）："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "HEZ6ET1AHAS3"
      },
      "outputs": [],
      "source": [
        "converter = tf.lite.TFLiteConverter.from_keras_model(model)\n",
        "converter.optimizations = [tf.lite.Optimize.DEFAULT]\n",
        "\n",
        "tflite_model_quant = converter.convert()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "o5wuE-RcdX_3"
      },
      "source": [
        "现在，进行了权重量化的模型要略小一些，但其他变量数据仍为浮点格式。"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "UgKDdnHQEhpb"
      },
      "source": [
        "### 使用浮点回退量化进行转换"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "rTe8avZJHMDO"
      },
      "source": [
        "要量化可变数据（例如模型输入/输出和层之间的中间体），您需要提供 [`RepresentativeDataset`](https://tensorflow.google.cn/api_docs/python/tf/lite/RepresentativeDataset)。这是一个生成器函数，它提供一组足够大的输入数据来代表典型值。转换器可以通过该函数估算所有可变数据的动态范围。（相比训练或评估数据集，此数据集不必唯一。）为了支持多个输入，每个代表性数据点都是一个列表，并且列表中的元素会根据其索引被馈送到模型。\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "FiwiWU3gHdkW"
      },
      "outputs": [],
      "source": [
        "def representative_data_gen():\n",
        "  for input_value in tf.data.Dataset.from_tensor_slices(train_images).batch(1).take(100):\n",
        "    # Model has only one input so each data point has one element.\n",
        "    yield [input_value]\n",
        "\n",
        "converter = tf.lite.TFLiteConverter.from_keras_model(model)\n",
        "converter.optimizations = [tf.lite.Optimize.DEFAULT]\n",
        "converter.representative_dataset = representative_data_gen\n",
        "\n",
        "tflite_model_quant = converter.convert()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_GC3HFlptf7x"
      },
      "source": [
        "现在，所有权重和可变数据都已量化，并且与原始 TensorFlow Lite 模型相比，该模型要小得多。\n",
        "\n",
        "但是，为了与传统上使用浮点模型输入和输出张量的应用保持兼容，TensorFlow Lite 转换器将模型的输入和输出张量保留为浮点："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "id1OEKFELQwp"
      },
      "outputs": [],
      "source": [
        "interpreter = tf.lite.Interpreter(model_content=tflite_model_quant)\n",
        "input_type = interpreter.get_input_details()[0]['dtype']\n",
        "print('input: ', input_type)\n",
        "output_type = interpreter.get_output_details()[0]['dtype']\n",
        "print('output: ', output_type)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "RACBJuj2XO8x"
      },
      "source": [
        "这通常对兼容性有利，但它无法兼容执行全整数运算的设备（如 Edge TPU）。\n",
        "\n",
        "此外，如果 TensorFlow Lite 不包括某个运算的量化实现，则上述过程可能会将该运算保留为浮点格式。您仍能通过此策略完成转换，并得到一个更小、更高效的模型，但它还是不兼容仅支持整数的硬件。（此 MNIST 模型中的所有算子都有量化的实现。）\n",
        "\n",
        "因此，为了确保端到端全整数模型，您还需要几个参数…"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "FQgTqbvPvxGJ"
      },
      "source": [
        "### 使用仅整数量化进行转换"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "mwR9keYAwArA"
      },
      "source": [
        "为了量化输入和输出张量，并让转换器在遇到无法量化的运算时引发错误，使用一些附加参数再次转换模型："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "kzjEjcDs3BHa"
      },
      "outputs": [],
      "source": [
        "def representative_data_gen():\n",
        "  for input_value in tf.data.Dataset.from_tensor_slices(train_images).batch(1).take(100):\n",
        "    yield [input_value]\n",
        "\n",
        "converter = tf.lite.TFLiteConverter.from_keras_model(model)\n",
        "converter.optimizations = [tf.lite.Optimize.DEFAULT]\n",
        "converter.representative_dataset = representative_data_gen\n",
        "# Ensure that if any ops can't be quantized, the converter throws an error\n",
        "converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]\n",
        "# Set the input and output tensors to uint8 (APIs added in r2.3)\n",
        "converter.inference_input_type = tf.uint8\n",
        "converter.inference_output_type = tf.uint8\n",
        "\n",
        "tflite_model_quant = converter.convert()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "wYd6NxD03yjB"
      },
      "source": [
        "内部量化与上文相同，但您可以看到输入和输出张量现在是整数格式：\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "PaNkOS-twz4k"
      },
      "outputs": [],
      "source": [
        "interpreter = tf.lite.Interpreter(model_content=tflite_model_quant)\n",
        "input_type = interpreter.get_input_details()[0]['dtype']\n",
        "print('input: ', input_type)\n",
        "output_type = interpreter.get_output_details()[0]['dtype']\n",
        "print('output: ', output_type)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "TO17AP84wzBb"
      },
      "source": [
        "现在，您有了一个整数量化模型，该模型使用整数数据作为模型的输入和输出张量，因此它兼容仅支持整数的硬件（如 [Edge TPU](https://coral.ai)）。"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "sse224YJ4KMm"
      },
      "source": [
        "### 将模型另存为文件"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "4_9nZ4nv4b9P"
      },
      "source": [
        "您需要 `.tflite` 文件才能在其他设备上部署模型。因此，我们将转换的模型保存为文件，然后在下面运行推断时加载它们。"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "BEY59dC14uRv"
      },
      "outputs": [],
      "source": [
        "import pathlib\n",
        "\n",
        "tflite_models_dir = pathlib.Path(\"/tmp/mnist_tflite_models/\")\n",
        "tflite_models_dir.mkdir(exist_ok=True, parents=True)\n",
        "\n",
        "# Save the unquantized/float model:\n",
        "tflite_model_file = tflite_models_dir/\"mnist_model.tflite\"\n",
        "tflite_model_file.write_bytes(tflite_model)\n",
        "# Save the quantized model:\n",
        "tflite_model_quant_file = tflite_models_dir/\"mnist_model_quant.tflite\"\n",
        "tflite_model_quant_file.write_bytes(tflite_model_quant)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9t9yaTeF9fyM"
      },
      "source": [
        "## 运行 TensorFlow Lite 模型"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "L8lQHMp_asCq"
      },
      "source": [
        "现在，我们使用 TensorFlow Lite [`Interpreter`](https://tensorflow.google.cn/api_docs/python/tf/lite/Interpreter) 运行推断来比较模型的准确率。\n",
        "\n",
        "首先，我们需要一个函数，该函数使用给定的模型和图像运行推断，然后返回预测值：\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "X092SbeWfd1A"
      },
      "outputs": [],
      "source": [
        "# Helper function to run inference on a TFLite model\n",
        "def run_tflite_model(tflite_file, test_image_indices):\n",
        "  global test_images\n",
        "\n",
        "  # Initialize the interpreter\n",
        "  interpreter = tf.lite.Interpreter(model_path=str(tflite_file))\n",
        "  interpreter.allocate_tensors()\n",
        "\n",
        "  input_details = interpreter.get_input_details()[0]\n",
        "  output_details = interpreter.get_output_details()[0]\n",
        "\n",
        "  predictions = np.zeros((len(test_image_indices),), dtype=int)\n",
        "  for i, test_image_index in enumerate(test_image_indices):\n",
        "    test_image = test_images[test_image_index]\n",
        "    test_label = test_labels[test_image_index]\n",
        "\n",
        "    # Check if the input type is quantized, then rescale input data to uint8\n",
        "    if input_details['dtype'] == np.uint8:\n",
        "      input_scale, input_zero_point = input_details[\"quantization\"]\n",
        "      test_image = test_image / input_scale + input_zero_point\n",
        "\n",
        "    test_image = np.expand_dims(test_image, axis=0).astype(input_details[\"dtype\"])\n",
        "    interpreter.set_tensor(input_details[\"index\"], test_image)\n",
        "    interpreter.invoke()\n",
        "    output = interpreter.get_tensor(output_details[\"index\"])[0]\n",
        "\n",
        "    predictions[i] = output.argmax()\n",
        "\n",
        "  return predictions\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "2opUt_JTdyEu"
      },
      "source": [
        "### 在单个图像上测试模型\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "QpPpFPaz7eEM"
      },
      "source": [
        "现在，我们来比较一下浮点模型和量化模型的性能：\n",
        "\n",
        "- `tflite_model_file` 是使用浮点数据的原始 TensorFlow Lite 模型。\n",
        "- `tflite_model_quant_file` 是我们使用全整数量化转换的上一个模型（它使用 uint8 数据作为输入和输出）。\n",
        "\n",
        "我们来创建另一个函数打印预测值："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "zR2cHRUcUZ6e"
      },
      "outputs": [],
      "source": [
        "import matplotlib.pylab as plt\n",
        "\n",
        "# Change this to test a different image\n",
        "test_image_index = 1\n",
        "\n",
        "## Helper function to test the models on one image\n",
        "def test_model(tflite_file, test_image_index, model_type):\n",
        "  global test_labels\n",
        "\n",
        "  predictions = run_tflite_model(tflite_file, [test_image_index])\n",
        "\n",
        "  plt.imshow(test_images[test_image_index])\n",
        "  template = model_type + \" Model \\n True:{true}, Predicted:{predict}\"\n",
        "  _ = plt.title(template.format(true= str(test_labels[test_image_index]), predict=str(predictions[0])))\n",
        "  plt.grid(False)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "A5OTJ_6Vcslt"
      },
      "source": [
        "现在测试浮点模型："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "iTK0x980coto"
      },
      "outputs": [],
      "source": [
        "test_model(tflite_model_file, test_image_index, model_type=\"Float\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "o3N6-UGl1dfE"
      },
      "source": [
        "然后测试量化模型："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "rc1i9umMcp0t"
      },
      "outputs": [],
      "source": [
        "test_model(tflite_model_quant_file, test_image_index, model_type=\"Quantized\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "LwN7uIdCd8Gw"
      },
      "source": [
        "### 在所有图像上评估模型"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "RFKOD4DG8XmU"
      },
      "source": [
        "现在，我们使用在本教程开始时加载的所有测试图像来运行两个模型："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "05aeAuWjvjPx"
      },
      "outputs": [],
      "source": [
        "# Helper function to evaluate a TFLite model on all images\n",
        "def evaluate_model(tflite_file, model_type):\n",
        "  global test_images\n",
        "  global test_labels\n",
        "\n",
        "  test_image_indices = range(test_images.shape[0])\n",
        "  predictions = run_tflite_model(tflite_file, test_image_indices)\n",
        "\n",
        "  accuracy = (np.sum(test_labels== predictions) * 100) / len(test_images)\n",
        "\n",
        "  print('%s model accuracy is %.4f%% (Number of test samples=%d)' % (\n",
        "      model_type, accuracy, len(test_images)))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "xnFilQpBuMh5"
      },
      "source": [
        "评估浮点模型："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "T5mWkSbMcU5z"
      },
      "outputs": [],
      "source": [
        "evaluate_model(tflite_model_file, model_type=\"Float\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Km3cY9ry8ZlG"
      },
      "source": [
        "评估量化模型："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "-9cnwiPp6EGm"
      },
      "outputs": [],
      "source": [
        "evaluate_model(tflite_model_quant_file, model_type=\"Quantized\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "L7lfxkor8pgv"
      },
      "source": [
        "现在您有了一个整数量化模型，该模型的准确率与浮点模型相比几乎没有差别。\n",
        "\n",
        "要详细了解其他量化策略，请阅读 [TensorFlow Lite 模型优化](https://tensorflow.google.cn/lite/performance/model_optimization)。"
      ]
    }
  ],
  "metadata": {
    "colab": {
      "collapsed_sections": [],
      "name": "post_training_integer_quant.ipynb",
      "toc_visible": true
    },
    "kernelspec": {
      "display_name": "Python 3",
      "name": "python3"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}