005-keras_saving_and_serializing_model.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# tensorflow2教程-keras模型保持和序列化\n",
    "\n",
    "## 1.保持序列模型和函数模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"3_layer_mlp\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "digits (InputLayer)          [(None, 784)]             0         \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 64)                50240     \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 64)                4160      \n",
      "_________________________________________________________________\n",
      "predictions (Dense)          (None, 10)                650       \n",
      "=================================================================\n",
      "Total params: 55,050\n",
      "Trainable params: 55,050\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "60000/60000 [==============================] - 2s 29us/sample - loss: 0.3116\n"
     ]
    }
   ],
   "source": [
    "# 构建一个简单的模型并训练\n",
    "from __future__ import absolute_import, division, print_function\n",
    "import tensorflow as tf\n",
    "tf.keras.backend.clear_session()\n",
    "from tensorflow import keras\n",
    "from tensorflow.keras import layers\n",
    "\n",
    "inputs = keras.Input(shape=(784,), name='digits')\n",
    "x = layers.Dense(64, activation='relu', name='dense_1')(inputs)\n",
    "x = layers.Dense(64, activation='relu', name='dense_2')(x)\n",
    "outputs = layers.Dense(10, activation='softmax', name='predictions')(x)\n",
    "\n",
    "model = keras.Model(inputs=inputs, outputs=outputs, name='3_layer_mlp')\n",
    "model.summary()\n",
    "(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()\n",
    "x_train = x_train.reshape(60000, 784).astype('float32') / 255\n",
    "x_test = x_test.reshape(10000, 784).astype('float32') / 255\n",
    "\n",
    "model.compile(loss='sparse_categorical_crossentropy',\n",
    "              optimizer=keras.optimizers.RMSprop())\n",
    "history = model.fit(x_train, y_train,\n",
    "                    batch_size=64,\n",
    "                    epochs=1)\n",
    "\n",
    "predictions = model.predict(x_test)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.1保持全模型\n",
    "可以对整个模型进行保存，其保持的内容包括：\n",
    "- 该模型的架构\n",
    "- 模型的权重（在训练期间学到的）\n",
    "- 模型的训练配置（你传递给编译的），如果有的话\n",
    "- 优化器及其状态（如果有的话）（这使您可以从中断的地方重新启动训练）\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "model.save('the_save_model.h5')\n",
    "new_model = keras.models.load_model('the_save_model.h5')\n",
    "new_prediction = new_model.predict(x_test)\n",
    "np.testing.assert_allclose(predictions, new_prediction, atol=1e-6) # 预测结果一样\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.2 保持为SavedModel文件"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "WARNING: Logging before flag parsing goes to stderr.\n",
      "W0311 23:50:02.399847 139796031059712 deprecation.py:323] From /home/czy/anaconda3/envs/tf2_0/lib/python3.6/site-packages/tensorflow/python/saved_model/signature_def_utils_impl.py:253: build_tensor_info (from tensorflow.python.saved_model.utils_impl) is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "This function will only be available through the v1 compatibility library as tf.compat.v1.saved_model.utils.build_tensor_info or tf.compat.v1.saved_model.build_tensor_info.\n",
      "W0311 23:50:02.400944 139796031059712 tf_logging.py:161] Export includes no default signature!\n",
      "W0311 23:50:02.721019 139796031059712 tf_logging.py:161] Export includes no default signature!\n"
     ]
    }
   ],
   "source": [
    "keras.experimental.export_saved_model(model, 'saved_model')\n",
    "new_model = keras.experimental.load_from_saved_model('saved_model')\n",
    "new_prediction = new_model.predict(x_test)\n",
    "np.testing.assert_allclose(predictions, new_prediction, atol=1e-6) # 预测结果一样\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.3仅保存网络结构\n",
    "仅保持网络结构，这样导出的模型并未包含训练好的参数"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "config = model.get_config()\n",
    "reinitialized_model = keras.Model.from_config(config)\n",
    "new_prediction = reinitialized_model.predict(x_test)\n",
    "assert abs(np.sum(predictions-new_prediction)) >0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "也可以使用json保存网络结构"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "json_config = model.to_json()\n",
    "reinitialized_model = keras.models.model_from_json(json_config)\n",
    "new_prediction = reinitialized_model.predict(x_test)\n",
    "assert abs(np.sum(predictions-new_prediction)) >0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.4仅保存网络参数"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "weights = model.get_weights()\n",
    "model.set_weights(weights)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 可以把结构和参数保存结合起来\n",
    "config = model.get_config()\n",
    "weights = model.get_weights()\n",
    "new_model = keras.Model.from_config(config) # config只能用keras.Model的这个api\n",
    "new_model.set_weights(weights)\n",
    "new_predictions = new_model.predict(x_test)\n",
    "np.testing.assert_allclose(predictions, new_predictions, atol=1e-6)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.5完整的模型保持方法"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "json_config = model.to_json()\n",
    "with open('model_config.json', 'w') as json_file:\n",
    "    json_file.write(json_config)\n",
    "\n",
    "model.save_weights('path_to_my_weights.h5')\n",
    "\n",
    "with open('model_config.json') as json_file:\n",
    "    json_config = json_file.read()\n",
    "new_model = keras.models.model_from_json(json_config)\n",
    "new_model.load_weights('path_to_my_weights.h5')\n",
    "\n",
    "new_predictions = new_model.predict(x_test)\n",
    "np.testing.assert_allclose(predictions, new_predictions, atol=1e-6)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 当然也可以一步到位\n",
    "model.save('path_to_my_model.h5')\n",
    "del model\n",
    "model = keras.models.load_model('path_to_my_model.h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.6保存网络权重为SavedModel格式"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save_weights('weight_tf_savedmodel')\n",
    "model.save_weights('weight_tf_savedmodel_h5', save_format='h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.7子类模型参数保存\n",
    "子类模型的结构无法保存和序列化，只能保持参数"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 构建模型\n",
    "class ThreeLayerMLP(keras.Model):\n",
    "  \n",
    "    def __init__(self, name=None):\n",
    "        super(ThreeLayerMLP, self).__init__(name=name)\n",
    "        self.dense_1 = layers.Dense(64, activation='relu', name='dense_1')\n",
    "        self.dense_2 = layers.Dense(64, activation='relu', name='dense_2')\n",
    "        self.pred_layer = layers.Dense(10, activation='softmax', name='predictions')\n",
    "\n",
    "    def call(self, inputs):\n",
    "        x = self.dense_1(inputs)\n",
    "        x = self.dense_2(x)\n",
    "        return self.pred_layer(x)\n",
    "\n",
    "def get_model():\n",
    "    return ThreeLayerMLP(name='3_layer_mlp')\n",
    "\n",
    "model = get_model()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "60000/60000 [==============================] - 2s 28us/sample - loss: 0.3217\n"
     ]
    }
   ],
   "source": [
    "# 训练模型\n",
    "(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()\n",
    "x_train = x_train.reshape(60000, 784).astype('float32') / 255\n",
    "x_test = x_test.reshape(10000, 784).astype('float32') / 255\n",
    "\n",
    "model.compile(loss='sparse_categorical_crossentropy',\n",
    "              optimizer=keras.optimizers.RMSprop())\n",
    "history = model.fit(x_train, y_train,\n",
    "                    batch_size=64,\n",
    "                    epochs=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 保持权重参数\n",
    "model.save_weights('my_model_weights', save_format='tf')\n",
    "\n",
    "# 输出结果，供后面对比\n",
    "\n",
    "predictions = model.predict(x_test)\n",
    "first_batch_loss = model.train_on_batch(x_train[:64], y_train[:64])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 读取保存的模型参数\n",
    "new_model = get_model()\n",
    "new_model.compile(loss='sparse_categorical_crossentropy',\n",
    "                  optimizer=keras.optimizers.RMSprop())\n",
    "\n",
    "#new_model.train_on_batch(x_train[:1], y_train[:1])\n",
    "\n",
    "new_model.load_weights('my_model_weights')\n",
    "\n",
    "new_predictions = new_model.predict(x_test)\n",
    "np.testing.assert_allclose(predictions, new_predictions, atol=1e-6)\n",
    "\n",
    "\n",
    "new_first_batch_loss = new_model.train_on_batch(x_train[:64], y_train[:64])\n",
    "assert first_batch_loss == new_first_batch_loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}