{
  "cells": [
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "q9KfUf1BI6Kl"
      },
      "outputs": [],
      "source": [
        "##### Copyright 2021 The TensorFlow Authors."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "cellView": "form",
        "id": "WvqLCVQ6I58i",
        "vscode": {
          "languageId": "python"
        }
      },
      "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": "ZoFRICPTNUca"
      },
      "source": [
        "# 迁移 SavedModel 工作流\n",
        "\n",
        "<table class=\"tfo-notebook-buttons\" align=\"left\">\n",
        "  <td>     <a target=\"_blank\" href=\"https://tensorflow.google.cn/guide/migrate/saved_model\"><img src=\"https://tensorflow.google.cn/images/tf_logo_32px.png\">在 TensorFlow.org 上查看</a>   </td>\n",
        "  <td>     <a target=\"_blank\" href=\"https://colab.research.google.com/github/tensorflow/docs-l10n/blob/master/site/zh-cn/guide/migrate/saved_model.ipynb\"><img src=\"https://tensorflow.google.cn/images/colab_logo_32px.png\">在 Google Colab 运行</a> </td>\n",
        "  <td>     <a target=\"_blank\" href=\"https://github.com/tensorflow/docs-l10n/blob/master/site/zh-cn/guide/migrate/saved_model.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/guide/migrate/saved_model.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": "nGyIb7MYJfaM"
      },
      "source": [
        "将模型从 TensorFlow 1 的计算图和会话迁移到 TensorFlow 2 API（例如 `tf.function`、`tf.Module` 和 `tf.keras.Model`）后，您可以迁移模型保存和加载代码。此笔记本提供了如何在 TensorFlow 1 和 TensorFlow 2 中以 SavedModel 格式保存和加载的示例。下面是从 TensorFlow 1 迁移到 TensorFlow 2 的相关 API 变更的快速概览：\n",
        "\n",
        " | TensorFlow 1 | 迁移到 TensorFlow 2\n",
        "--- | --- | ---\n",
        "**保存** | `tf.compat.v1.saved_model.Builder`<br>`tf.compat.v1.saved_model.simple_save` | `tf.saved_model.save`<br>Keras：`tf.keras.models.save_model`\n",
        "**加载** | `tf.compat.v1.saved_model.load` | `tf.saved_model.load`<br>Keras：`tf.keras.models.load_model`\n",
        "**签名**：一组输入<br>和输出张量，<br>可用于运行<br> | 使用 `*.signature_def` 效用函数生成<br>例如 `tf.compat.v1.saved_model.predict_signature_def`） | 编写一个 `tf.function` 并使用 `tf.saved_model.save` 中的 <code>signatures</code> 参数将其导出。<br>\n",
        "**分类<br>和回归**：<br>特殊类型的签名 | 使用<br>`tf.compat.v1.saved_model.classification_signature_def`、<br>`tf.compat.v1.saved_model.regression_signature_def` <br>和某些 Estimator 导出生成。 | 这两种签名类型已从 TensorFlow 2 中移除。<br>如果应用库需要这些方法名称，<br>可以使用 `tf.compat.v1.saved_model.signature_def_utils.MethodNameUpdater`。\n",
        "\n",
        "有关映射的更深入解释，请参阅下面的从 TensorFlow 1 到 TensorFlow 2 的变化部分。"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "r5mR2xsNAGsB"
      },
      "source": [
        "## 安装\n",
        "\n",
        "下面的示例展示了如何使用 TensorFlow 1 和 TensorFlow 2 API 将相同的虚拟 TensorFlow 模型（定义为下面的 `add_two`）导出并加载到 SavedModel 格式。首先，设置导入和效用函数："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "B94QZyy-kOGQ",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "import tensorflow as tf\n",
        "import tensorflow.compat.v1 as tf1\n",
        "import shutil\n",
        "\n",
        "def remove_dir(path):\n",
        "  try:\n",
        "    shutil.rmtree(path)\n",
        "  except:\n",
        "    pass\n",
        "\n",
        "def add_two(input):\n",
        "  return input + 2"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ZNVpH5tOCgd9"
      },
      "source": [
        "## TensorFlow 1：保存和导出 SavedModel\n",
        "\n",
        "在 TensorFlow 1 中，使用 `tf.compat.v1.saved_model.Builder`、`tf.compat.v1.saved_model.simple_save` 和 `tf.estimator.Estimator.export_saved_model` API 来构建、保存及导出 TensorFlow 计算图和会话："
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "THRLul5ijmTE"
      },
      "source": [
        "### 1. 使用 SavedModelBuilder 将计算图保存为 SavedModel"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "dcZDQaI8jl3h",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "remove_dir(\"saved-model-builder\")\n",
        "\n",
        "with tf.Graph().as_default() as g:\n",
        "  with tf1.Session() as sess:\n",
        "    input = tf1.placeholder(tf.float32, shape=[])\n",
        "    output = add_two(input)\n",
        "    print(\"add two output: \", sess.run(output, {input: 3.}))\n",
        "\n",
        "    # Save with SavedModelBuilder\n",
        "    builder = tf1.saved_model.Builder('saved-model-builder')\n",
        "    sig_def = tf1.saved_model.predict_signature_def(\n",
        "        inputs={'input': input},\n",
        "        outputs={'output': output})\n",
        "    builder.add_meta_graph_and_variables(\n",
        "        sess, tags=[\"serve\"], signature_def_map={\n",
        "            tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY: sig_def\n",
        "    })\n",
        "    builder.save()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "PwtC27VFlwCa",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "!saved_model_cli run --dir saved-model-builder --tag_set serve \\\n",
        " --signature_def serving_default --input_exprs input=10"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "gnBDNTxKG_vR"
      },
      "source": [
        "### 2. 为应用构建 SavedModel"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "jtMxe2rjHSq9",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "remove_dir(\"simple-save\")\n",
        "\n",
        "with tf.Graph().as_default() as g:\n",
        "  with tf1.Session() as sess:\n",
        "    input = tf1.placeholder(tf.float32, shape=[])\n",
        "    output = add_two(input)\n",
        "    print(\"add_two output: \", sess.run(output, {input: 3.}))\n",
        "\n",
        "    tf1.saved_model.simple_save(\n",
        "        sess, 'simple-save',\n",
        "        inputs={'input': input},\n",
        "        outputs={'output': output})"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "AdnqemvIHb2P",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "!saved_model_cli run --dir simple-save --tag_set serve \\\n",
        " --signature_def serving_default --input_exprs input=10"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "r0BNzzAHjnkp"
      },
      "source": [
        "### 3. 将 Estimator 推断计算图导出为 SavedModel\n",
        "\n",
        "在 Estimator `model_fn`（定义如下）的定义中，您可以通过在 `tf.estimator.EstimatorSpec` 中返回 `export_outputs` 来定义模型中的签名。下面是不同类型的输出：\n",
        "\n",
        "- `tf.estimator.export.ClassificationOutput`\n",
        "- `tf.estimator.export.RegressionOutput`\n",
        "- `tf.estimator.export.PredictOutput`\n",
        "\n",
        "这些输出将分别产生分类、回归和预测特征类型。\n",
        "\n",
        "使用 `tf.estimator.Estimator.export_saved_model` 导出 Estimator 时，这些签名将随模型一起保存。"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "3nQ5Stnxjhfs",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "def model_fn(features, labels, mode):\n",
        "  output = add_two(features['input'])\n",
        "  step = tf1.train.get_global_step()\n",
        "  return tf.estimator.EstimatorSpec(\n",
        "      mode,\n",
        "      predictions=output,\n",
        "      train_op=step.assign_add(1),\n",
        "      loss=tf.constant(0.),\n",
        "      export_outputs={\n",
        "          tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY: \\\n",
        "          tf.estimator.export.PredictOutput({'output': output})})\n",
        "est = tf.estimator.Estimator(model_fn, 'estimator-checkpoints')\n",
        "\n",
        "# Train for one step to create a checkpoint.\n",
        "def train_fn():\n",
        "  return tf.data.Dataset.from_tensors({'input': 3.})\n",
        "est.train(train_fn, steps=1)\n",
        "\n",
        "# This utility function `build_raw_serving_input_receiver_fn` takes in raw\n",
        "# tensor features and builds an \"input serving receiver function\", which\n",
        "# creates placeholder inputs to the model.\n",
        "serving_input_fn = tf.estimator.export.build_raw_serving_input_receiver_fn(\n",
        "    {'input': tf.constant(3.)})  # Pass in a dummy input batch.\n",
        "estimator_path = est.export_saved_model('exported-estimator', serving_input_fn)\n",
        "\n",
        "# Estimator's export_saved_model creates a time stamped directory. Move this\n",
        "# to a set path so it can be inspected with `saved_model_cli` in the cell below.\n",
        "!rm -rf estimator-model\n",
        "import shutil\n",
        "shutil.move(estimator_path, 'estimator-model')"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "8_gD2gkE7CMu",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "!saved_model_cli run --dir estimator-model --tag_set serve \\\n",
        " --signature_def serving_default --input_exprs input=[10]"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "DyBvrNQgIhIo"
      },
      "source": [
        "## TensorFlow 2：保存和导出 SavedModel"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "BZmFH-eIjqjB"
      },
      "source": [
        "### 保存并导出使用 tf.Module 定义的 SavedModel\n",
        "\n",
        "要在 TensorFlow 2 中导出模型，必须定义 `tf.Module` 或 `tf.keras.Model` 来保存模型的所有变量和函数。随后，可以调用 `tf.saved_model.save` 来创建 SavedModel。请参阅[使用 SavedModel 格式](../saved_model.ipynb)指南中的*保存自定义模型*部分来了解详情。"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "_j-PwgP_jrgw",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "class MyModel(tf.Module):\n",
        "  @tf.function\n",
        "  def __call__(self, input):\n",
        "    return add_two(input)\n",
        "\n",
        "model = MyModel()\n",
        "\n",
        "@tf.function\n",
        "def serving_default(input):\n",
        "  return {'output': model(input)}\n",
        "\n",
        "signature_function = serving_default.get_concrete_function(\n",
        "    tf.TensorSpec(shape=[], dtype=tf.float32))\n",
        "tf.saved_model.save(\n",
        "    model, 'tf2-save', signatures={\n",
        "        tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature_function})"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "slvU4vZN756F",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "!saved_model_cli run --dir tf2-save --tag_set serve \\\n",
        " --signature_def serving_default --input_exprs input=10"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "UYpSfbBJjr33"
      },
      "source": [
        "### 保存并导出使用 Keras 定义的 SavedModel\n",
        "\n",
        "已弃用：对于 Keras 对象，建议使用新的高级 `.keras` 格式和 `tf.keras.Model.export`，如[此处](https://tensorflow.google.cn/guide/keras/save_and_serialize)的指南所示。对于现有代码，继续支持低级 SavedModel 格式。\n",
        "\n",
        "用于保存和导出的 Keras API（`Mode.save` 或 `tf.keras.models.save_model`）可以从 `tf.keras.Model` 导出 SavedModel。请查看[保存和加载 Keras 模型](../..guide/keras/save_and_serialize)，了解更多详细信息。"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "mMcjhzyRjvp6",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "inp = tf.keras.Input(3)\n",
        "out = add_two(inp)\n",
        "model = tf.keras.Model(inputs=inp, outputs=out)\n",
        "\n",
        "@tf.function(input_signature=[tf.TensorSpec(shape=[], dtype=tf.float32)])\n",
        "def serving_default(input):\n",
        "  return {'output': model(input)}\n",
        "\n",
        "model.save('keras-model', save_format='tf', signatures={\n",
        "        tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY: serving_default})"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "4P93WP5R7-VT",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "!saved_model_cli run --dir keras-model --tag_set serve \\\n",
        " --signature_def serving_default --input_exprs input=10"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "SEKe9rGgoGCw"
      },
      "source": [
        "## 加载 SavedModel\n",
        "\n",
        "可以使用 TensorFlow 1 或 TensorFlow 2 API 加载使用上述任何 API 保存的 SavedModel。\n",
        "\n",
        "TensorFlow 1 SavedModel 在加载到 TensorFlow 2 时通常可用于推断，但只有在 SavedModel 包含*资源变量*时才能进行训练（生成梯度）。您可以检查变量的数据类型，如果变量数据类型包含“_ref\"”，则它是引用变量。\n",
        "\n",
        "只要 SavedModel 随签名一起保存，就可以在 TensorFlow 1 中加载和执行 TensorFlow 2 SavedModel。\n",
        "\n",
        "下面的部分包含代码示例，展示了如何加载前面部分中保存的 SavedModel 以及调用导出的签名。"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "hLztK_0YoTEP"
      },
      "source": [
        "### TensorFlow 1：使用 tf.saved_model.load 加载 SavedModel\n",
        "\n",
        "在 TensorFlow 1 中，可以使用 `tf.saved_model.load` 将 SavedModel 直接导入当前计算图和会话。可以在张量输入和输出名称上调用 `Session.run`："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "IMO0laj-m0p9",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "def load_tf1(path, input):\n",
        "  print('Loading from', path)\n",
        "  with tf.Graph().as_default() as g:\n",
        "    with tf1.Session() as sess:\n",
        "      meta_graph = tf1.saved_model.load(sess, [\"serve\"], path)\n",
        "      sig_def = meta_graph.signature_def[tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY]\n",
        "      input_name = sig_def.inputs['input'].name\n",
        "      output_name = sig_def.outputs['output'].name\n",
        "      print('  Output with input', input, ': ', \n",
        "            sess.run(output_name, feed_dict={input_name: input}))\n",
        "\n",
        "load_tf1('saved-model-builder', 5.)\n",
        "load_tf1('simple-save', 5.)\n",
        "load_tf1('estimator-model', [5.])  # Estimator's input must be batched.\n",
        "load_tf1('tf2-save', 5.)\n",
        "load_tf1('keras-model', 5.)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "FbR3sfvooVBN"
      },
      "source": [
        "### TensorFlow 2：加载使用 tf.saved_model 保存的模型\n",
        "\n",
        "在 TensorFlow 2 中，对象会加载到存储变量和函数的 Python 对象中。这与从 TensorFlow 1 保存的模型兼容。\n",
        "\n",
        "查看 `tf.saved_model.load` API 文档和[使用 SavedModel 格式](../..guide/saved_model)指南中的[加载和使用自定义模型](../../guide/saved_model#loading_and_using_a_custom_model)部分来了解详情。"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "OA52ezWV_KgL",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "def load_tf2(path, input):\n",
        "  print('Loading from', path)\n",
        "  loaded = tf.saved_model.load(path)\n",
        "  out = loaded.signatures[tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY](\n",
        "      tf.constant(input))['output']\n",
        "  print('  Output with input', input, ': ', out)\n",
        "\n",
        "load_tf2('saved-model-builder', 5.)\n",
        "load_tf2('simple-save', 5.)\n",
        "load_tf2('estimator-model', [5.])  # Estimator's input must be batched.\n",
        "load_tf2('tf2-save', 5.)\n",
        "load_tf2('keras-model', 5.)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Gz3VFn5aAfmK"
      },
      "source": [
        "使用 TensorFlow 2 API 保存的模型还可以访问附加到模型的 `tf.function` 和变量（而不是作为签名导出的那些条目）。例如："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "IfMTp-TGAfOs",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "loaded = tf.saved_model.load('tf2-save')\n",
        "print('restored __call__:', loaded.__call__)\n",
        "print('output with input 5.', loaded(5))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "VMoErNKHoXEg"
      },
      "source": [
        "### TensorFlow 2：加载使用 Keras 保存的模型\n",
        "\n",
        "已弃用：对于 Keras 对象，建议使用新的高级 `.keras` 格式和 `tf.keras.Model.export`，如[此处](https://tensorflow.google.cn/guide/keras/save_and_serialize)的指南所示。对于现有代码，继续支持低级 SavedModel 格式。\n",
        "\n",
        "Keras 加载 API (`tf.keras.models.load_model`) 允许您将保存的模型重新加载回 Keras 模型对象。请注意，这只允许您加载使用 Keras 保存的 SavedModel（`Model.save` 或 `tf.keras.models.save_model`）\n",
        "\n",
        "使用 `tf.saved_model.save` 保存的模型应使用 `tf.saved_model.load` 加载。可以使用 `tf.saved_model.load` 加载使用  `Model.save` 保存的 Keras 模型，但这样做只会获得 TensorFlow 计算图。有关详情，请参阅 `tf.keras.models.load_model` API 文档与[保存和加载 Keras 模型](https://tensorflow.google.cn/guide/keras/save_and_serialize#savedmodel_format)指南。"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "ZFUAxK0YeIAe",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "loaded_model = tf.keras.models.load_model('keras-model')\n",
        "loaded_model.predict_on_batch(tf.constant([1, 3, 4]))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Tz4eAAGY19MM"
      },
      "source": [
        "## GraphDef 和 MetaGraphDef\n",
        "\n",
        "&lt;a name=\"graphdef_and_metagraphdef\"&gt;\n",
        "\n",
        "没有直接方式可以将原始 `GraphDef` 或 `MetaGraphDef` 加载到 TF2。但是，可以使用 [`v1.wrap_function`](https://tensorflow.google.cn/api_docs/python/tf/compat/v1/wrap_function) 将导入计算图的 TF1 代码转换为 TF2 [`concrete_function`](https://tensorflow.org/guide/concrete_function)。\n",
        "\n",
        "首先，保存 MetaGraphDef："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "grKam9zGnNRZ",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "# Save a simple multiplication computation:\n",
        "with tf.Graph().as_default() as g:\n",
        "  x = tf1.placeholder(tf.float32, shape=[], name='x')\n",
        "  v = tf.Variable(3.0, name='v')\n",
        "  y = tf.multiply(x, v, name='y')\n",
        "  with tf1.Session() as sess:\n",
        "    sess.run(v.initializer)\n",
        "    print(sess.run(y, feed_dict={x: 5}))\n",
        "    s = tf1.train.Saver()\n",
        "    s.export_meta_graph('multiply.pb', as_text=True)\n",
        "    s.save(sess, 'multiply_values.ckpt')"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "bJnCe7eYrXev"
      },
      "source": [
        "利用 TF1 API，可以使用 `tf1.train.import_meta_graph` 导入计算图并恢复值："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "3bbcGZ4CoVDL",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "with tf.Graph().as_default() as g:\n",
        "  meta = tf1.train.import_meta_graph('multiply.pb')\n",
        "  x = g.get_tensor_by_name('x:0')\n",
        "  y = g.get_tensor_by_name('y:0')\n",
        "  with tf1.Session() as sess:\n",
        "    meta.restore(sess, 'multiply_values.ckpt')\n",
        "    print(sess.run(y, feed_dict={x: 5}))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "s9OnHOLDrnco"
      },
      "source": [
        "没有用于加载计算图的 TF2 API，但您仍然可以将其导入到可以在 Eager 模式下执行的具体函数中："
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "yAyGmDLlpVBX",
        "vscode": {
          "languageId": "python"
        }
      },
      "outputs": [],
      "source": [
        "def import_multiply():\n",
        "  # Any graph-building code is allowed here.\n",
        "  tf1.train.import_meta_graph('multiply.pb')\n",
        "\n",
        "# Creates a tf.function with all the imported elements in the function graph.\n",
        "wrapped_import = tf1.wrap_function(import_multiply, [])\n",
        "import_graph = wrapped_import.graph\n",
        "x = import_graph.get_tensor_by_name('x:0')\n",
        "y = import_graph.get_tensor_by_name('y:0')\n",
        "\n",
        "# Restore the variable values.\n",
        "tf1.train.Saver(wrapped_import.variables).restore(\n",
        "    sess=None, save_path='multiply_values.ckpt')\n",
        "\n",
        "# Create a concrete function by pruning the wrap_function (similar to sess.run).\n",
        "multiply_fn = wrapped_import.prune(feeds=x, fetches=y)\n",
        "\n",
        "# Run this function\n",
        "multiply_fn(tf.constant(5.))  # inputs to concrete functions must be Tensors."
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "GZ5vGJ0IDorc"
      },
      "source": [
        "## 从 TensorFlow 1 到 TensorFlow 2 的变化\n",
        "\n",
        "&lt;a id=\"changes_from_tf1_to_tf2\"&gt;\n",
        "\n",
        "本部分列出了 TensorFlow 1 中的关键保存和加载术语、它们的 TensorFlow 2 等效项以及发生了哪些变化。\n",
        "\n",
        "### SavedModel\n",
        "\n",
        "[SavedModel](../../guide/saved_model.ipynb) 是一种存储带参数和计算的完整 TensorFlow 程序的格式。它包含应用平台用来运行模型的签名。\n",
        "\n",
        "文件格式本身没有重大变化，因此可以使用 TensorFlow 1 或 TensorFlow 2 API 加载和应用 SavedModel。\n",
        "\n",
        "**TensorFlow 1 和 TensorFlow 2 的区别**\n",
        "\n",
        "除了 API 变更外，TensorFlow 2 中的*应用*和*推断*用例没有更新。在*重用*和*组合*从 SavedModel 加载的模型的能力中引入了改进。\n",
        "\n",
        "在 TensorFlow 2 中，程序由 `tf.Variable`、`tf.Module` 或更高级别的 Keras 模型 (`tf.keras.Model`) 和层 (`tf.keras.layers`) 等对象表示。不再具有在会话中存储值的全局变量，并且计算图现在存在于不同的 `tf.function` 中。因此，在模型导出期间，SavedModel 会分别保存每个组件和函数计算图。\n",
        "\n",
        "使用 TensorFlow Python API 编写 TensorFlow 程序时，您必须构建一个对象来管理变量、函数和其他资源。通常，可以通过使用 Keras API 来完成此目的，但也可以通过创建或子类化 `tf.Module` 来构建对象。\n",
        "\n",
        "Keras 模型 (`tf.keras.Model`) 和 `tf.Module` 会自动跟踪附加到它们的变量和函数。SavedModel 会保存各个模块、变量和函数之间的相应连接，以便在加载时可以恢复。\n",
        "\n",
        "### 签名\n",
        "\n",
        "签名是 SavedModel 的端点 – 它们告诉用户如何运行模型以及需要哪些输入。\n",
        "\n",
        "在 TensorFlow 1 中，签名是通过列出输入和输出张量来创建的。在 TensorFlow 2 中，签名是通过传入*具体函数*来生成的。（在[计算图和 tf.function 简介](../intro_to_graphs.ipynb)指南中阅读更多关于 TensorFlow 函数的信息，特别是*多态性：一个函数，多个计算图*部分。）简而言之，具体函数是从 `tf.function` 生成的：\n",
        "\n",
        "```python\n",
        "# Option 1: Specify an input signature.\n",
        "@tf.function(input_signature=[...])\n",
        "def fn(...):\n",
        "  ...\n",
        "  return outputs\n",
        "\n",
        "tf.saved_model.save(model, path, signatures={\n",
        "    'name': fn\n",
        "})\n",
        "```\n",
        "\n",
        "```python\n",
        "# Option 2: Call `get_concrete_function`\n",
        "@tf.function\n",
        "def fn(...):\n",
        "  ...\n",
        "  return outputs\n",
        "\n",
        "tf.saved_model.save(model, path, signatures={\n",
        "    'name': fn.get_concrete_function(...)\n",
        "})\n",
        "```\n",
        "\n",
        "### `Session.run`\n",
        "\n",
        "在 TensorFlow 1 中，只要您已经知道张量名称，就可以使用导入的计算图调用 `Session.run`。这样就可以检索恢复的变量值，或者运行未在签名中导出的模型部分。\n",
        "\n",
        "在 TensorFlow 2 中，可以直接访问变量，例如权重矩阵 (`kernel`)：\n",
        "\n",
        "```python\n",
        "model = tf.Module()\n",
        "model.dense_layer = tf.keras.layers.Dense(...)\n",
        "tf.saved_model.save('my_saved_model')\n",
        "loaded = tf.saved_model.load('my_saved_model')\n",
        "loaded.dense_layer.kernel\n",
        "```\n",
        "\n",
        "或者调用附加到模型对象的 `tf.function`：例如 `loaded.__call__`。\n",
        "\n",
        "与 TF1 不同，没有办法提取函数的一部分并访问中间值。您*必须*在保存的对象中导出所有需要的功能。\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "b6NG9JvUwJxn"
      },
      "source": [
        "## TensorFlow Serving 迁移指南\n",
        "\n",
        "SavedModel 最初是为了与 [TensorFlow Serving](https://tensorflow.google.cn/tfx/guide/serving) 一起使用而创建的。此平台提供不同类型的预测请求：分类、回归和预测。\n",
        "\n",
        "**TensorFlow 1** API 允许您使用效用函数创建这些类型的签名：\n",
        "\n",
        "- `tf.compat.v1.saved_model.classification_signature_def`\n",
        "- `tf.compat.v1.saved_model.regression_signature_def`\n",
        "- `tf.compat.v1.saved_model.predict_signature_def`\n",
        "\n",
        "[分类](https://tensorflow.google.cn/tfx/serving/signature_defs#classification_signaturedef) (`classification_signature_def`) 和[回归](https://tensorflow.google.cn/tfx/serving/signature_defs#regression_signaturedef) (`regression_signature_def`) 会限制输入和输出，因此输入必须是 `tf.Example`，输出必须是 `classes`、`scores` 或 `prediction`。同时，[预测签名](https://tensorflow.google.cn/tfx/serving/signature_defs#predict_signaturedef) (`predict_signature_def`) 没有限制。\n",
        "\n",
        "使用 **TensorFlow 2** API 导出的 SavedModel 与 TensorFlow Serving 兼容，但仅包含预测签名。分类和回归签名已被移除。\n",
        "\n",
        "如果您需要使用分类和回归签名，则可以使用 `tf.compat.v1.saved_model.signature_def_utils.MethodNameUpdater` 修改导出的 SavedModel。"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "a3acd3b86215"
      },
      "source": [
        "## 后续步骤\n",
        "\n",
        "要详细了解 TensorFlow 2 中的 SavedModel，请查看以下指南：\n",
        "\n",
        "- [使用 SavedModel 格式](https://tensorflow.google.cn/guide/saved_model)\n",
        "- [保存和加载 Keras 模型](https://tensorflow.google.cn/guide/keras/save_and_serialize)\n",
        "\n",
        "如果您使用的是 TensorFlow Hub，则可能会发现下列指南十分有用：\n",
        "\n",
        "- [TensorFlow Hub：TensorFlow 1/TensorFlow 2 的模型兼容性](https://tensorflow.google.cn/hub/model_compatibility)\n",
        "- [使用 TensorFlow Hub 从 TensorFlow 1 迁移到 TensorFlow 2](https://tensorflow.google.cn/hub/migration_tf2)"
      ]
    }
  ],
  "metadata": {
    "colab": {
      "collapsed_sections": [],
      "name": "saved_model.ipynb",
      "toc_visible": true
    },
    "kernelspec": {
      "display_name": "Python 3",
      "name": "python3"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}