AI - TensorFlow - 示例05：保存和恢复模型

保存和恢复模型（Save and restore models）

官网示例：https://www.tensorflow.org/tutorials/keras/save_and_restore_models

在训练期间保存检查点

在训练期间或训练结束时自动保存检查点。
权重存储在检查点格式的文件集合中，这些文件仅包含经过训练的权重（采用二进制格式）。
可以使用经过训练的模型，而无需重新训练该模型，或从上次暂停的地方继续训练，以防训练过程中断

• 检查点回调用法：创建检查点回调，训练模型并将ModelCheckpoint回调传递给该模型，得到检查点文件集合，用于分享权重
• 检查点回调选项：该回调提供了多个选项，用于为生成的检查点提供独一无二的名称，以及调整检查点创建频率。

手动保存权重

使用 Model.save_weights 方法即可手动保存权重

保存整个模型

整个模型可以保存到一个文件中，其中包含权重值、模型配置（架构）、优化器配置。
可以为模型设置检查点，并稍后从完全相同的状态继续训练，而无需访问原始代码。
Keras通过检查架构来保存模型，使用HDF5标准提供基本的保存格式。
特别注意：

• 目前无法保存TensorFlow优化器（来自tf.train）。
• 使用此类优化器时，需要在加载模型后对其进行重新编译，使优化器的状态变松散。

MNIST数据集

MNIST（Mixed National Institute of Standards and Technology database）是一个计算机视觉数据集

• 官方下载地址：http://yann.lecun.com/exdb/mnist/
• 包含70000张手写数字的灰度图片，其中60000张为训练图像和10000张为测试图像
• 每一张图片都是28*28个像素点大小的灰度图像
• https://keras.io/datasets/#mnist-database-of-handwritten-digits
• TensorFlow：https://www.tensorflow.org/api_docs/python/tf/keras/datasets/mnist

示例

脚本内容

GitHub：https://github.com/anliven/Hello-AI/blob/master/Google-Learn-and-use-ML/5_save_and_restore_models.py

 # coding=utf-8
 import tensorflow as tf
 from tensorflow import keras
 import numpy as np
 import pathlib
 import os

 os.environ['TF_CPP_MIN_LOG_LEVEL'] = ''
 print("# TensorFlow version: {}  - tf.keras version: {}".format(tf.VERSION, tf.keras.__version__))  # 查看版本

 # ### 获取示例数据集

 ds_path = str(pathlib.Path.cwd()) + "\\datasets\\mnist\\"  # 数据集路径
 np_data = np.load(ds_path + "mnist.npz")  # 加载numpy格式数据
 print("# np_data keys: ", list(np_data.keys()))  # 查看所有的键

 # 加载mnist数据集
 (train_images, train_labels), (test_images, test_labels) = keras.datasets.mnist.load_data(path=ds_path + "mnist.npz")
 train_labels = train_labels[:1000]
 test_labels = test_labels[:1000]
 train_images = train_images[:1000].reshape(-1, 28 * 28) / 255.0
 test_images = test_images[:1000].reshape(-1, 28 * 28) / 255.0

 # ### 定义模型
 def create_model():
     model = tf.keras.models.Sequential([
         keras.layers.Dense(512, activation=tf.nn.relu, input_shape=(784,)),
         keras.layers.Dropout(0.2),
         keras.layers.Dense(10, activation=tf.nn.softmax)
     ])  # 构建一个简单的模型
     model.compile(optimizer=tf.keras.optimizers.Adam(),
                   loss=tf.keras.losses.sparse_categorical_crossentropy,
                   metrics=['accuracy'])
     return model

 mod = create_model()
 mod.summary()

 # ### 在训练期间保存检查点

 # 检查点回调用法
 checkpoint_path = "training_1/cp.ckpt"
 checkpoint_dir = os.path.dirname(checkpoint_path)  # 检查点存放目录
 cp_callback = tf.keras.callbacks.ModelCheckpoint(checkpoint_path,
                                                  save_weights_only=True,
                                                  verbose=2)  # 创建检查点回调
 model1 = create_model()
 model1.fit(train_images, train_labels,
            epochs=10,
            validation_data=(test_images, test_labels),
            verbose=0,
            callbacks=[cp_callback]  # 将ModelCheckpoint回调传递给该模型
            )  # 训练模型，将创建一个TensorFlow检查点文件集合，这些文件在每个周期结束时更新

 model2 = create_model()  # 创建一个未经训练的全新模型（与原始模型架构相同，才能分享权重）
 loss, acc = model2.evaluate(test_images, test_labels)  # 使用测试集进行评估
 print("# Untrained model2, accuracy: {:5.2f}%".format(100 * acc))  # 未训练模型的表现（准确率约为10%）

 model2.load_weights(checkpoint_path)  # 从检查点加载权重
 loss, acc = model2.evaluate(test_images, test_labels)  # 使用测试集，重新进行评估
 print("# Restored model2, accuracy: {:5.2f}%".format(100 * acc))  # 模型表现得到大幅提升

 # 检查点回调选项
 checkpoint_path2 = "training_2/cp-{epoch:04d}.ckpt"  # 使用“str.format”方式为每个检查点设置唯一名称
 checkpoint_dir2 = os.path.dirname(checkpoint_path)
 cp_callback2 = tf.keras.callbacks.ModelCheckpoint(checkpoint_path2,
                                                   verbose=1,
                                                   save_weights_only=True,
                                                   period=5  # 每隔5个周期保存一次检查点
                                                   )  # 创建检查点回调
 model3 = create_model()
 model3.fit(train_images, train_labels,
            epochs=50,
            callbacks=[cp_callback2],  # 将ModelCheckpoint回调传递给该模型
            validation_data=(test_images, test_labels),
            verbose=0)  # 训练一个新模型，每隔5个周期保存一次检查点并设置唯一名称
 latest = tf.train.latest_checkpoint(checkpoint_dir2)
 print("# latest checkpoint: {}".format(latest))  # 查看最新的检查点

 model4 = create_model()  # 重新创建一个全新的模型
 loss, acc = model2.evaluate(test_images, test_labels)  # 使用测试集进行评估
 print("# Untrained model4, accuracy: {:5.2f}%".format(100 * acc))  # 未训练模型的表现（准确率约为10%）

 model4.load_weights(latest)  # 加载最新的检查点
 loss, acc = model4.evaluate(test_images, test_labels)  #
 print("# Restored model4, accuracy: {:5.2f}%".format(100 * acc))  # 模型表现得到大幅提升

 # ### 手动保存权重
 model5 = create_model()
 model5.fit(train_images, train_labels,
            epochs=10,
            validation_data=(test_images, test_labels),
            verbose=0)  # 训练模型
 model5.save_weights('./training_3/my_checkpoint')  # 手动保存权重

 model6 = create_model()
 loss, acc = model6.evaluate(test_images, test_labels)
 print("# Restored model6, accuracy: {:5.2f}%".format(100 * acc))
 model6.load_weights('./training_3/my_checkpoint')
 loss, acc = model6.evaluate(test_images, test_labels)
 print("# Restored model6, accuracy: {:5.2f}%".format(100 * acc))

 # ### 保存整个模型
 model7 = create_model()
 model7.fit(train_images, train_labels, epochs=5)
 model7.save('my_model.h5')  # 保存整个模型到HDF5文件

 model8 = keras.models.load_model('my_model.h5')  # 重建完全一样的模型，包括权重和优化器
 model8.summary()
 loss, acc = model8.evaluate(test_images, test_labels)
 print("Restored model8, accuracy: {:5.2f}%".format(100 * acc))

运行结果

C:\Users\anliven\AppData\Local\conda\conda\envs\mlcc\python.exe D:/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML/5_save_and_restore_models.py
# TensorFlow version: 1.12.0  - tf.keras version: 2.1.6-tf
# np_data keys:  ['x_test', 'x_train', 'y_train', 'y_test']
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
dense (Dense)                (None, 512)               401920
_________________________________________________________________
dropout (Dropout)            (None, 512)               0
_________________________________________________________________
dense_1 (Dense)              (None, 10)                5130
=================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
_________________________________________________________________

Epoch 00001: saving model to training_1/cp.ckpt
Epoch 00002: saving model to training_1/cp.ckpt
Epoch 00003: saving model to training_1/cp.ckpt
Epoch 00004: saving model to training_1/cp.ckpt
Epoch 00005: saving model to training_1/cp.ckpt
Epoch 00006: saving model to training_1/cp.ckpt
Epoch 00007: saving model to training_1/cp.ckpt
Epoch 00008: saving model to training_1/cp.ckpt
Epoch 00009: saving model to training_1/cp.ckpt
Epoch 00010: saving model to training_1/cp.ckpt

  32/1000 [..............................] - ETA: 3s
1000/1000 [==============================] - 0s 140us/step
# Untrained model2, accuracy:  8.20%

  32/1000 [..............................] - ETA: 0s
1000/1000 [==============================] - 0s 40us/step
# Restored model2, accuracy: 86.40%

Epoch 00005: saving model to training_2/cp-0005.ckpt
Epoch 00010: saving model to training_2/cp-0010.ckpt
Epoch 00015: saving model to training_2/cp-0015.ckpt
Epoch 00020: saving model to training_2/cp-0020.ckpt
Epoch 00025: saving model to training_2/cp-0025.ckpt
Epoch 00030: saving model to training_2/cp-0030.ckpt
Epoch 00035: saving model to training_2/cp-0035.ckpt
Epoch 00040: saving model to training_2/cp-0040.ckpt
Epoch 00045: saving model to training_2/cp-0045.ckpt
Epoch 00050: saving model to training_2/cp-0050.ckpt

# latest checkpoint: training_1\cp.ckpt

  32/1000 [..............................] - ETA: 3s
1000/1000 [==============================] - 0s 140us/step
# Untrained model4, accuracy: 86.40%

  32/1000 [..............................] - ETA: 2s
1000/1000 [==============================] - 0s 110us/step
# Restored model4, accuracy: 86.40%

  32/1000 [..............................] - ETA: 5s
1000/1000 [==============================] - 0s 220us/step
# Restored model6, accuracy: 18.20%

  32/1000 [..............................] - ETA: 0s
1000/1000 [==============================] - 0s 40us/step
# Restored model6, accuracy: 87.40%
Epoch 1/5

  32/1000 [..............................] - ETA: 9s - loss: 2.4141 - acc: 0.0625
 320/1000 [========>.....................] - ETA: 0s - loss: 1.8229 - acc: 0.4469
 576/1000 [================>.............] - ETA: 0s - loss: 1.4932 - acc: 0.5694
 864/1000 [========================>.....] - ETA: 0s - loss: 1.2624 - acc: 0.6481
1000/1000 [==============================] - 1s 530us/step - loss: 1.1978 - acc: 0.6620
Epoch 2/5

  32/1000 [..............................] - ETA: 0s - loss: 0.5490 - acc: 0.8750
 320/1000 [========>.....................] - ETA: 0s - loss: 0.4832 - acc: 0.8594
 576/1000 [================>.............] - ETA: 0s - loss: 0.4630 - acc: 0.8715
 864/1000 [========================>.....] - ETA: 0s - loss: 0.4356 - acc: 0.8808
1000/1000 [==============================] - 0s 200us/step - loss: 0.4298 - acc: 0.8790
Epoch 3/5

  32/1000 [..............................] - ETA: 0s - loss: 0.1681 - acc: 0.9688
 320/1000 [========>.....................] - ETA: 0s - loss: 0.2826 - acc: 0.9437
 576/1000 [================>.............] - ETA: 0s - loss: 0.2774 - acc: 0.9340
 832/1000 [=======================>......] - ETA: 0s - loss: 0.2740 - acc: 0.9327
1000/1000 [==============================] - 0s 200us/step - loss: 0.2781 - acc: 0.9280
Epoch 4/5

  32/1000 [..............................] - ETA: 0s - loss: 0.1589 - acc: 0.9688
 288/1000 [=======>......................] - ETA: 0s - loss: 0.2169 - acc: 0.9410
 608/1000 [=================>............] - ETA: 0s - loss: 0.2186 - acc: 0.9457
 864/1000 [========================>.....] - ETA: 0s - loss: 0.2231 - acc: 0.9479
1000/1000 [==============================] - 0s 200us/step - loss: 0.2164 - acc: 0.9480
Epoch 5/5

  32/1000 [..............................] - ETA: 0s - loss: 0.1095 - acc: 1.0000
 352/1000 [=========>....................] - ETA: 0s - loss: 0.1631 - acc: 0.9744
 608/1000 [=================>............] - ETA: 0s - loss: 0.1671 - acc: 0.9638
 864/1000 [========================>.....] - ETA: 0s - loss: 0.1545 - acc: 0.9688
1000/1000 [==============================] - 0s 210us/step - loss: 0.1538 - acc: 0.9670
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
dense_14 (Dense)             (None, 512)               401920
_________________________________________________________________
dropout_7 (Dropout)          (None, 512)               0
_________________________________________________________________
dense_15 (Dense)             (None, 10)                5130
=================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
_________________________________________________________________

  32/1000 [..............................] - ETA: 3s
1000/1000 [==============================] - 0s 150us/step
Restored model8, accuracy: 86.10%

Process finished with exit code 0

生成的文件

anliven@ANLIVEN MINGW64 /d/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML
$ ll training_1
total 1601
-rw-r--r-- 1 anliven 197121      71 5月   5 23:36 checkpoint
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:36 cp.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:36 cp.ckpt.index

anliven@ANLIVEN MINGW64 /d/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML
$

anliven@ANLIVEN MINGW64 /d/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML
$

anliven@ANLIVEN MINGW64 /d/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML
$

anliven@ANLIVEN MINGW64 /d/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML
$ ls -l training_1
total 1601
-rw-r--r-- 1 anliven 197121      71 5月   5 23:36 checkpoint
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:36 cp.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:36 cp.ckpt.index

anliven@ANLIVEN MINGW64 /d/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML
$ ls -l training_2
total 16001
-rw-r--r-- 1 anliven 197121      81 5月   5 23:37 checkpoint
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:36 cp-0005.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:36 cp-0005.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:36 cp-0010.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:36 cp-0010.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:36 cp-0015.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:36 cp-0015.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:36 cp-0020.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:36 cp-0020.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:36 cp-0025.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:36 cp-0025.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:37 cp-0030.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:37 cp-0030.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:37 cp-0035.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:37 cp-0035.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:37 cp-0040.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:37 cp-0040.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:37 cp-0045.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:37 cp-0045.ckpt.index
-rw-r--r-- 1 anliven 197121 1631508 5月   5 23:37 cp-0050.ckpt.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:37 cp-0050.ckpt.index

anliven@ANLIVEN MINGW64 /d/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML
$ ls -l training_3
total 1601
-rw-r--r-- 1 anliven 197121      83 5月   5 23:37 checkpoint
-rw-r--r-- 1 anliven 197121 1631517 5月   5 23:37 my_checkpoint.data-00000-of-00001
-rw-r--r-- 1 anliven 197121     647 5月   5 23:37 my_checkpoint.index

anliven@ANLIVEN MINGW64 /d/Anliven/Anliven-Code/PycharmProjects/Google-Learn-and-use-ML
$ ls -l my_model.h5
-rw-r--r-- 1 anliven 197121 4909112 5月   5 23:37 my_model.h5

问题处理

问题描述：出现如下告警信息。

WARNING:tensorflow:This model was compiled with a Keras optimizer (<tensorflow.python.keras.optimizers.Adam object at 0x00000280FD318780>) but is being saved in TensorFlow format with `save_weights`. The model's weights will be saved, but unlike with TensorFlow optimizers in the TensorFlow format the optimizer's state will not be saved.

Consider using a TensorFlow optimizer from `tf.train`.

问题处理：

正常告警，对脚本运行和结果无影响，暂不关注。