我使用ModelCheckpoint回调函数来保存我正在训练的模型的最佳epoch。保存时没有错误，但当我尝试加载时，出现了以下错误：

2019-07-27 22:58:04.713951: W tensorflow/core/util/tensor_slice_reader.cc:95] Could not open C:\Users\Riley\PycharmProjects\myNN\cp.ckpt: Data loss: not an sstable (bad magic number): perhaps your file is in a different file format and you need to use a different restore operator?

我尝试使用绝对路径/完整路径，但没有成功。我确定可以使用EarlyStopping，但我想了解为什么会出现这个错误。以下是我的代码：

from __future__ import absolute_import, division, print_functionimport tensorflow as tffrom tensorflow import kerasimport numpy as npimport matplotlib.pyplot as pltimport datetimeimport statistics(train_images, train_labels), (test_images, test_labels) = np.load("dataset.npy", allow_pickle=True)train_images = train_images / 255test_images = test_images / 255train_labels = list(map(float, train_labels))test_labels = list(map(float, test_labels))train_labels = [i/10 for i in train_labels]test_labels = [i/10 for i in test_labels]'''model = keras.Sequential([    keras.layers.Flatten(input_shape=(128, 128)),    keras.layers.Dense(64, activation=tf.nn.relu),    keras.layers.Dense(1)  ])'''start_time = datetime.datetime.now()model = keras.Sequential([    keras.layers.Conv2D(32, kernel_size=(5, 5), strides=(1, 1), activation='relu', input_shape=(128, 128, 1)),    keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2)),    keras.layers.Dropout(0.2),    keras.layers.Conv2D(64, (5, 5), activation='relu'),    keras.layers.MaxPooling2D(pool_size=(2, 2)),    keras.layers.Dropout(0.2),    keras.layers.Flatten(),    keras.layers.Dropout(0.5),    keras.layers.Dense(1000, activation='relu'),    keras.layers.Dense(1)])model.compile(loss='mean_absolute_error',    optimizer=keras.optimizers.SGD(lr=0.01),    metrics=['mean_absolute_error', 'mean_squared_error'])train_images = train_images.reshape(328, 128, 128, 1)test_images = test_images.reshape(82, 128, 128, 1)model.fit(train_images, train_labels, epochs=100, callbacks=[keras.callbacks.ModelCheckpoint("cp.ckpt", monitor='mean_absolute_error', save_best_only=True, verbose=1)])model.load_weights("cp.ckpt")predictions = model.predict(test_images)totalDifference = 0for i in range(82):    print("%s: %s" % (test_labels[i] * 10, predictions[i] * 10))    totalDifference += abs(test_labels[i] - predictions[i])avgDifference = totalDifference / 8.2print("\n%s\n" % avgDifference)print("Time Elapsed:")print(datetime.datetime.now() - start_time)

回答：

简而言之，你在保存整个模型，但尝试加载的只是权重，这样是不行的。

解释

你的模型的fit方法如下：

model.fit(    train_images,    train_labels,    epochs=100,    callbacks=[        keras.callbacks.ModelCheckpoint(            "cp.ckpt", monitor="mean_absolute_error", save_best_only=True, verbose=1        )    ],)

由于ModelCheckpoint中的save_weights=False是默认设置，你正在将整个模型保存到.ckpt文件中。

顺便说一下，文件应该命名为.hdf5或.hf5，因为这是Hierarchical Data Format 5。由于Windows不是对扩展名无感知的操作系统，如果tensorflow/keras依赖于扩展名，你可能会遇到一些问题。

另一方面，你只是在加载模型的权重，而文件中包含了整个模型：

model.load_weights("cp.ckpt")

Tensorflow的检查点机制（.cp）与Keras的（.hdf5）不同，所以要注意这一点（有计划将它们更紧密地整合在一起，参见这里和这里）。

解决方案

因此，要么继续使用当前的回调函数，但是使用model.load("model.hdf5")，要么在ModelCheckpoint中添加save_weights_only=True参数：

model.fit(    train_images,    train_labels,    epochs=100,    callbacks=[        keras.callbacks.ModelCheckpoint(            "weights.hdf5",            monitor="mean_absolute_error",            save_best_only=True,            verbose=1,            save_weights_only=True,  # 指定此参数        )    ],)

然后你可以使用你的model.load_weights("weights.hdf5")。