更新至 2018-9-14 版本
模型进度可以在训练期间和训练之后保存。这意味着模型可以从中断的地方恢复,并避免长时间的训练。保存模型也意味着你可以共享你的模型,而其他人可以重新创建你的工作。在发布研究模型和技术时,大多数机器学习从业者共享:
- 用于创建模型的代码
- 模型的训练权重或参数
共享此数据有助于其他人了解模型的工作原理,并使用新数据自行尝试。
保存 TensorFlow 模型有多种方法 —— 取决于你使用的API。本指南使用 tf.keras
,一个用于在 TensorFlow 中构建和训练模型的高级 API。有关其他方法,请参阅TensorFlow 保存和还原 指南或 保存在 eager 中。
设置
安装和导入
安装并导入 TensorFlow 和依赖项:
!pip install -q h5py pyyaml
获取样本数据集
我们将使用 MNIST 数据集 来训练我们的模型,并演示如何保存权重。我们仅使用前 1000 个样本:
from __future__ import absolute_import, division, print_function
import os
import tensorflow as tf
from tensorflow import keras
tf.__version__
'1.11.0-rc0'
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.mnist.load_data()
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
Downloading data from https://s3.amazonaws.com/img-datasets/mnist.npz
11493376/11490434 [==============================] - 2s 0us/step
定义模型
让我们构建一个简单的模型,并用它来演示保存和加载权重。
# 返回一个简单的序列模型
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)
])
model.compile(optimizer=tf.keras.optimizers.Adam(),
loss=tf.keras.losses.sparse_categorical_crossentropy,
metrics=['accuracy'])
return model
# 创建一个基础模型实例
model = create_model()
model.summary()
_________________________________________________________________
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
_________________________________________________________________
训练时保存检查点
在训练期间和训练结束时自动保存检查点。通过这种方式,你可以使用已训练的模型,而无需重新训练,或者在你中断的地方继续训练 —— 万一训练过程中断。
tf.keras.callbacks.ModelCheckpoint 是执行此任务的回调。回调需要几个参数来配置检查点。
检查点回调用法
训练模型并将 ModelCheckpoint
回调传递给它:
checkpoint_path = "training_1/cp.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)
# 创建检查点回调
cp_callback = tf.keras.callbacks.ModelCheckpoint(checkpoint_path,
save_weights_only=True,
verbose=1)
model = create_model()
model.fit(train_images, train_labels, epochs = 10,
validation_data = (test_images,test_labels),
callbacks = [cp_callback]) # 训练时传递回调
Train on 1000 samples, validate on 1000 samples
Epoch 1/10
800/1000 [=======================>......] - ETA: 0s - loss: 1.2836 - acc: 0.6400
Epoch 00001: saving model to training_1/cp.ckpt
WARNING:tensorflow:This model was compiled with a Keras optimizer (<tensorflow.python.keras.optimizers.Adam object at 0x7f8d67c20198>) 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.
1000/1000 [==============================] - 0s 195us/step - loss: 0.0368 - acc: 1.0000 - val_loss: 0.4113 - val_acc: 0.8660
这将创建一个 TensorFlow 检查点文件集合,这些文件在每个周期结束时更新:
!ls {checkpoint_dir}
checkpoint cp.ckpt.data-00000-of-00001 cp.ckpt.index
创建一个新的未经训练的模型时,如果仅从权重恢复模型,则必须具有与原始模型相同的架构。由于它是相同的模型架构,我们可以共享权重,尽管它是模型的不同实例。
现在重建一个新的未经训练的模型,并在测试集上进行评估(准确度约为10%):
model = create_model()
loss, acc = model.evaluate(test_images, test_labels)
print("Untrained model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 121us/step
Untrained model, accuracy: 12.50%
然后从检查点加载权重,并重新评估:
model.load_weights(checkpoint_path)
loss,acc = model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 36us/step
Restored model, accuracy: 86.60%
检查点回调选项
回调提供了几个选项,可以为生成的检查点提供唯一的名称,并调整检查点的频率。
训练一个新模型,每 5 个周期保存一次唯一命名的检查点:
# 在文件名中包含周期。 (使用 `str.format`)
checkpoint_path = "training_2/cp-{epoch:04d}.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)
cp_callback = tf.keras.callbacks.ModelCheckpoint(
checkpoint_path, verbose=1, save_weights_only=True,
# 每 5 个周期保存权重
period=5)
model = create_model()
model.fit(train_images, train_labels,
epochs = 50, callbacks = [cp_callback],
validation_data = (test_images,test_labels),
verbose=0)
Epoch 00005: saving model to training_2/cp-0005.ckpt
WARNING:tensorflow:This model was compiled with a Keras optimizer (<tensorflow.python.keras.optimizers.Adam object at 0x7f8dc5d86198>) 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.
...
Epoch 00050: saving model to training_2/cp-0050.ckpt
WARNING:tensorflow:This model was compiled with a Keras optimizer (<tensorflow.python.keras.optimizers.Adam object at 0x7f8dc5d86198>) 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.
现在,查看生成的检查点(按修改日期排序):
import pathlib
# 按修改日期排序检查点
checkpoints = pathlib.Path(checkpoint_dir).glob("*.index")
checkpoints = sorted(checkpoints, key=lambda cp:cp.stat().st_mtime)
checkpoints = [cp.with_suffix('') for cp in checkpoints]
latest = str(checkpoints[-1])
checkpoints
[PosixPath('training_2/cp-0005.ckpt'),
PosixPath('training_2/cp-0010.ckpt'),
PosixPath('training_2/cp-0015.ckpt'),
PosixPath('training_2/cp-0020.ckpt'),
PosixPath('training_2/cp-0025.ckpt'),
PosixPath('training_2/cp-0030.ckpt'),
PosixPath('training_2/cp-0035.ckpt'),
PosixPath('training_2/cp-0040.ckpt'),
PosixPath('training_2/cp-0045.ckpt'),
PosixPath('training_2/cp-0050.ckpt')]
注意:默认的 tensorflow 格式仅保存最近的5个检查点。
要进行测试,需要重置模型并加载最新的检查点:
model = create_model()
model.load_weights(latest)
loss, acc = model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 94us/step
Restored model, accuracy: 87.60%
这些文件是什么?
上述代码将权重存储到 检查点 格式文件的集合中,这些文件仅包含二进制格式的训练权重。
检查点包含:
- 一个或多个包含模型权重的分片。
- 索引文件,指示哪些权重存储在哪个分片中。
如果你只在一台机器上训练模型,那么你将有一个带有 .data-00000-of-00001
后缀的分片。
手动保存权重
上文讲述了如何将权重加载到模型中。
手动保存权重同样很简单,需要使用 Model.save_weights
方法。
# 保存权重
model.save_weights('./checkpoints/my_checkpoint')
# 恢复权重
model = create_model()
model.load_weights('./checkpoints/my_checkpoint')
loss,acc = model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
WARNING:tensorflow:This model was compiled with a Keras optimizer (<tensorflow.python.keras.optimizers.Adam object at 0x7f8dc5155748>) 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.
1000/1000 [==============================] - 0s 100us/step
Restored model, accuracy: 87.60%
保存完整模型
可以保存整个模型到文件,包含权重值、模型配置甚至优化器配置。这允许你检查模型并稍后从完全相同的状态恢复培训,而无需访问原始代码。
在 Keras 中保存功能齐全的模型非常有用,你可以在 TensorFlow.js 中加载它们,然后在 Web 浏览器中训练和运行它们。
Keras 使用 HDF5 标准提供基本保存格式。对于我们的模型,可以将其视为单个二进制 blob。
model = create_model()
model.fit(train_images, train_labels, epochs=5)
# 将整个模型保存至 HDF5 文件
model.save('my_model.h5')
Epoch 1/5
1000/1000 [==============================] - 0s 419us/step - loss: 1.1345 - acc: 0.6800
Epoch 2/5
1000/1000 [==============================] - 0s 162us/step - loss: 0.4104 - acc: 0.8880
Epoch 3/5
1000/1000 [==============================] - 0s 159us/step - loss: 0.2768 - acc: 0.9270
Epoch 4/5
1000/1000 [==============================] - 0s 162us/step - loss: 0.2194 - acc: 0.9530
Epoch 5/5
1000/1000 [==============================] - 0s 157us/step - loss: 0.1678 - acc: 0.9600
现在从该文件重新创建模型:
# 重新创建完全相同的模型,包括权重和优化器
new_model = keras.models.load_model('my_model.h5')
new_model.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_12 (Dense) (None, 512) 401920
_________________________________________________________________
dropout_6 (Dropout) (None, 512) 0
_________________________________________________________________
dense_13 (Dense) (None, 10) 5130
=================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
_________________________________________________________________
检查其准确率:
loss, acc = new_model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 120us/step
Restored model, accuracy: 84.70%
这项技术可以保存所有东西:
- 权重值
- 模型的配置(架构)
- 优化器配置
Keras 通过检查架构来保存模型。目前,它无法保存 TensorFlow 优化器(来自 tf.train
)。使用这些时,你需要在加载后重新编译模型,并且你将失去优化器的状态。
下一步可以做什么?
本教程是使用 tf.keras
保存和加载模型的快速指南。
- tf.keras 指南 包含更多有关保存和加载模型的内容。
- 参阅 在 eager 中保存。
- 保存和恢复 指南包含有关 TensorFlow 保存的低层详细信息。