使用 tf.train.Saver 中断和恢复 Tensorflow 的训练数据

在 TensorFlow 1.x 的图模式中，tf.train.Saver 可以把当前会话里的变量保存成 checkpoint，并在新的会话中恢复这些变量。这个机制适合处理训练被中断、需要分阶段训练、或者需要把训练好的模型参数保存下来再继续调试的场景。

下面的例子使用 MNIST 数据集训练一个两层全连接网络：第一次会话先训练 3 个 epoch 并保存模型；第二次会话重新初始化图变量后，再从 checkpoint 恢复权重并继续训练 7 个 epoch。

需要注意：这段代码使用的是 TensorFlow 1.x API，包括 tf.Session()、tf.placeholder() 和 tensorflow.examples.tutorials.mnist.input_data。如果在 TensorFlow 2.x 环境中运行，应使用兼容模式，例如禁用 eager execution 并通过 tf.compat.v1 调用对应接口；具体可用性请以本地安装版本为准。

'''
Save and Restore a model using TensorFlow.
This example is using the MNIST database of handwritten digits
(http://yann.lecun.com/exdb/mnist/)

Author: Aymeric Damien
Project: https://github.com/aymericdamien/TensorFlow-Examples/
'''

from __future__ import print_function

# Import MNIST data
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)

import tensorflow as tf

# Parameters
learning_rate = 0.001
batch_size = 100
display_step = 1
model_path = "/tmp/model.ckpt"

# Network Parameters
n_hidden_1 = 256 # 1st layer number of features
n_hidden_2 = 256 # 2nd layer number of features
n_input = 784 # MNIST data input (img shape: 28*28)
n_classes = 10 # MNIST total classes (0-9 digits)

# tf Graph input
x = tf.placeholder("float", [None, n_input])
y = tf.placeholder("float", [None, n_classes])

# Create model
def multilayer_perceptron(x, weights, biases):
    # Hidden layer with RELU activation
    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])
    layer_1 = tf.nn.relu(layer_1)
    # Hidden layer with RELU activation
    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])
    layer_2 = tf.nn.relu(layer_2)
    # Output layer with linear activation
    out_layer = tf.matmul(layer_2, weights['out']) + biases['out']
    return out_layer

# Store layers weight & bias
weights = {
    'h1': tf.Variable(tf.random_normal([n_input, n_hidden_1])),
    'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2])),
    'out': tf.Variable(tf.random_normal([n_hidden_2, n_classes]))
}
biases = {
    'b1': tf.Variable(tf.random_normal([n_hidden_1])),
    'b2': tf.Variable(tf.random_normal([n_hidden_2])),
    'out': tf.Variable(tf.random_normal([n_classes]))
}

# Construct model
pred = multilayer_perceptron(x, weights, biases)

# Define loss and optimizer
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

# Initializing the variables
init = tf.global_variables_initializer()

# 'Saver' op to save and restore all the variables
saver = tf.train.Saver()

# Running first session
print("Starting 1st session...")
with tf.Session() as sess:
    # Initialize variables
    sess.run(init)

    # Training cycle
    for epoch in range(3):
        avg_cost = 0.
        total_batch = int(mnist.train.num_examples/batch_size)
        # Loop over all batches
        for i in range(total_batch):
            batch_x, batch_y = mnist.train.next_batch(batch_size)
            # Run optimization op (backprop) and cost op (to get loss value)
            _, c = sess.run([optimizer, cost], feed_dict={x: batch_x,
                                                          y: batch_y})
            # Compute average loss
            avg_cost += c / total_batch
        # Display logs per epoch step
        if epoch % display_step == 0:
            print("Epoch:", '%04d' % (epoch+1), "cost=", \
                "{:.9f}".format(avg_cost))
    print("First Optimization Finished!")

    # Test model
    correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
    # Calculate accuracy
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    print("Accuracy:", accuracy.eval({x: mnist.test.images, y: mnist.test.labels}))

    # Save model weights to disk
    save_path = saver.save(sess, model_path)
    print("Model saved in file: %s" % save_path)

# Running a new session
print("Starting 2nd session...")
with tf.Session() as sess:
    # Initialize variables
    sess.run(init)

    # Restore model weights from previously saved model
    saver.restore(sess, model_path)
    print("Model restored from file: %s" % save_path)

    # Resume training
    for epoch in range(7):
        avg_cost = 0.
        total_batch = int(mnist.train.num_examples / batch_size)
        # Loop over all batches
        for i in range(total_batch):
            batch_x, batch_y = mnist.train.next_batch(batch_size)
            # Run optimization op (backprop) and cost op (to get loss value)
            _, c = sess.run([optimizer, cost], feed_dict={x: batch_x,
                                                          y: batch_y})
            # Compute average loss
            avg_cost += c / total_batch
        # Display logs per epoch step
        if epoch % display_step == 0:
            print("Epoch:", '%04d' % (epoch + 1), "cost=", \
                "{:.9f}".format(avg_cost))
    print("Second Optimization Finished!")

    # Test model
    correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
    # Calculate accuracy
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    print("Accuracy:", accuracy.eval(
        {x: mnist.test.images, y: mnist.test.labels}))

核心流程可以概括为三步：

在图构建完成后创建 saver = tf.train.Saver()。
在训练会话中调用 saver.save(sess, model_path) 保存变量。
在新的会话中先构建同样的图，再调用 saver.restore(sess, model_path) 恢复变量。

Saver 保存的是图中变量的值，而不是 Python 训练循环本身。因此，如果需要精确恢复到某一个训练步，还应额外保存当前 epoch、global step、学习率调度状态等训练元数据。常见做法是定义 global_step = tf.Variable(0, trainable=False)，并在优化器的 minimize() 中传入 global_step=global_step，这样 checkpoint 中也会包含训练步数。

运行后，/tmp/model.ckpt 附近通常会出现一组 checkpoint 文件。恢复时不需要手动读取这些文件，只要 model_path 指向保存时使用的路径即可。

Leave a Reply Cancel reply