如何调整TensorFlow API中检查点的保存间隔（1000步改500步）

1. Keras API（tf.keras）场景 #

在ModelCheckpoint回调中设置save_freq参数即可指定每500步保存一次：

• 示例代码：

from tensorflow.keras.callbacks import ModelCheckpoint

# 定义检查点回调
checkpoint_callback = ModelCheckpoint(
    filepath='./checkpoints/model_{epoch}_{step}',
    save_freq=500,  # 核心配置：每500步保存一次
    save_weights_only=False,
    verbose=1
)

# 训练时传入该回调
model.fit(
    x_train, y_train,
    epochs=10,
    callbacks=[checkpoint_callback]
)

• 注意：save_freq设为整数时按训练步数计数，设为'epoch'则按轮次保存，默认是轮次。

2. Estimator API场景 #

通过RunConfig的save_checkpoints_steps参数配置：

• 示例代码：

import tensorflow as tf

# 配置训练参数
run_config = tf.estimator.RunConfig(
    save_checkpoints_steps=500,  # 每500步保存检查点
    model_dir='./checkpoints'
)

# 初始化Estimator
estimator = tf.estimator.DNNClassifier(
    feature_columns=feature_cols,
    hidden_units=[128, 64],
    n_classes=10,
    config=run_config
)

# 启动训练
estimator.train(input_fn=train_input_fn, steps=10000)

3. 自定义训练循环（tf.GradientTape）场景 #

需要手动在步数整除500时触发保存操作：

• 示例代码：

checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
manager = tf.train.CheckpointManager(checkpoint, './checkpoints', max_to_keep=5)

step = 0
for epoch in range(epochs):
    for x, y in train_dataset:
        with tf.GradientTape() as tape:
            predictions = model(x)
            loss = loss_fn(y, predictions)
        gradients = tape.gradient(loss, model.trainable_variables)
        optimizer.apply_gradients(zip(gradients, model.trainable_variables))
        
        step += 1
        if step % 500 == 0:
            manager.save()
            print(f"检查点已保存，当前步数：{step}")

内容的提问来源于stack exchange，提问作者Shaw AN