Keras训练每隔一个epoch所有日志指标均为0问题求助

核心原因 #

你手动指定了steps_per_epoch=90和validation_steps=10，但ImageDataGenerator的生成器是状态化的——第一个epoch结束后，生成器的内部指针已指向训练数据末尾。第二个epoch时，生成器未自动重置，导致指定步数内未读取到有效样本，因此loss、accuracy等指标全部为0，训练时间也几乎为0。

修复方案 #

方案1：移除固定步数参数（推荐） #

Keras可自动根据生成器的总样本数和batch_size计算每个epoch的步数，无需手动指定。修改model.fit()代码如下：

history1=model1.fit(
    train_generator_CD,
    validation_data = test_generator_CD,
    epochs = 20,
    callbacks=[myCallback()]
)

方案2：手动重置生成器状态 #

如果必须保留固定步数，可自定义回调函数在每个epoch开始前重置生成器状态：

class ResetGeneratorCallback(keras.callbacks.Callback):
    def on_epoch_begin(self, epoch, logs=None):
        train_generator_CD.reset()
        test_generator_CD.reset()

history1=model1.fit(
    train_generator_CD,
    validation_data = test_generator_CD,
    epochs = 20,
    steps_per_epoch = 90,
    validation_steps = 10,
    callbacks=[myCallback(), ResetGeneratorCallback()]
)

补充说明 #

你的训练数据集有22500张图（25000的90%），batch_size=250，22500/250=90，所以steps_per_epoch=90的数值是正确的，但生成器不会自动重置指针，才导致偶数epoch无数据可读。移除手动步数参数后，Keras会在每个epoch结束后自动重置生成器，彻底避免这个问题。

原始异常日志 #

Epoch 1/20
90/90 ━━━━━━━━━━━━━━━━━━━━ 170s 2s/step - accuracy: 0.9974 - loss: 0.4769 - val_accuracy: 0.7968 - val_loss: 0.9699
Epoch 2/20
90/90 ━━━━━━━━━━━━━━━━━━━━ 0s 1ms/step - accuracy: 0.0000e+00 - loss: 0.0000e+00 - val_accuracy: 0.0000e+00 - val_loss: 0.0000e+00
Epoch 3/20
90/90 ━━━━━━━━━━━━━━━━━━━━ 175s 2s/step - accuracy: 0.9988 - loss: 0.4260 - val_accuracy: 0.8052 - val_loss: 0.9283
Epoch 4/20
90/90 ━━━━━━━━━━━━━━━━━━━━ 0s 728us/step - accuracy: 0.0000e+00 - loss: 0.0000e+00 - val_accuracy: 0.0000e+00 - val_loss: 0.0000e+00
etc...

原始模型代码 #

model1 = keras.models.Sequential([
    keras.layers.Conv2D(16,(3,3), activation='relu', input_shape=(150, 150, 3)),
    keras.layers.MaxPooling2D(2,2),
    keras.layers.BatchNormalization(),
    keras.layers.Conv2D(32,(3,3), activation='relu'),
    keras.layers.MaxPooling2D(2,2),
    keras.layers.BatchNormalization(),
    keras.layers.Conv2D(64,(3,3), activation='relu'),
    keras.layers.MaxPooling2D(2,2),
    keras.layers.BatchNormalization(),
    keras.layers.Flatten(),
    keras.layers.Dense(512, activation='relu', kernel_regularizer=regularizers.l2(0.001)),
    keras.layers.Dropout(0.2),
    keras.layers.Dense(1, activation='sigmoid')
])
model1.compile(optimizer=Adam(learning_rate=0.0002), loss='binary_crossentropy', metrics=['accuracy'])

train_CD = ImageDataGenerator(rescale=1.0/255.)
train_generator_CD = train_CD.flow_from_directory(
    './images/cat_dog/train_data/',
    target_size = (150, 150),
    batch_size = 250,
    class_mode = 'binary')

test_CD = ImageDataGenerator(rescale=1.0/255.)
test_generator_CD = test_CD.flow_from_directory(
    './images/cat_dog/test_data/',
    target_size = (150, 150),
    batch_size = 250,
    class_mode = 'binary')

history1=model1.fit(
    train_generator_CD,
    validation_data = test_generator_CD,
    epochs = 20,
    steps_per_epoch = 90,
    validation_steps = 10,
    callbacks=[myCallback()]
    )

内容的提问来源于stack exchange，提问作者Arden McArthur