TF Keras ModelCheckpoint 文件路径批次号

Question

我习惯ModelCheckpoint在每个时期每 500 个批次保存一次检查点。它记录在此处https://www.tensorflow.org/api_docs/python/tf/keras/callbacks/ModelCheckpoint。

我该如何设置filepath以包含批次号？我知道我可以在 中使用{epoch}和 参数logs。

Answer 1

假设你使用的是tf.keras.callbacks.ModelCheckpointwith save_freq = int（需要在一定数量的批次后保存），你可以创建一个类，继承ModelCheckpoint并修改类方法on_train_batch_end

class CustomCallback(tf.keras.callbacks.ModelCheckpoint):
    def __init__(self, filepath, save_freq):
        self.model_name = filepath
        self.save_freq = save_freq
        super().__init__(self.model_name, save_freq=self.save_freq)
            
    def on_train_batch_end(self, batch, logs=None):
        if self._should_save_on_batch(batch):
            filename = self.model_name + "epoch_" + str(self._current_epoch+1) + "_batch_ " + str(batch+1) + '.tf'
            self.model.save_weights(filename)
            print("\nsaved checkpoint: " + filename + "\n")

然后在 model.fit 中添加该类的实例。

SAVE_FREQ = 200 # number of batches 
custom_callback = CustomCallback(filepath="checkpoint_", save_freq=SAVE_FREQ)
model.fit(..., callbacks=[custom_callback])

这会将纪元和批次号添加到检查点文件名中。

Epoch 1/3
199/422 [=============>................] - ETA: 6s - loss: 0.0261 - accuracy: 0.9915
saved checkpoint: checkpoint_epoch_0_batch_200.tf

399/422 [===========================>..] - ETA: 0s - loss: 0.0263 - accuracy: 0.9914
saved checkpoint: checkpoint_epoch_0_batch_400.tf

422/422 [==============================] - 13s 31ms/step - loss: 0.0264 - accuracy: 0.9914 - val_loss: 0.0311 - val_accuracy: 0.9920
Epoch 2/3
177/422 [===========>..................] - ETA: 7s - loss: 0.0254 - accuracy: 0.9913
saved checkpoint: checkpoint_epoch_1_batch_178.tf

377/422 [=========================>....] - ETA: 1s - loss: 0.0252 - accuracy: 0.9912
saved checkpoint: checkpoint_epoch_1_batch_378.tf

422/422 [==============================] - 13s 32ms/step - loss: 0.0252 - accuracy: 0.9912 - val_loss: 0.0306 - val_accuracy: 0.9925
Epoch 3/3
156/422 [==========>...................] - ETA: 7s - loss: 0.0253 - accuracy: 0.9914
saved checkpoint: checkpoint_epoch_2_batch_156.tf

355/422 [========================>.....] - ETA: 2s - loss: 0.0246 - accuracy: 0.9919
saved checkpoint: checkpoint_epoch_2_batch_356.tf

422/422 [==============================] - 13s 31ms/step - loss: 0.0245 - accuracy: 0.9919 - val_loss: 0.0294 - val_accuracy: 0.9922