keras.callbacks.Callback()回调函数ModelCheckpoint

model_checkpoint = tf.keras.callbacks.ModelCheckpoint('net.hdf5', monitor='loss',verbose=1, save_best_only=True)
model.fit(data_gen,steps_per_epoch=10,epochs=80,verbose=0,callbacks=[model_checkpoint])

或

cp = keras.callbacks.ModelCheckpoint(model_path, monitor = "val_acc", save_best_only = True, mode="max")
callbacks = [cp]
history = model.fit(train_db, epochs = 20, validation_data = valid_db, callbacks = callbacks)

"val_acc",既然是准确率，模式当然是最大值。

“val_loss”，模式是“min”

verbose: 0, 1 或 2。日志显示模式。 0 = 安静模式, 1 = 进度条, 2 = 每轮一行

其他

steps_per_epoch=100，如果是yield方法生成批数据，那么steps_per_epoch需指定一个值。

一般说来，这样训练

model_checkpoint = tf.keras.callbacks.ModelCheckpoint(process_str_id+".hdf5", verbose=1,monitor = "val_loss", save_best_only = True, mode="min")

model_history=model.fit(train_gen,steps_per_epoch=100,epochs=8,validation_data = val_gen, validation_steps=20, callbacks=[model_checkpoint]) 

设置一个验证好处很多，至少不用频繁储存系数。