如何将更多参数传递给调度程序？

Question

我正在研究示例 in this link，但我不确定调度程序函数是如何同时接收纪元和学习率 (lr) 的。他们是如何通过的？我怎样才能传递更多参数？

我尝试按照此示例进行操作，但收到一条错误消息，指出调度程序收到了一个额外的参数“lr”，因此我不确定如何解决该问题。

Answer 1

你通常使用tf.keras.optimizers.schedules and pass them directly to the model optimizer. The link you are referring to is actually a callback that needs some kind of scheduler function. Here is an example of an tf.keras.optimizers.schedules.ExponentialDecay function according to the docs:

initial_learning_rate = 0.1
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
    initial_learning_rate,
    decay_steps=100000,
    decay_rate=0.96,
    staircase=True)

model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=lr_schedule),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

还有这个link shows you a good example of how to use a custom callable, which uses the initial learning rate defined in the optimizer. In this case, it is 0.01.

如何将更多参数传递给调度程序？

How to pass more arguments to scheduler?

python

scheduler

tensorflow