A LearningRateSchedule that uses a cosine decay schedule.

Inherits From: LearningRateSchedule

See [Loshchilov & Hutter, ICLR2016], SGDR: Stochastic Gradient Descent with Warm Restarts.

When training a model, it is often recommended to lower the learning rate as the training progresses. This schedule applies a cosine decay function to an optimizer step, given a provided initial learning rate. It requires a step value to compute the decayed learning rate. You can just pass a TensorFlow variable that you increment at each training step.

The schedule a 1-arg callable that produces a decayed learning rate when passed the current optimizer step. This can be useful for changing the learning rate value across different invocations of optimizer functions. It is computed as:

def decayed_learning_rate(step):
  step = min(step, decay_steps)
  cosine_decay = 0.5 * (1 + cos(pi * step / decay_steps))
  decayed = (1 - alpha) * cosine_decay + alpha
  return initial_learning_rate * decayed

Example usage:

decay_steps = 1000
lr_decayed_fn = tf.keras.experimental.CosineDecay(
    initial_learning_rate, decay_steps)

You can pass this schedule directly into a tf.keras.optimizers.Optimizer as the learning rate. The learning rate schedule is also serializable and deserializable using tf.keras.optimizers.schedules.serialize and tf.keras.optimizers.schedules.deserialize.

A 1-arg callable learning rate schedule that takes the current optimizer step and outputs the decayed learning rate, a scalar Tensor of the same type as initial_learning_rate.

initial_learning_rate A scalar float32 or float64 Tensor or a Python number. The initial learning rate.
decay_steps A scalar int32 or int64 Tensor or a Python number. Number of steps to decay over.
alpha A scalar