A LearningRateSchedule that uses a cosine decay schedule

  alpha = 0,
  name = NULL



A scalar float32 or float64 Tensor or a R number. The initial learning rate.


A scalar int32 or int64 Tensor or an R number. Number of steps to decay over.


A scalar float32 or float64 Tensor or an R number. Minimum learning rate value as a fraction of initial_learning_rate.


For backwards and forwards compatibility


String. Optional name of the operation. Defaults to 'CosineDecay'.


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

When training a model, it is often useful 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 is 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:

decayed_learning_rate <- function(step) {
  step <- min(step, decay_steps)
  cosine_decay = <- 0.5 * (1 + cos(pi * step / decay_steps))
  decayed <- (1 - alpha) * cosine_decay + alpha
  initial_learning_rate * decayed

Example usage:

decay_steps <- 1000
lr_decayed_fn <-
  learning_rate_schedule_cosine_decay(initial_learning_rate, decay_steps)

You can pass this schedule directly into a keras Optimizer as the learning_rate.