tf.compat.v1.train.cosine_decay

Applies cosine decay to the learning rate.

tf.compat.v1.train.cosine_decay(
    learning_rate, global_step, decay_steps, alpha=0.0, name=None
)

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

The function returns the decayed learning rate. It is computed as:

global_step = min(global_step, decay_steps)
cosine_decay = 0.5 * (1 + cos(pi * global_step / decay_steps))
decayed = (1 - alpha) * cosine_decay + alpha
decayed_learning_rate = learning_rate * decayed

Example usage:

decay_steps = 1000
lr_decayed = cosine_decay(learning_rate, global_step, decay_steps)

References:

Stochastic Gradient Descent with Warm Restarts: Loshchilov et al., 2017 (pdf)

Eager Compatibility

When eager execution is enabled, this function returns a function which in turn returns the decayed learning rate Tensor. This can be useful for changing the learning rate value across different invocations of optimizer functions.