tf.compat.v1.train.linear_cosine_decay

View source on GitHub

Applies linear cosine decay to the learning rate.

tf.compat.v1.train.linear_cosine_decay(
    learning_rate, global_step, decay_steps, num_periods=0.5, alpha=0.0, beta=0.001,
    name=None
)

See [Bello et al., ICML2017] Neural Optimizer Search with RL. https://arxiv.org/abs/1709.07417

For the idea of warm starts here controlled by num_periods, see [Loshchilov & Hutter, ICLR2016] SGDR: Stochastic Gradient Descent with Warm Restarts. https://arxiv.org/abs/1608.03983

Note that linear cosine decay is more aggressive than cosine decay and larger initial learning rates can typically be used.

When training a model, it is often recommended to lower the learning rate as the training progresses. This function applies a linear 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)
linear_decay = (decay_steps - global_step) / decay_steps)
cosine_decay = 0.5 * (
    1 + cos(pi * 2 * num_periods * global_step / decay_steps))
decayed = (alpha + linear_decay) * cosine_decay + beta
decayed_learning_rate = learning_rate * decayed

Example usage:

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

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.