The following code is taken from tensorflow official tpu repository
def cosine_learning_rate_with_linear_warmup(global_step,
init_learning_rate,
warmup_learning_rate,
warmup_steps,
total_steps):
"""Creates the cosine learning rate tensor with linear warmup."""
global_step = tf.cast(global_step, dtype=tf.float32)
linear_warmup = (warmup_learning_rate + global_step / warmup_steps *
(init_learning_rate - warmup_learning_rate))
cosine_learning_rate = (
init_learning_rate * (tf.cos(
np.pi * (global_step - warmup_steps) / (total_steps - warmup_steps))
+ 1.0) / 2.0)
learning_rate = tf.where(global_step < warmup_steps,
linear_warmup, cosine_learning_rate)
return learning_rate
For the meaning of the five parameters, just look at the picture. It’s easy to take a look at the code.
In the warmup stage, the learning rate changes from warmup_learning_rate
to init_learning_rate
. In this stage, the learning rate increases or decreases linearly.
In the cosine decay stage, the learning rate decays like this:
l r = c o s ( g l − w t − w π ) + 1 2 ∗ i n i t _ l e a r n i n g _ r a t e lr = \frac{ cos \left ( \frac{gl-w} {t-w} \pi \right ) + 1 }{ 2 } * init\_learning\_rate lr=2cos(t−wg l − wp )+1∗init_learning_rate
c o s cos Variables in cos :
- g l gl g l是global _ step global\_stepglobal_step
- w w w 是 w a r m u p _ s t e p s warmup\_steps warmup_steps
- t t t 是 t o t a l _ s t e p total\_step total_step
The attenuation curve is shown in the blue box in the figure below:
the degree of decline gradually accelerates at first, then gradually slows down, and converges to a very small value.