21cd0774c8c3d41a8464427c81629075c953e7e3,models/official/retinanet/retinanet_model.py,,learning_rate_schedule,#Any#Any#Any#Any#Any#Any#,117

Before Change

  Handles linear scaling rule, gradual warmup, and LR decay.
  // lr_warmup_init is the starting learning rate; the learning rate is linearly
  // scaled up to the full learning rate after `lr_warmup_steps` before decaying.
  linear_warmup = [(lr_warmup_init + float(step) / lr_warmup_step *
                    (1 - lr_warmup_init), step)
                   for step in range(lr_warmup_step)]
  lr_schedule = linear_warmup + [[1.0, lr_warmup_step],
                                 [0.1, first_lr_drop_step],
                                 [0.01, second_lr_drop_step]]

After Change

  linear_warmup = (lr_warmup_init +
                   (tf.cast(global_step, dtype=tf.float32) / lr_warmup_step *
                    (adjusted_learning_rate - lr_warmup_init)))
  learning_rate = tf.where(global_step < lr_warmup_step,
                           linear_warmup, adjusted_learning_rate)
  lr_schedule = [[1.0, lr_warmup_step],
                 [0.1, first_lr_drop_step],
                 [0.01, second_lr_drop_step]]
  for mult, start_global_step in lr_schedule:

Instances