Allow multi_step_lr_scheduler to load a new LR schedule when restoring state
This commit is contained in:
parent
e37726f302
commit
6b45b35447
|
@ -54,6 +54,12 @@ class MultiStepLR_Restart(_LRScheduler):
|
|||
for group in self.optimizer.param_groups
|
||||
]
|
||||
|
||||
# Allow this scheduler to use newly appointed milestones partially through a training run..
|
||||
def load_state_dict(self, s):
|
||||
milestones_cache = self.milestones
|
||||
super(MultiStepLR_Restart, self).load_state_dict(s)
|
||||
self.milestones = milestones_cache
|
||||
|
||||
|
||||
class CosineAnnealingLR_Restart(_LRScheduler):
|
||||
def __init__(self, optimizer, T_period, restarts=None, weights=None, eta_min=0, last_epoch=-1):
|
||||
|
|
Loading…
Reference in New Issue
Block a user