3v324v23's picture
add
c310e19
raw
history blame
2.29 kB
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
from bisect import bisect_right
import torch
# FIXME ideally this would be achieved with a CombinedLRScheduler,
# separating MultiStepLR with WarmupLR
# but the current LRScheduler design doesn't allow it
class WarmupMultiStepLR(torch.optim.lr_scheduler._LRScheduler):
def __init__(
self,
optimizer,
milestones,
gamma=0.1,
warmup_factor=1.0 / 3,
warmup_iters=500,
warmup_method="linear",
last_epoch=-1,
pow_schedule_mode = False,
max_iter = 300000,
lr_pow = 0.9
):
if not list(milestones) == sorted(milestones):
raise ValueError(
"Milestones should be a list of" " increasing integers. Got {}",
milestones,
)
if warmup_method not in ("constant", "linear"):
raise ValueError(
"Only 'constant' or 'linear' warmup_method accepted"
"got {}".format(warmup_method)
)
self.milestones = milestones
self.gamma = gamma
self.warmup_factor = warmup_factor
self.warmup_iters = warmup_iters
self.warmup_method = warmup_method
self.pow_schedule_mode = pow_schedule_mode
self.max_iter = max_iter
self.lr_pow = lr_pow
super(WarmupMultiStepLR, self).__init__(optimizer, last_epoch)
def get_lr(self):
warmup_factor = 1
if self.last_epoch < self.warmup_iters:
if self.warmup_method == "constant":
warmup_factor = self.warmup_factor
elif self.warmup_method == "linear":
alpha = self.last_epoch / self.warmup_iters
warmup_factor = self.warmup_factor * (1 - alpha) + alpha
if self.pow_schedule_mode:
scale_running_lr = ((1. - float(self.last_epoch) / self.max_iter) ** self.lr_pow)
return [
base_lr * warmup_factor * scale_running_lr
for base_lr in self.base_lrs
]
else:
return [
base_lr
* warmup_factor
* self.gamma ** bisect_right(self.milestones, self.last_epoch)
for base_lr in self.base_lrs
]