Spaces:
No application file
No application file
File size: 4,134 Bytes
430de99 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 |
# Copyright (c) Facebook, Inc. and its affiliates.
import math
from bisect import bisect_right
from typing import List
import torch
# NOTE: PyTorch's LR scheduler interface uses names that assume the LR changes
# only on epoch boundaries. We typically use iteration based schedules instead.
# As a result, "epoch" (e.g., as in self.last_epoch) should be understood to mean
# "iteration" instead.
# 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: torch.optim.Optimizer,
milestones: List[int],
gamma: float = 0.1,
warmup_factor: float = 0.001,
warmup_iters: int = 1000,
warmup_method: str = "linear",
last_epoch: int = -1,
):
if not list(milestones) == sorted(milestones):
raise ValueError(
"Milestones should be a list of" " increasing integers. Got {}", milestones
)
self.milestones = milestones
self.gamma = gamma
self.warmup_factor = warmup_factor
self.warmup_iters = warmup_iters
self.warmup_method = warmup_method
super().__init__(optimizer, last_epoch)
def get_lr(self) -> List[float]:
warmup_factor = _get_warmup_factor_at_iter(
self.warmup_method, self.last_epoch, self.warmup_iters, self.warmup_factor
)
return [
base_lr * warmup_factor * self.gamma ** bisect_right(self.milestones, self.last_epoch)
for base_lr in self.base_lrs
]
def _compute_values(self) -> List[float]:
# The new interface
return self.get_lr()
class WarmupCosineLR(torch.optim.lr_scheduler._LRScheduler):
def __init__(
self,
optimizer: torch.optim.Optimizer,
max_iters: int,
warmup_factor: float = 0.001,
warmup_iters: int = 1000,
warmup_method: str = "linear",
last_epoch: int = -1,
):
self.max_iters = max_iters
self.warmup_factor = warmup_factor
self.warmup_iters = warmup_iters
self.warmup_method = warmup_method
super().__init__(optimizer, last_epoch)
def get_lr(self) -> List[float]:
warmup_factor = _get_warmup_factor_at_iter(
self.warmup_method, self.last_epoch, self.warmup_iters, self.warmup_factor
)
# Different definitions of half-cosine with warmup are possible. For
# simplicity we multiply the standard half-cosine schedule by the warmup
# factor. An alternative is to start the period of the cosine at warmup_iters
# instead of at 0. In the case that warmup_iters << max_iters the two are
# very close to each other.
return [
base_lr
* warmup_factor
* 0.5
* (1.0 + math.cos(math.pi * self.last_epoch / self.max_iters))
for base_lr in self.base_lrs
]
def _compute_values(self) -> List[float]:
# The new interface
return self.get_lr()
def _get_warmup_factor_at_iter(
method: str, iter: int, warmup_iters: int, warmup_factor: float
) -> float:
"""
Return the learning rate warmup factor at a specific iteration.
See :paper:`ImageNet in 1h` for more details.
Args:
method (str): warmup method; either "constant" or "linear".
iter (int): iteration at which to calculate the warmup factor.
warmup_iters (int): the number of warmup iterations.
warmup_factor (float): the base warmup factor (the meaning changes according
to the method used).
Returns:
float: the effective warmup factor at the given iteration.
"""
if iter >= warmup_iters:
return 1.0
if method == "constant":
return warmup_factor
elif method == "linear":
alpha = iter / warmup_iters
return warmup_factor * (1 - alpha) + alpha
else:
raise ValueError("Unknown warmup method: {}".format(method))
|