94 lines
3.4 KiB
Python
94 lines
3.4 KiB
Python
# MIT License
|
|
#
|
|
# Copyright (c) 2021 Soohwan Kim
|
|
#
|
|
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
# of this software and associated documentation files (the "Software"), to deal
|
|
# in the Software without restriction, including without limitation the rights
|
|
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
# copies of the Software, and to permit persons to whom the Software is
|
|
# furnished to do so, subject to the following conditions:
|
|
#
|
|
# The above copyright notice and this permission notice shall be included in all
|
|
# copies or substantial portions of the Software.
|
|
#
|
|
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
# SOFTWARE.
|
|
|
|
import math
|
|
import torch
|
|
from typing import Optional
|
|
from torch.optim import Optimizer
|
|
|
|
from src.train.lr_scheduler_v2.lr_scheduler import LearningRateScheduler
|
|
|
|
|
|
class TransformerLRScheduler(LearningRateScheduler):
|
|
r"""
|
|
Transformer Learning Rate Scheduler proposed in "Attention Is All You Need"
|
|
|
|
Args:
|
|
optimizer (Optimizer): Optimizer.
|
|
init_lr (float): Initial learning rate.
|
|
peak_lr (float): Maximum learning rate.
|
|
final_lr (float): Final learning rate.
|
|
final_lr_scale (float): Final learning rate scale
|
|
warmup_steps (int): Warmup the learning rate linearly for the first N updates
|
|
decay_steps (int): Steps in decay stages
|
|
"""
|
|
def __init__(
|
|
self,
|
|
optimizer: Optimizer,
|
|
init_lr: float,
|
|
peak_lr: float,
|
|
final_lr: float,
|
|
final_lr_scale: float,
|
|
warmup_steps: int,
|
|
decay_steps: int,
|
|
) -> None:
|
|
assert isinstance(warmup_steps, int), "warmup_steps should be integer type"
|
|
assert isinstance(decay_steps, int), "total_steps should be integer type"
|
|
|
|
super(TransformerLRScheduler, self).__init__(optimizer, init_lr)
|
|
self.final_lr = final_lr
|
|
self.peak_lr = peak_lr
|
|
self.warmup_steps = warmup_steps
|
|
self.decay_steps = decay_steps
|
|
|
|
self.warmup_rate = self.peak_lr / self.warmup_steps
|
|
self.decay_factor = -math.log(final_lr_scale) / self.decay_steps
|
|
|
|
self.init_lr = init_lr
|
|
self.update_steps = 0
|
|
|
|
def _decide_stage(self):
|
|
if self.update_steps < self.warmup_steps:
|
|
return 0, self.update_steps
|
|
|
|
if self.warmup_steps <= self.update_steps < self.warmup_steps + self.decay_steps:
|
|
return 1, self.update_steps - self.warmup_steps
|
|
|
|
return 2, None
|
|
|
|
def step(self, val_loss: Optional[torch.FloatTensor] = None):
|
|
self.update_steps += 1
|
|
stage, steps_in_stage = self._decide_stage()
|
|
|
|
if stage == 0:
|
|
self.lr = self.update_steps * self.warmup_rate
|
|
elif stage == 1:
|
|
self.lr = self.peak_lr * math.exp(-self.decay_factor * steps_in_stage)
|
|
elif stage == 2:
|
|
self.lr = self.final_lr
|
|
else:
|
|
raise ValueError("Undefined stage")
|
|
|
|
self.set_lr(self.optimizer, self.lr)
|
|
|
|
return self.lr
|