mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-08 17:42:21 +00:00
137 lines
4.8 KiB
Python
137 lines
4.8 KiB
Python
# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey)
|
|
#
|
|
# See ../LICENSE for clarification regarding multiple authors
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
import logging
|
|
from typing import Dict, List, Optional, Union
|
|
from torch.optim import Optimizer
|
|
|
|
|
|
class LRScheduler(object):
|
|
"""
|
|
Base-class for learning rate schedulers where the learning-rate depends on both the
|
|
batch and the epoch.
|
|
"""
|
|
|
|
def __init__(self, optimizer: Optimizer, verbose: bool = False):
|
|
# Attach optimizer
|
|
if not isinstance(optimizer, Optimizer):
|
|
raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__))
|
|
self.optimizer = optimizer
|
|
self.verbose = verbose
|
|
|
|
for group in optimizer.param_groups:
|
|
group.setdefault("base_lr", group["lr"])
|
|
|
|
self.base_lrs = [group["base_lr"] for group in optimizer.param_groups]
|
|
|
|
self.epoch = 0
|
|
self.batch = 0
|
|
|
|
def state_dict(self):
|
|
"""Returns the state of the scheduler as a :class:`dict`.
|
|
|
|
It contains an entry for every variable in self.__dict__ which
|
|
is not the optimizer.
|
|
"""
|
|
return {
|
|
"base_lrs": self.base_lrs,
|
|
"epoch": self.epoch,
|
|
"batch": self.batch,
|
|
}
|
|
|
|
def load_state_dict(self, state_dict):
|
|
"""Loads the schedulers state.
|
|
|
|
Args:
|
|
state_dict (dict): scheduler state. Should be an object returned
|
|
from a call to :meth:`state_dict`.
|
|
"""
|
|
self.__dict__.update(state_dict)
|
|
|
|
def get_last_lr(self) -> List[float]:
|
|
"""Return last computed learning rate by current scheduler. Will be a list of float."""
|
|
return self._last_lr
|
|
|
|
def get_lr(self):
|
|
# Compute list of learning rates from self.epoch and self.batch and
|
|
# self.base_lrs; this must be overloaded by the user.
|
|
# e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
|
|
raise NotImplementedError
|
|
|
|
def step_batch(self, batch: Optional[int] = None) -> None:
|
|
# Step the batch index, or just set it. If `batch` is specified, it
|
|
# must be the batch index from the start of training, i.e. summed over
|
|
# all epochs.
|
|
# You can call this in any order; if you don't provide 'batch', it should
|
|
# of course be called once per batch.
|
|
if batch is not None:
|
|
self.batch = batch
|
|
else:
|
|
self.batch = self.batch + 1
|
|
self._set_lrs()
|
|
|
|
def step_epoch(self, epoch: Optional[int] = None):
|
|
# Step the epoch index, or just set it. If you provide the 'epoch' arg,
|
|
# you should call this at the start of the epoch; if you don't provide the 'epoch'
|
|
# arg, you should call it at the end of the epoch.
|
|
if epoch is not None:
|
|
self.epoch = epoch
|
|
else:
|
|
self.epoch = self.epoch + 1
|
|
self._set_lrs()
|
|
|
|
def _set_lrs(self):
|
|
values = self.get_lr()
|
|
assert len(values) == len(self.optimizer.param_groups)
|
|
|
|
for i, data in enumerate(zip(self.optimizer.param_groups, values)):
|
|
param_group, lr = data
|
|
param_group["lr"] = lr
|
|
self.print_lr(self.verbose, i, lr)
|
|
self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
|
|
|
|
def print_lr(self, is_verbose, group, lr):
|
|
"""Display the current learning rate."""
|
|
if is_verbose:
|
|
logging.warn(
|
|
f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
|
|
f" of group {group} to {lr:.4e}."
|
|
)
|
|
|
|
class Noam(LRScheduler):
|
|
"""
|
|
The LR scheduler proposed by Noam
|
|
Ref: "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
optimizer: Optimizer,
|
|
warmup_batches: Union[int, float] = 10000.0,
|
|
verbose: bool = False,
|
|
):
|
|
super().__init__(optimizer, verbose)
|
|
self.warmup_batches = warmup_batches
|
|
self.normalize = self.warmup_batches ** (-0.5)
|
|
|
|
def get_lr(self):
|
|
warmup_factor = 0 if self.batch == 0 else min(
|
|
self.batch ** (-0.5),
|
|
self.batch * self.warmup_batches ** (-1.5)
|
|
) / self.normalize
|
|
|
|
return [x * warmup_factor for x in self.base_lrs]
|