github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-30 22:24:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

[misc] update pre-commit and run all files (#4752)

Browse Source 
* [misc] update pre-commit

* [misc] run pre-commit

* [misc] remove useless configuration files

* [misc] ignore cuda for clang-format
This commit is contained in:
Hongxin Liu
2023-09-19 14:20:26 +08:00
committed by GitHub
parent 3c6b831c26
commit 079bf3cb26
1268 changed files with 50037 additions and 38444 deletions
Download Patch File Download Diff File Expand all files Collapse all files

102
examples/tutorial/sequence_parallel/lr_scheduler/annealing_lr.py
View File

@@ -21,16 +21,17 @@ import math
class AnnealingLR(object):
"""Anneals the learning rate."""
def __init__(self,
optimizer,
max_lr,
min_lr,
warmup_steps,
decay_steps,
decay_style,
use_checkpoint_lr_scheduler=True,
override_lr_scheduler=False):
def __init__(
self,
optimizer,
max_lr,
min_lr,
warmup_steps,
decay_steps,
decay_style,
use_checkpoint_lr_scheduler=True,
override_lr_scheduler=False,
):
# Class values.
self.optimizer = optimizer
@@ -50,23 +51,21 @@ class AnnealingLR(object):
self.override_lr_scheduler = override_lr_scheduler
self.use_checkpoint_lr_scheduler = use_checkpoint_lr_scheduler
if self.override_lr_scheduler:
assert not self.use_checkpoint_lr_scheduler, 'both override and '\
'use-checkpoint are set.'
assert not self.use_checkpoint_lr_scheduler, "both override and " "use-checkpoint are set."
# Set the learning rate
self.step(0)
def get_lr(self):
"""Learning rate decay functions from:
https://openreview.net/pdf?id=BJYwwY9ll pg. 4"""
https://openreview.net/pdf?id=BJYwwY9ll pg. 4"""
# Use linear warmup for the initial part.
if self.warmup_steps > 0 and self.num_steps <= self.warmup_steps:
return self.max_lr * float(self.num_steps) / \
float(self.warmup_steps)
return self.max_lr * float(self.num_steps) / float(self.warmup_steps)
# If the learning rate is constant, just return the initial value.
if self.decay_style == 'constant':
if self.decay_style == "constant":
return self.max_lr
# For any steps larger than `self.decay_steps`, use `self.min_lr`.
@@ -81,13 +80,12 @@ class AnnealingLR(object):
assert decay_ratio <= 1.0
delta_lr = self.max_lr - self.min_lr
if self.decay_style == 'linear':
coeff = (1.0 - decay_ratio)
elif self.decay_style == 'cosine':
if self.decay_style == "linear":
coeff = 1.0 - decay_ratio
elif self.decay_style == "cosine":
coeff = 0.5 * (math.cos(math.pi * decay_ratio) + 1.0)
else:
raise Exception('{} decay style is not supported.'.format(
self.decay_style))
raise Exception("{} decay style is not supported.".format(self.decay_style))
return self.min_lr + coeff * delta_lr
@@ -96,16 +94,16 @@ class AnnealingLR(object):
self.num_steps += increment
new_lr = self.get_lr()
for group in self.optimizer.param_groups:
group['lr'] = new_lr
group["lr"] = new_lr
def state_dict(self):
state_dict = {
'max_lr': self.max_lr,
'warmup_steps': self.warmup_steps,
'num_steps': self.num_steps,
'decay_style': self.decay_style,
'decay_steps': self.decay_steps,
'min_lr': self.min_lr
"max_lr": self.max_lr,
"warmup_steps": self.warmup_steps,
"num_steps": self.num_steps,
"decay_style": self.decay_style,
"decay_steps": self.decay_steps,
"min_lr": self.min_lr,
}
return state_dict
@@ -116,43 +114,35 @@ class AnnealingLR(object):
return cls_value
if not self.use_checkpoint_lr_scheduler:
assert cls_value == sd_value, \
f'AnnealingLR: class input value {cls_value} and checkpoint' \
f'value {sd_value} for {name} do not match'
assert cls_value == sd_value, (
f"AnnealingLR: class input value {cls_value} and checkpoint" f"value {sd_value} for {name} do not match"
)
return sd_value
def load_state_dict(self, sd):
if 'start_lr' in sd:
max_lr_ = sd['start_lr']
if "start_lr" in sd:
max_lr_ = sd["start_lr"]
else:
max_lr_ = sd['max_lr']
self.max_lr = self._check_and_set(self.max_lr, max_lr_,
'learning rate')
max_lr_ = sd["max_lr"]
self.max_lr = self._check_and_set(self.max_lr, max_lr_, "learning rate")
self.min_lr = self._check_and_set(self.min_lr, sd['min_lr'],
'minimum learning rate')
self.min_lr = self._check_and_set(self.min_lr, sd["min_lr"], "minimum learning rate")
if 'warmup_iter' in sd:
warmup_steps_ = sd['warmup_iter']
if "warmup_iter" in sd:
warmup_steps_ = sd["warmup_iter"]
else:
warmup_steps_ = sd['warmup_steps']
self.warmup_steps = self._check_and_set(self.warmup_steps,
warmup_steps_,
'warmup iterations')
warmup_steps_ = sd["warmup_steps"]
self.warmup_steps = self._check_and_set(self.warmup_steps, warmup_steps_, "warmup iterations")
if 'end_iter' in sd:
decay_steps_ = sd['end_iter']
if "end_iter" in sd:
decay_steps_ = sd["end_iter"]
else:
decay_steps_ = sd['decay_steps']
self.decay_steps = self._check_and_set(self.decay_steps, decay_steps_,
'total number of iterations')
self.decay_style = self._check_and_set(self.decay_style,
sd['decay_style'],
'decay style')
decay_steps_ = sd["decay_steps"]
self.decay_steps = self._check_and_set(self.decay_steps, decay_steps_, "total number of iterations")
self.decay_style = self._check_and_set(self.decay_style, sd["decay_style"], "decay style")
if 'num_iters' in sd:
num_steps = sd['num_iters']
if "num_iters" in sd:
num_steps = sd["num_iters"]
else:
num_steps = sd['num_steps']
num_steps = sd["num_steps"]
self.step(increment=num_steps)