Spaces:
Running
Running
File size: 23,978 Bytes
0883aa1 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 |
# This module is from [WeNet](https://github.com/wenet-e2e/wenet).
# ## Citations
# ```bibtex
# @inproceedings{yao2021wenet,
# title={WeNet: Production oriented Streaming and Non-streaming End-to-End Speech Recognition Toolkit},
# author={Yao, Zhuoyuan and Wu, Di and Wang, Xiong and Zhang, Binbin and Yu, Fan and Yang, Chao and Peng, Zhendong and Chen, Xiaoyu and Xie, Lei and Lei, Xin},
# booktitle={Proc. Interspeech},
# year={2021},
# address={Brno, Czech Republic },
# organization={IEEE}
# }
# @article{zhang2022wenet,
# title={WeNet 2.0: More Productive End-to-End Speech Recognition Toolkit},
# author={Zhang, Binbin and Wu, Di and Peng, Zhendong and Song, Xingchen and Yao, Zhuoyuan and Lv, Hang and Xie, Lei and Yang, Chao and Pan, Fuping and Niu, Jianwei},
# journal={arXiv preprint arXiv:2203.15455},
# year={2022}
# }
#
from typing import Union
import math
import warnings
import torch
from torch.optim.lr_scheduler import _LRScheduler
class WarmupLR(_LRScheduler):
"""The WarmupLR scheduler
This scheduler is almost same as NoamLR Scheduler except for following
difference:
NoamLR:
lr = optimizer.lr * model_size ** -0.5
* min(step ** -0.5, step * warmup_step ** -1.5)
WarmupLR:
lr = optimizer.lr * warmup_step ** 0.5
* min(step ** -0.5, step * warmup_step ** -1.5)
Note that the maximum lr equals to optimizer.lr in this scheduler.
"""
def __init__(
self,
optimizer: torch.optim.Optimizer,
warmup_steps: Union[int, float] = 25000,
last_epoch: int = -1,
):
self.warmup_steps = warmup_steps
# __init__() must be invoked before setting field
# because step() is also invoked in __init__()
super().__init__(optimizer, last_epoch)
def __repr__(self):
return f"{self.__class__.__name__}(warmup_steps={self.warmup_steps})"
def get_lr(self):
step_num = self.last_epoch + 1
if self.warmup_steps == 0:
return [lr * step_num**-0.5 for lr in self.base_lrs]
else:
return [
lr
* self.warmup_steps**0.5
* min(step_num**-0.5, step_num * self.warmup_steps**-1.5)
for lr in self.base_lrs
]
def set_step(self, step: int):
self.last_epoch = step
class WarmupPolicy(_LRScheduler):
"""Adds warmup kwargs and warmup logic to lr policy.
All arguments should be passed as kwargs for clarity,
Args:
warmup_steps: Number of training steps in warmup stage
warmup_ratio: Ratio of warmup steps to total steps
max_steps: Total number of steps while training or `None` for
infinite training
"""
def __init__(
self,
optimizer,
*,
warmup_steps=None,
warmup_ratio=None,
max_steps=None,
min_lr=0.0,
last_epoch=-1,
):
assert not (
warmup_steps is not None and warmup_ratio is not None
), "Either use particular number of step or ratio"
assert (
warmup_ratio is None or max_steps is not None
), "If there is a ratio, there should be a total steps"
# It is necessary to assign all attributes *before* __init__,
# as class is wrapped by an inner class.
self.max_steps = max_steps
if warmup_steps is not None:
self.warmup_steps = warmup_steps
elif warmup_ratio is not None:
self.warmup_steps = int(warmup_ratio * max_steps)
else:
self.warmup_steps = 0
self.min_lr = min_lr
super().__init__(optimizer, last_epoch)
def get_lr(self):
if not self._get_lr_called_within_step:
warnings.warn(
"To get the last learning rate computed "
"by the scheduler, please use `get_last_lr()`.",
UserWarning,
stacklevel=2,
)
step = self.last_epoch
if step <= self.warmup_steps and self.warmup_steps > 0:
return self._get_warmup_lr(step)
if step > self.max_steps:
return [self.min_lr for _ in self.base_lrs]
return self._get_lr(step)
def _get_warmup_lr(self, step):
lr_val = (step + 1) / (self.warmup_steps + 1)
return [initial_lr * lr_val for initial_lr in self.base_lrs]
def _get_lr(self, step):
"""Simple const lr policy"""
return self.base_lrs
class SquareRootConstantPolicy(_LRScheduler):
"""Adds warmup kwargs and warmup logic to lr policy.
All arguments should be passed as kwargs for clarity,
Args:
warmup_steps: Number of training steps in warmup stage
warmup_ratio: Ratio of warmup steps to total steps
max_steps: Total number of steps while training or `None` for
infinite training
"""
def __init__(
self,
optimizer,
*,
constant_steps=None,
constant_ratio=None,
max_steps=None,
min_lr=0.0,
last_epoch=-1,
):
assert not (
constant_steps is not None and constant_ratio is not None
), "Either use particular number of step or ratio"
assert (
constant_ratio is None or max_steps is not None
), "If there is a ratio, there should be a total steps"
# It is necessary to assign all attributes *before* __init__,
# as class is wrapped by an inner class.
self.max_steps = max_steps
if constant_steps is not None:
self.constant_steps = constant_steps
elif constant_ratio is not None:
self.constant_steps = int(constant_ratio * max_steps)
else:
self.constant_steps = 0
self.constant_lr = 1 / (constant_steps**0.5)
self.min_lr = min_lr
super().__init__(optimizer, last_epoch)
def get_lr(self):
if not self._get_lr_called_within_step:
warnings.warn(
"To get the last learning rate computed "
"by the scheduler, please use `get_last_lr()`.",
UserWarning,
stacklevel=2,
)
step = self.last_epoch
if step <= self.constant_steps:
return [self.constant_lr for _ in self.base_lrs]
if step > self.max_steps:
return [self.min_lr for _ in self.base_lrs]
return self._get_lr(step)
def _get_lr(self, step):
"""Simple const lr policy"""
return self.base_lrs
class WarmupHoldPolicy(WarmupPolicy):
"""Variant of WarmupPolicy which maintains high
learning rate for a defined number of steps.
All arguments should be passed as kwargs for clarity,
Args:
warmup_steps: Number of training steps in warmup stage
warmup_ratio: Ratio of warmup steps to total steps
hold_steps: Number of training steps to
hold the learning rate after warm up
hold_ratio: Ratio of hold steps to total steps
max_steps: Total number of steps while training or `None` for
infinite training
"""
def __init__(
self,
optimizer,
*,
warmup_steps=None,
warmup_ratio=None,
hold_steps=None,
hold_ratio=None,
max_steps=None,
min_lr=0.0,
last_epoch=-1,
):
assert not (
hold_steps is not None and hold_ratio is not None
), "Either use particular number of step or ratio"
assert (
hold_ratio is None or max_steps is not None
), "If there is a ratio, there should be a total steps"
self.min_lr = min_lr
self._last_warmup_lr = 0.0
# Necessary to duplicate as class attributes are hidden in inner class
self.max_steps = max_steps
if warmup_steps is not None:
self.warmup_steps = warmup_steps
elif warmup_ratio is not None:
self.warmup_steps = int(warmup_ratio * max_steps)
else:
self.warmup_steps = 0
if hold_steps is not None:
self.hold_steps = hold_steps + self.warmup_steps
elif hold_ratio is not None:
self.hold_steps = int(hold_ratio * max_steps) + self.warmup_steps
else:
self.hold_steps = 0
super().__init__(
optimizer,
warmup_steps=warmup_steps,
warmup_ratio=warmup_ratio,
max_steps=max_steps,
last_epoch=last_epoch,
min_lr=min_lr,
)
def get_lr(self):
if not self._get_lr_called_within_step:
warnings.warn(
"To get the last learning rate computed by the scheduler,"
" "
"please use `get_last_lr()`.",
UserWarning,
stacklevel=2,
)
step = self.last_epoch
# Warmup phase
if step <= self.warmup_steps and self.warmup_steps > 0:
return self._get_warmup_lr(step)
# Hold phase
if (step >= self.warmup_steps) and (step < self.hold_steps):
return self.base_lrs
if step > self.max_steps:
return [self.min_lr for _ in self.base_lrs]
return self._get_lr(step)
class WarmupAnnealHoldPolicy(_LRScheduler):
"""Adds warmup kwargs and warmup logic to lr policy.
All arguments should be passed as kwargs for clarity,
Args:
warmup_steps: Number of training steps in warmup stage
warmup_ratio: Ratio of warmup steps to total steps
max_steps: Total number of steps while training or `None` for
infinite training
min_lr: Minimum lr to hold the learning rate after decay at.
constant_steps: Number of steps to keep lr constant at.
constant_ratio: Ratio of steps to keep lr constant.
"""
def __init__(
self,
optimizer,
*,
warmup_steps=None,
warmup_ratio=None,
constant_steps=None,
constant_ratio=None,
max_steps=None,
min_lr=0.0,
last_epoch=-1,
):
assert not (
warmup_steps is not None and warmup_ratio is not None
), "Either use particular number of step or ratio"
assert not (
constant_steps is not None and constant_ratio is not None
), "Either use constant_steps or constant_ratio"
assert (
warmup_ratio is None or max_steps is not None
), "If there is a ratio, there should be a total steps"
# It is necessary to assign all attributes *before* __init__,
# as class is wrapped by an inner class.
self.max_steps = max_steps
if warmup_steps is not None:
self.warmup_steps = warmup_steps
elif warmup_ratio is not None:
self.warmup_steps = int(warmup_ratio * max_steps)
else:
self.warmup_steps = 0
if constant_steps is not None:
self.constant_steps = constant_steps
elif constant_ratio is not None:
self.constant_steps = int(constant_ratio * max_steps)
else:
self.constant_steps = 0
self.decay_steps = max_steps - (self.constant_steps + self.warmup_steps)
self.min_lr = min_lr
super().__init__(optimizer, last_epoch)
def get_lr(self):
if not self._get_lr_called_within_step:
warnings.warn(
"To get the last learning rate computed "
"by the scheduler, please use `get_last_lr()`.",
UserWarning,
stacklevel=2,
)
step = self.last_epoch
# Warmup steps
if self.warmup_steps > 0 and step <= self.warmup_steps:
return self._get_warmup_lr(step)
# Constant steps after warmup and decay
if (
self.constant_steps > 0
and (self.warmup_steps + self.decay_steps) < step <= self.max_steps
):
return self._get_constant_lr(step)
# Min lr after max steps of updates
if step > self.max_steps:
return [self.min_lr for _ in self.base_lrs]
return self._get_lr(step)
def _get_warmup_lr(self, step):
lr_val = (step + 1) / (self.warmup_steps + 1)
return [initial_lr * lr_val for initial_lr in self.base_lrs]
def _get_constant_lr(self, step):
return [self.min_lr for _ in self.base_lrs]
def _get_lr(self, step):
"""Simple const lr policy"""
return self.base_lrs
def _squareroot_annealing(initial_lr, step, max_steps, min_lr):
mult = ((max_steps - step) / max_steps) ** 0.5
out_lr = initial_lr * mult
out_lr = max(out_lr, min_lr)
return out_lr
def _square_annealing(initial_lr, step, max_steps, min_lr):
mult = ((max_steps - step) / max_steps) ** 2
out_lr = initial_lr * mult
out_lr = max(out_lr, min_lr)
return out_lr
def _cosine_annealing(initial_lr, step, max_steps, min_lr):
mult = 0.5 * (1 + math.cos(math.pi * step / max_steps))
out_lr = (initial_lr - min_lr) * mult + min_lr
return out_lr
def _linear_warmup_with_cosine_annealing(
max_lr, warmup_steps, step, decay_steps, min_lr
):
assert max_lr > min_lr
# Use linear warmup for the initial part.
if warmup_steps > 0 and step <= warmup_steps:
return max_lr * float(step) / float(warmup_steps)
# For any steps larger than `decay_steps`, use `min_lr`.
if step > warmup_steps + decay_steps:
return min_lr
# If we are done with the warmup period, use the decay style.
num_steps_ = step - warmup_steps
decay_steps_ = decay_steps
decay_ratio = float(num_steps_) / float(decay_steps_)
assert decay_ratio >= 0.0
assert decay_ratio <= 1.0
delta_lr = max_lr - min_lr
coeff = 0.5 * (math.cos(math.pi * decay_ratio) + 1.0)
return min_lr + coeff * delta_lr
def _poly_decay(initial_lr, step, decay_steps, power, min_lr, cycle):
if cycle:
multiplier = 1.0 if step == 0 else math.ceil(step / decay_steps)
decay_steps *= multiplier
else:
step = min(step, decay_steps)
p = step / decay_steps
lr = (initial_lr - min_lr) * math.pow(1.0 - p, power)
lr += min_lr
return lr
def _noam_hold_annealing(
initial_lr, step, warmup_steps, hold_steps, decay_rate, min_lr
):
# hold_steps = total number of steps
# to hold the LR, not the warmup + hold steps.
T_warmup_decay = max(1, warmup_steps**decay_rate)
T_hold_decay = max(1, (step - hold_steps) ** decay_rate)
lr = (initial_lr * T_warmup_decay) / T_hold_decay
lr = max(lr, min_lr)
return lr
class SquareAnnealing(WarmupPolicy):
def __init__(self, optimizer, *, max_steps, min_lr=1e-5, last_epoch=-1, **kwargs):
super().__init__(
optimizer=optimizer,
max_steps=max_steps,
last_epoch=last_epoch,
min_lr=min_lr,
**kwargs,
)
def _get_lr(self, step):
new_lrs = [
_square_annealing(
initial_lr=initial_lr,
step=step - self.warmup_steps,
max_steps=self.max_steps - self.warmup_steps,
min_lr=self.min_lr,
)
for initial_lr in self.base_lrs
]
return new_lrs
class SquareRootAnnealing(WarmupPolicy):
def __init__(self, optimizer, *, max_steps, min_lr=0, last_epoch=-1, **kwargs):
super().__init__(
optimizer=optimizer,
max_steps=max_steps,
last_epoch=last_epoch,
min_lr=min_lr,
**kwargs,
)
def _get_lr(self, step):
new_lrs = [
_squareroot_annealing(
initial_lr=initial_lr,
step=step,
max_steps=self.max_steps,
min_lr=self.min_lr,
)
for initial_lr in self.base_lrs
]
return new_lrs
class CosineAnnealing(WarmupAnnealHoldPolicy):
def __init__(self, optimizer, *, max_steps, min_lr=0, last_epoch=-1, **kwargs):
super().__init__(
optimizer=optimizer,
max_steps=max_steps,
last_epoch=last_epoch,
min_lr=min_lr,
**kwargs,
)
def _get_lr(self, step):
for initial_lr in self.base_lrs:
if initial_lr < self.min_lr:
raise ValueError(
f"{self} received an initial learning rate "
f"that was lower than the minimum learning rate."
)
if self.constant_steps is None or self.constant_steps == 0:
new_lrs = [
_cosine_annealing(
initial_lr=initial_lr,
step=step - self.warmup_steps,
max_steps=self.max_steps - self.warmup_steps,
min_lr=self.min_lr,
)
for initial_lr in self.base_lrs
]
else:
new_lrs = self._get_linear_warmup_with_cosine_annealing_lr(step)
return new_lrs
def _get_warmup_lr(self, step):
if self.constant_steps is None or self.constant_steps == 0:
return super()._get_warmup_lr(step)
else:
# Use linear warmup for the initial part.
return self._get_linear_warmup_with_cosine_annealing_lr(step)
def _get_constant_lr(self, step):
# Only called when `constant_steps` > 0.
return self._get_linear_warmup_with_cosine_annealing_lr(step)
def _get_linear_warmup_with_cosine_annealing_lr(self, step):
# Cosine Schedule for Megatron LM,
# slightly different warmup schedule + constant LR at the end.
new_lrs = [
_linear_warmup_with_cosine_annealing(
max_lr=self.base_lrs[0],
warmup_steps=self.warmup_steps,
step=step,
decay_steps=self.decay_steps,
min_lr=self.min_lr,
)
for _ in self.base_lrs
]
return new_lrs
class NoamAnnealing(_LRScheduler):
def __init__(
self,
optimizer,
*,
d_model,
warmup_steps=None,
warmup_ratio=None,
max_steps=None,
min_lr=0.0,
last_epoch=-1,
):
self._normalize = d_model ** (-0.5)
assert not (
warmup_steps is not None and warmup_ratio is not None
), "Either use particular number of step or ratio"
assert (
warmup_ratio is None or max_steps is not None
), "If there is a ratio, there should be a total steps"
# It is necessary to assign all attributes *before* __init__,
# as class is wrapped by an inner class.
self.max_steps = max_steps
if warmup_steps is not None:
self.warmup_steps = warmup_steps
elif warmup_ratio is not None:
self.warmup_steps = int(warmup_ratio * max_steps)
else:
self.warmup_steps = 0
self.min_lr = min_lr
super().__init__(optimizer, last_epoch)
def get_lr(self):
if not self._get_lr_called_within_step:
warnings.warn(
"To get the last learning rate computed "
"by the scheduler, please use `get_last_lr()`.",
UserWarning,
stacklevel=2,
)
step = max(1, self.last_epoch)
for initial_lr in self.base_lrs:
if initial_lr < self.min_lr:
raise ValueError(
f"{self} received an initial learning rate "
f"that was lower than the minimum learning rate."
)
new_lrs = [
self._noam_annealing(initial_lr=initial_lr, step=step)
for initial_lr in self.base_lrs
]
return new_lrs
def _noam_annealing(self, initial_lr, step):
if self.warmup_steps > 0:
mult = self._normalize * min(
step ** (-0.5), step * (self.warmup_steps ** (-1.5))
)
else:
mult = self._normalize * step ** (-0.5)
out_lr = initial_lr * mult
if step > self.warmup_steps:
out_lr = max(out_lr, self.min_lr)
return out_lr
class NoamHoldAnnealing(WarmupHoldPolicy):
def __init__(
self,
optimizer,
*,
max_steps,
decay_rate=0.5,
min_lr=0.0,
last_epoch=-1,
**kwargs,
):
"""
From Nemo:
Implementation of the Noam Hold Annealing policy
from the SqueezeFormer paper.
Unlike NoamAnnealing, the peak learning rate
can be explicitly set for this scheduler.
The schedule first performs linear warmup,
then holds the peak LR, then decays with some schedule for
the remainder of the steps.
Therefore the min-lr is still dependent
on the hyper parameters selected.
It's schedule is determined by three factors-
Warmup Steps: Initial stage, where linear warmup
occurs uptil the peak LR is reached. Unlike NoamAnnealing,
the peak LR is explicitly stated here instead of a scaling factor.
Hold Steps: Intermediate stage, where the peak LR
is maintained for some number of steps. In this region,
the high peak LR allows the model to converge faster
if training is stable. However the high LR
may also cause instability during training.
Should usually be a significant fraction of training
steps (around 30-40% of the entire training steps).
Decay Steps: Final stage, where the LR rapidly decays
with some scaling rate (set by decay rate).
To attain Noam decay, use 0.5,
for Squeezeformer recommended decay, use 1.0.
The fast decay after prolonged high LR during
hold phase allows for rapid convergence.
References:
- [Squeezeformer:
An Efficient Transformer for Automatic Speech Recognition]
(https://arxiv.org/abs/2206.00888)
Args:
optimizer: Pytorch compatible Optimizer object.
warmup_steps: Number of training steps in warmup stage
warmup_ratio: Ratio of warmup steps to total steps
hold_steps: Number of training steps to
hold the learning rate after warm up
hold_ratio: Ratio of hold steps to total steps
max_steps: Total number of steps while training or `None` for
infinite training
decay_rate: Float value describing the polynomial decay
after the hold period. Default value
of 0.5 corresponds to Noam decay.
min_lr: Minimum learning rate.
"""
self.decay_rate = decay_rate
super().__init__(
optimizer=optimizer,
max_steps=max_steps,
last_epoch=last_epoch,
min_lr=min_lr,
**kwargs,
)
def _get_lr(self, step):
if self.warmup_steps is None or self.warmup_steps == 0:
raise ValueError("Noam scheduler cannot be used without warmup steps")
if self.hold_steps > 0:
hold_steps = self.hold_steps - self.warmup_steps
else:
hold_steps = 0
new_lrs = [
_noam_hold_annealing(
initial_lr,
step=step,
warmup_steps=self.warmup_steps,
hold_steps=hold_steps,
decay_rate=self.decay_rate,
min_lr=self.min_lr,
)
for initial_lr in self.base_lrs
]
return new_lrs
def set_step(self, step: int):
self.last_epoch = step
|