Spaces:
Running
Running
File size: 3,516 Bytes
85ab89d |
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 |
"""
Copyright (c) 2022, salesforce.com, inc.
All rights reserved.
SPDX-License-Identifier: BSD-3-Clause
For full license text, see the LICENSE_Lavis file in the repo root or https://opensource.org/licenses/BSD-3-Clause
"""
import math
from minigpt4.common.registry import registry
@registry.register_lr_scheduler("linear_warmup_step_lr")
class LinearWarmupStepLRScheduler:
def __init__(
self,
optimizer,
max_epoch,
min_lr,
init_lr,
decay_rate=1,
warmup_start_lr=-1,
warmup_steps=0,
**kwargs
):
self.optimizer = optimizer
self.max_epoch = max_epoch
self.min_lr = min_lr
self.decay_rate = decay_rate
self.init_lr = init_lr
self.warmup_steps = warmup_steps
self.warmup_start_lr = warmup_start_lr if warmup_start_lr >= 0 else init_lr
def step(self, cur_epoch, cur_step):
if cur_epoch == 0:
warmup_lr_schedule(
step=cur_step,
optimizer=self.optimizer,
max_step=self.warmup_steps,
init_lr=self.warmup_start_lr,
max_lr=self.init_lr,
)
else:
step_lr_schedule(
epoch=cur_epoch,
optimizer=self.optimizer,
init_lr=self.init_lr,
min_lr=self.min_lr,
decay_rate=self.decay_rate,
)
@registry.register_lr_scheduler("linear_warmup_cosine_lr")
class LinearWarmupCosineLRScheduler:
def __init__(
self,
optimizer,
max_epoch,
iters_per_epoch,
min_lr,
init_lr,
warmup_steps=0,
warmup_start_lr=-1,
**kwargs
):
self.optimizer = optimizer
self.max_epoch = max_epoch
self.iters_per_epoch = iters_per_epoch
self.min_lr = min_lr
self.init_lr = init_lr
self.warmup_steps = warmup_steps
self.warmup_start_lr = warmup_start_lr if warmup_start_lr >= 0 else init_lr
def step(self, cur_epoch, cur_step):
total_cur_step = cur_epoch * self.iters_per_epoch + cur_step
if total_cur_step < self.warmup_steps:
warmup_lr_schedule(
step=cur_step,
optimizer=self.optimizer,
max_step=self.warmup_steps,
init_lr=self.warmup_start_lr,
max_lr=self.init_lr,
)
else:
cosine_lr_schedule(
epoch=total_cur_step,
optimizer=self.optimizer,
max_epoch=self.max_epoch * self.iters_per_epoch,
init_lr=self.init_lr,
min_lr=self.min_lr,
)
def cosine_lr_schedule(optimizer, epoch, max_epoch, init_lr, min_lr):
"""Decay the learning rate"""
lr = (init_lr - min_lr) * 0.5 * (
1.0 + math.cos(math.pi * epoch / max_epoch)
) + min_lr
for param_group in optimizer.param_groups:
param_group["lr"] = lr
def warmup_lr_schedule(optimizer, step, max_step, init_lr, max_lr):
"""Warmup the learning rate"""
lr = min(max_lr, init_lr + (max_lr - init_lr) * step / max(max_step, 1))
for param_group in optimizer.param_groups:
param_group["lr"] = lr
def step_lr_schedule(optimizer, epoch, init_lr, min_lr, decay_rate):
"""Decay the learning rate"""
lr = max(min_lr, init_lr * (decay_rate**epoch))
for param_group in optimizer.param_groups:
param_group["lr"] = lr
|