Spaces:
Running
Running
File size: 7,439 Bytes
29f689c |
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 |
import math
from functools import partial
import numpy as np
from torch.optim import lr_scheduler
class StepLR(object):
def __init__(self,
step_each_epoch,
step_size,
warmup_epoch=0,
gamma=0.1,
last_epoch=-1,
**kwargs):
super(StepLR, self).__init__()
self.step_size = step_each_epoch * step_size
self.gamma = gamma
self.last_epoch = last_epoch
self.warmup_epoch = warmup_epoch
def __call__(self, optimizer):
return lr_scheduler.LambdaLR(optimizer, self.lambda_func,
self.last_epoch)
def lambda_func(self, current_step):
if current_step < self.warmup_epoch:
return float(current_step) / float(max(1, self.warmup_epoch))
return self.gamma**(current_step // self.step_size)
class MultiStepLR(object):
def __init__(self,
step_each_epoch,
milestones,
warmup_epoch=0,
gamma=0.1,
last_epoch=-1,
**kwargs):
super(MultiStepLR, self).__init__()
self.milestones = [step_each_epoch * e for e in milestones]
self.gamma = gamma
self.last_epoch = last_epoch
self.warmup_epoch = warmup_epoch
def __call__(self, optimizer):
return lr_scheduler.LambdaLR(optimizer, self.lambda_func,
self.last_epoch)
def lambda_func(self, current_step):
if current_step < self.warmup_epoch:
return float(current_step) / float(max(1, self.warmup_epoch))
return self.gamma**len(
[m for m in self.milestones if m <= current_step])
class ConstLR(object):
def __init__(self,
step_each_epoch,
warmup_epoch=0,
last_epoch=-1,
**kwargs):
super(ConstLR, self).__init__()
self.last_epoch = last_epoch
self.warmup_epoch = warmup_epoch * step_each_epoch
def __call__(self, optimizer):
return lr_scheduler.LambdaLR(optimizer, self.lambda_func,
self.last_epoch)
def lambda_func(self, current_step):
if current_step < self.warmup_epoch:
return float(current_step) / float(max(1.0, self.warmup_epoch))
return 1.0
class LinearLR(object):
def __init__(self,
epochs,
step_each_epoch,
warmup_epoch=0,
last_epoch=-1,
**kwargs):
super(LinearLR, self).__init__()
self.epochs = epochs * step_each_epoch
self.last_epoch = last_epoch
self.warmup_epoch = warmup_epoch * step_each_epoch
def __call__(self, optimizer):
return lr_scheduler.LambdaLR(optimizer, self.lambda_func,
self.last_epoch)
def lambda_func(self, current_step):
if current_step < self.warmup_epoch:
return float(current_step) / float(max(1, self.warmup_epoch))
return max(
0.0,
float(self.epochs - current_step) /
float(max(1, self.epochs - self.warmup_epoch)),
)
class CosineAnnealingLR(object):
def __init__(self,
epochs,
step_each_epoch,
warmup_epoch=0,
last_epoch=-1,
**kwargs):
super(CosineAnnealingLR, self).__init__()
self.epochs = epochs * step_each_epoch
self.last_epoch = last_epoch
self.warmup_epoch = warmup_epoch * step_each_epoch
def __call__(self, optimizer):
return lr_scheduler.LambdaLR(optimizer, self.lambda_func,
self.last_epoch)
def lambda_func(self, current_step, num_cycles=0.5):
if current_step < self.warmup_epoch:
return float(current_step) / float(max(1, self.warmup_epoch))
progress = float(current_step - self.warmup_epoch) / float(
max(1, self.epochs - self.warmup_epoch))
return max(
0.0, 0.5 *
(1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress)))
class OneCycleLR(object):
def __init__(self,
epochs,
step_each_epoch,
last_epoch=-1,
lr=0.00148,
warmup_epoch=1.0,
cycle_momentum=True,
**kwargs):
super(OneCycleLR, self).__init__()
self.epochs = epochs
self.last_epoch = last_epoch
self.step_each_epoch = step_each_epoch
self.lr = lr
self.pct_start = warmup_epoch / epochs
self.cycle_momentum = cycle_momentum
def __call__(self, optimizer):
return lr_scheduler.OneCycleLR(
optimizer,
max_lr=self.lr,
total_steps=self.epochs * self.step_each_epoch,
pct_start=self.pct_start,
cycle_momentum=self.cycle_momentum,
)
class PolynomialLR(object):
def __init__(self,
step_each_epoch,
epochs,
lr_end=1e-7,
power=1.0,
warmup_epoch=0,
last_epoch=-1,
**kwargs):
super(PolynomialLR, self).__init__()
self.lr_end = lr_end
self.power = power
self.epochs = epochs * step_each_epoch
self.warmup_epoch = warmup_epoch * step_each_epoch
self.last_epoch = last_epoch
def __call__(self, optimizer):
lr_lambda = partial(
self.lambda_func,
lr_init=optimizer.defaults['lr'],
)
return lr_scheduler.LambdaLR(optimizer, lr_lambda, self.last_epoch)
def lambda_func(self, current_step, lr_init):
if current_step < self.warmup_epoch:
return float(current_step) / float(max(1, self.warmup_epoch))
elif current_step > self.epochs:
return self.lr_end / lr_init # as LambdaLR multiplies by lr_init
else:
lr_range = lr_init - self.lr_end
decay_steps = self.epochs - self.warmup_epoch
pct_remaining = 1 - (current_step -
self.warmup_epoch) / decay_steps
decay = lr_range * pct_remaining**self.power + self.lr_end
return decay / lr_init # as LambdaLR multiplies by lr_init
class CdistNetLR(object):
def __init__(self,
step_each_epoch,
lr=0.0442,
n_warmup_steps=10000,
step2_epoch=7,
last_epoch=-1,
**kwargs):
super(CdistNetLR, self).__init__()
self.last_epoch = last_epoch
self.step2_epoch = step2_epoch * step_each_epoch
self.n_current_steps = 0
self.n_warmup_steps = n_warmup_steps
self.init_lr = lr
self.step2_lr = 0.00001
def __call__(self, optimizer):
return lr_scheduler.LambdaLR(optimizer, self.lambda_func,
self.last_epoch)
def lambda_func(self, current_step):
if current_step < self.step2_epoch:
return np.min([
np.power(current_step, -0.5),
np.power(self.n_warmup_steps, -1.5) * current_step,
])
return self.step2_lr / self.init_lr
|