Spaces:
Runtime error
Runtime error
File size: 8,911 Bytes
fb53ec8 |
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 |
import torch
import torch.optim as optim
import numpy as np
import copy
from ... import sync
from ...cfg_holder import cfg_unique_holder as cfguh
def singleton(class_):
instances = {}
def getinstance(*args, **kwargs):
if class_ not in instances:
instances[class_] = class_(*args, **kwargs)
return instances[class_]
return getinstance
@singleton
class get_scheduler(object):
def __init__(self):
self.lr_scheduler = {}
def register(self, lrsf, name):
self.lr_scheduler[name] = lrsf
def __call__(self, cfg):
if cfg is None:
return None
if isinstance(cfg, list):
schedulers = []
for ci in cfg:
t = ci.type
schedulers.append(
self.lr_scheduler[t](**ci.args))
if len(schedulers) == 0:
raise ValueError
else:
return compose_scheduler(schedulers)
t = cfg.type
return self.lr_scheduler[t](**cfg.args)
def register(name):
def wrapper(class_):
get_scheduler().register(class_, name)
return class_
return wrapper
class template_scheduler(object):
def __init__(self, step):
self.step = step
def __getitem__(self, idx):
raise ValueError
def set_lr(self, optim, new_lr, pg_lrscale=None):
"""
Set Each parameter_groups in optim with new_lr
New_lr can be find according to the idx.
pg_lrscale tells how to scale each pg.
"""
# new_lr = self.__getitem__(idx)
pg_lrscale = copy.deepcopy(pg_lrscale)
for pg in optim.param_groups:
if pg_lrscale is None:
pg['lr'] = new_lr
else:
pg['lr'] = new_lr * pg_lrscale.pop(pg['name'])
assert (pg_lrscale is None) or (len(pg_lrscale)==0), \
"pg_lrscale doesn't match pg"
@register('constant')
class constant_scheduler(template_scheduler):
def __init__(self, lr, step):
super().__init__(step)
self.lr = lr
def __getitem__(self, idx):
if idx >= self.step:
raise ValueError
return self.lr
@register('poly')
class poly_scheduler(template_scheduler):
def __init__(self, start_lr, end_lr, power, step):
super().__init__(step)
self.start_lr = start_lr
self.end_lr = end_lr
self.power = power
def __getitem__(self, idx):
if idx >= self.step:
raise ValueError
a, b = self.start_lr, self.end_lr
p, n = self.power, self.step
return b + (a-b)*((1-idx/n)**p)
@register('linear')
class linear_scheduler(template_scheduler):
def __init__(self, start_lr, end_lr, step):
super().__init__(step)
self.start_lr = start_lr
self.end_lr = end_lr
def __getitem__(self, idx):
if idx >= self.step:
raise ValueError
a, b, n = self.start_lr, self.end_lr, self.step
return b + (a-b)*(1-idx/n)
@register('multistage')
class constant_scheduler(template_scheduler):
def __init__(self, start_lr, milestones, gamma, step):
super().__init__(step)
self.start_lr = start_lr
m = [0] + milestones + [step]
lr_iter = start_lr
self.lr = []
for ms, me in zip(m[0:-1], m[1:]):
for _ in range(ms, me):
self.lr.append(lr_iter)
lr_iter *= gamma
def __getitem__(self, idx):
if idx >= self.step:
raise ValueError
return self.lr[idx]
class compose_scheduler(template_scheduler):
def __init__(self, schedulers):
self.schedulers = schedulers
self.step = [si.step for si in schedulers]
self.step_milestone = []
acc = 0
for i in self.step:
acc += i
self.step_milestone.append(acc)
self.step = sum(self.step)
def __getitem__(self, idx):
if idx >= self.step:
raise ValueError
ms = self.step_milestone
for idx, (mi, mj) in enumerate(zip(ms[:-1], ms[1:])):
if mi <= idx < mj:
return self.schedulers[idx-mi]
raise ValueError
####################
# lambda schedular #
####################
class LambdaWarmUpCosineScheduler(template_scheduler):
"""
note: use with a base_lr of 1.0
"""
def __init__(self,
base_lr,
warm_up_steps,
lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0):
cfgt = cfguh().cfg.train
bs = cfgt.batch_size
if 'gradacc_every' not in cfgt:
print('Warning, gradacc_every is not found in xml, use 1 as default.')
acc = cfgt.get('gradacc_every', 1)
self.lr_multi = base_lr * bs * acc
self.lr_warm_up_steps = warm_up_steps
self.lr_start = lr_start
self.lr_min = lr_min
self.lr_max = lr_max
self.lr_max_decay_steps = max_decay_steps
self.last_lr = 0.
self.verbosity_interval = verbosity_interval
def schedule(self, n):
if self.verbosity_interval > 0:
if n % self.verbosity_interval == 0:
print(f"current step: {n}, recent lr-multiplier: {self.last_lr}")
if n < self.lr_warm_up_steps:
lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start
self.last_lr = lr
return lr
else:
t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps)
t = min(t, 1.0)
lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * (
1 + np.cos(t * np.pi))
self.last_lr = lr
return lr
def __getitem__(self, idx):
return self.schedule(idx) * self.lr_multi
class LambdaWarmUpCosineScheduler2(template_scheduler):
"""
supports repeated iterations, configurable via lists
note: use with a base_lr of 1.0.
"""
def __init__(self,
base_lr,
warm_up_steps,
f_min, f_max, f_start, cycle_lengths, verbosity_interval=0):
cfgt = cfguh().cfg.train
# bs = cfgt.batch_size
# if 'gradacc_every' not in cfgt:
# print('Warning, gradacc_every is not found in xml, use 1 as default.')
# acc = cfgt.get('gradacc_every', 1)
# self.lr_multi = base_lr * bs * acc
self.lr_multi = base_lr
assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths)
self.lr_warm_up_steps = warm_up_steps
self.f_start = f_start
self.f_min = f_min
self.f_max = f_max
self.cycle_lengths = cycle_lengths
self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths))
self.last_f = 0.
self.verbosity_interval = verbosity_interval
def find_in_interval(self, n):
interval = 0
for cl in self.cum_cycles[1:]:
if n <= cl:
return interval
interval += 1
def schedule(self, n):
cycle = self.find_in_interval(n)
n = n - self.cum_cycles[cycle]
if self.verbosity_interval > 0:
if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
f"current cycle {cycle}")
if n < self.lr_warm_up_steps[cycle]:
f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
self.last_f = f
return f
else:
t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle])
t = min(t, 1.0)
f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * (
1 + np.cos(t * np.pi))
self.last_f = f
return f
def __getitem__(self, idx):
return self.schedule(idx) * self.lr_multi
@register('stable_diffusion_linear')
class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2):
def schedule(self, n):
cycle = self.find_in_interval(n)
n = n - self.cum_cycles[cycle]
if self.verbosity_interval > 0:
if n % self.verbosity_interval == 0:
print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
f"current cycle {cycle}")
if n < self.lr_warm_up_steps[cycle]:
f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
self.last_f = f
return f
else:
f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle])
self.last_f = f
return f |