File size: 9,611 Bytes
d04a061 |
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 |
import torch
import math
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LambdaLR
from torch.optim.adamw import adamw
try:
import deepspeed
from deepspeed.ops.adam import FusedAdam
from deepspeed.ops.adam import DeepSpeedCPUAdam
except:
pass
def get_optimizer(cfg, params):
if cfg.optim.type == 'adam':
return torch.optim.Adam(
params=params,
lr=cfg.optim.lr,
weight_decay=cfg.optim.weight_decay,
betas=(cfg.optim.beta1, cfg.optim.beta2)
)
elif cfg.optim.type == 'adamw':
return AdamW(
params=params,
lr=cfg.optim.lr,
weight_decay=cfg.optim.weight_decay,
betas=(cfg.optim.beta1, cfg.optim.beta2)
)
elif cfg.type == 'fusedadam':
return FusedAdam(
params=params,
lr=cfg.lr,
weight_decay=cfg.weight_decay,
betas=cfg.betas,
)
else:
raise NotImplementedError('Optimizer not supported: %s' % cfg.type)
class AdamW(torch.optim.AdamW):
@torch.no_grad()
def step(self, closure=None):
"""Performs a single optimization step.
Args:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
self._cuda_graph_capture_health_check()
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
params_with_grad = []
grads = []
exp_avgs = []
exp_avg_sqs = []
max_exp_avg_sqs = []
state_steps = []
amsgrad = group['amsgrad']
beta1, beta2 = group['betas']
for p in group['params']:
if p.grad is None:
continue
params_with_grad.append(p)
if p.grad.is_sparse:
raise RuntimeError('AdamW does not support sparse gradients')
grads.append(p.grad)
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = torch.zeros((1,), dtype=torch.float, device=p.device) \
if self.defaults['capturable'] else torch.tensor(0.)
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
if amsgrad:
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
exp_avgs.append(state['exp_avg'])
exp_avg_sqs.append(state['exp_avg_sq'])
if amsgrad:
max_exp_avg_sqs.append(state['max_exp_avg_sq'])
state_steps.append(state['step'].cpu())
adamw(params_with_grad,
grads,
exp_avgs,
exp_avg_sqs,
max_exp_avg_sqs,
state_steps,
amsgrad=amsgrad,
beta1=beta1,
beta2=beta2,
lr=group['lr'],
weight_decay=group['weight_decay'],
eps=group['eps'],
maximize=group['maximize'],
foreach=group['foreach'],
capturable=group['capturable'])
return loss
def get_scheduler(cfg, optimizer):
if cfg.optim.scheduler is None:
return BlackHole()
elif cfg.optim.scheduler == 'plateau':
return (
torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
mode=cfg.mode,
factor=cfg.factor,
patience=cfg.patience,
min_lr=cfg.min_lr,
),
{'monitor': "val/loss", 'interval': 'epoch'}
)
elif cfg.optim.scheduler == 'noam':
return (
NoamScheduler(
optimizer,
lr=cfg.lr,
warmup_steps=cfg.warmup_steps,
model_size=cfg.model_size,
warmup_init_lr=cfg.get('warmup_init_lr')
),
{'frequency': 1, 'interval': 'step'}
)
elif cfg.optim.scheduler == 'polynomial':
return (
PolyNomialLRScheduler(
optimizer,
total_steps=cfg.training.max_steps,
warmup_steps=cfg.training.warmup_steps,
lr=cfg.optim.lr,
lr_end=cfg.optim.lr_end,
warmup_init_lr=cfg.optim.warmup_init_lr,
power=cfg.optim.power
),
{'frequency': 1, 'interval': 'step'}
)
elif cfg.optim.scheduler == 'multistep':
return torch.optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=cfg.milestones,
gamma=cfg.gamma,
)
elif cfg.optim.scheduler == 'exp':
return torch.optim.lr_scheduler.ExponentialLR(
optimizer,
gamma=cfg.gamma,
)
elif cfg.optim.scheduler == 'progen_ft':
sched = CosineToFrac(
optimizer=optimizer,
total_steps=cfg.training.max_steps,
final_frac=0.2, # decay to lr/5
)
return (sched, {'frequency': 1, 'interval': 'step'})
elif cfg.optim.scheduler is None:
return BlackHole()
else:
raise NotImplementedError('Scheduler not supported: %s' % cfg.optim.scheduler)
class BlackHole(object):
def __setattr__(self, name, value):
pass
def __call__(self, *args, **kwargs):
return self
def __getattr__(self, name):
return self
# -------# DPLM Scheduler #-------- #
def polynomial_lr_schedule(step, total_steps, warmup_steps, warmup_init_lr, lr, lr_end, power):
if step < warmup_steps:
return warmup_init_lr + (lr - warmup_init_lr) * step / warmup_steps
elif step > total_steps:
return lr_end
else:
return lr_end + (lr - lr_end) * (1 - (step - warmup_steps) / (total_steps - warmup_steps)) ** power
class PolyNomialLRScheduler(LambdaLR):
def __init__(
self,
optimizer: Optimizer,
total_steps: int = 1000,
warmup_steps: int = 0,
lr: float = 0.00004, # 5e-04,
lr_end: float = 1e-5, #1e-07,
warmup_init_lr: float = 1e-07, # 1e-07,
power: float = 1.0,
) -> None:
self.warmup_init_lr = warmup_init_lr
self.warmup_steps = warmup_steps
def lr_lambda(step):
return polynomial_lr_schedule(
step, total_steps, warmup_steps, warmup_init_lr, lr, lr_end, power
) / lr
super().__init__(optimizer, lr_lambda=lr_lambda)
# -------# ProGen2 Fine-Tuning Scheduler #-------- #
def cosine_frac_scheduler(step, total_steps, final_frac):
s = min(max(step, 0), total_steps)
cos = 0.5 * (1.0 + math.cos(math.pi * s / total_steps)) # 1 --> 0
return final_frac + (1.0 - final_frac) * cos # multiplier goes from 1.0 down to final_frac
class CosineToFrac(LambdaLR):
"""
Cosine decay of the LR multiplier from 1.0 -> final_frac over total_steps (no warmup).
For ProGen fine-tuning, final_frac=0.2 implements decay to lr/5.
"""
def __init__(self, optimizer, total_steps, final_frac=0.2):
self.total_steps = max(int(total_steps), 1)
self.final_frac = float(final_frac)
def lr_lambda(step):
return cosine_frac_scheduler(
step=step,
total_steps=self.total_steps,
final_frac=self.final_frac
)
super().__init__(optimizer, lr_lambda=lr_lambda)
def inverse_sqrt_lr_schedule(step, warmup_steps, warmup_init_lr, lr_step, decay_step):
if step == 0:
step = 1
if step < warmup_steps:
return warmup_init_lr + lr_step * step
else:
return decay_step * step ** -0.5
class InverseSqrtLRScheduler(LambdaLR):
def __init__(
self,
optimizer: Optimizer,
warmup_steps: int = 0,
lr: float = 5e-04,
warmup_init_lr: float = 1e-07,
) -> None:
self.warmup_init_lr = warmup_init_lr
self.warmup_steps = warmup_steps
self.lr_step = (lr - warmup_init_lr) / warmup_steps
self.decay_step = lr * warmup_steps ** 0.5
def lr_lambda(step):
return inverse_sqrt_lr_schedule(
step, warmup_steps, warmup_init_lr, self.lr_step, self.decay_step
) / lr
super().__init__(optimizer, lr_lambda=lr_lambda)
def noam_lr_schedule(step, warmup_steps, factor, model_size):
if step == 0:
step = 1
return factor * (model_size ** (-0.5) * min(step ** (-0.5), step * warmup_steps ** (-1.5)))
class NoamScheduler(LambdaLR):
def __init__(
self,
optimizer: Optimizer,
lr,
warmup_init_lr,
model_size: int = 128,
warmup_steps: int = 0,
factor: int = 2,
) -> None:
# dummy_lr = self.base_lrs[0]
def lr_lambda(step):
return noam_lr_schedule(step, warmup_steps, factor, model_size) / lr
super().__init__(optimizer, lr_lambda=lr_lambda)
|