forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtraining_utils.py
More file actions
324 lines (269 loc) 路 13.1 KB
/
Copy pathtraining_utils.py
File metadata and controls
324 lines (269 loc) 路 13.1 KB
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
import copy
import os
import random
from typing import Any, Dict, Iterable, Optional, Union
import numpy as np
import torch
from .utils import deprecate
def enable_full_determinism(seed: int):
"""
Helper function for reproducible behavior during distributed training. See
- https://pytorch.org/docs/stable/notes/randomness.html for pytorch
"""
# set seed first
set_seed(seed)
# Enable PyTorch deterministic mode. This potentially requires either the environment
# variable 'CUDA_LAUNCH_BLOCKING' or 'CUBLAS_WORKSPACE_CONFIG' to be set,
# depending on the CUDA version, so we set them both here
os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":16:8"
torch.use_deterministic_algorithms(True)
# Enable CUDNN deterministic mode
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def set_seed(seed: int):
"""
Args:
Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`.
seed (`int`): The seed to set.
"""
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# ^^ safe to call this function even if cuda is not available
# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
class EMAModel:
"""
Exponential Moving Average of models weights
"""
def __init__(
self,
parameters: Iterable[torch.nn.Parameter],
decay: float = 0.9999,
min_decay: float = 0.0,
update_after_step: int = 0,
use_ema_warmup: bool = False,
inv_gamma: Union[float, int] = 1.0,
power: Union[float, int] = 2 / 3,
model_cls: Optional[Any] = None,
model_config: Dict[str, Any] = None,
**kwargs,
):
"""
Args:
parameters (Iterable[torch.nn.Parameter]): The parameters to track.
decay (float): The decay factor for the exponential moving average.
min_decay (float): The minimum decay factor for the exponential moving average.
update_after_step (int): The number of steps to wait before starting to update the EMA weights.
use_ema_warmup (bool): Whether to use EMA warmup.
inv_gamma (float):
Inverse multiplicative factor of EMA warmup. Default: 1. Only used if `use_ema_warmup` is True.
power (float): Exponential factor of EMA warmup. Default: 2/3. Only used if `use_ema_warmup` is True.
device (Optional[Union[str, torch.device]]): The device to store the EMA weights on. If None, the EMA
weights will be stored on CPU.
@crowsonkb's notes on EMA Warmup:
If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are good values for models you plan
to train for a million or more steps (reaches decay factor 0.999 at 31.6K steps, 0.9999 at 1M steps),
gamma=1, power=3/4 for models you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999
at 215.4k steps).
"""
if isinstance(parameters, torch.nn.Module):
deprecation_message = (
"Passing a `torch.nn.Module` to `ExponentialMovingAverage` is deprecated. "
"Please pass the parameters of the module instead."
)
deprecate(
"passing a `torch.nn.Module` to `ExponentialMovingAverage`",
"1.0.0",
deprecation_message,
standard_warn=False,
)
parameters = parameters.parameters()
# set use_ema_warmup to True if a torch.nn.Module is passed for backwards compatibility
use_ema_warmup = True
if kwargs.get("max_value", None) is not None:
deprecation_message = "The `max_value` argument is deprecated. Please use `decay` instead."
deprecate("max_value", "1.0.0", deprecation_message, standard_warn=False)
decay = kwargs["max_value"]
if kwargs.get("min_value", None) is not None:
deprecation_message = "The `min_value` argument is deprecated. Please use `min_decay` instead."
deprecate("min_value", "1.0.0", deprecation_message, standard_warn=False)
min_decay = kwargs["min_value"]
parameters = list(parameters)
self.shadow_params = [p.clone().detach() for p in parameters]
if kwargs.get("device", None) is not None:
deprecation_message = "The `device` argument is deprecated. Please use `to` instead."
deprecate("device", "1.0.0", deprecation_message, standard_warn=False)
self.to(device=kwargs["device"])
self.temp_stored_params = None
self.decay = decay
self.min_decay = min_decay
self.update_after_step = update_after_step
self.use_ema_warmup = use_ema_warmup
self.inv_gamma = inv_gamma
self.power = power
self.optimization_step = 0
self.cur_decay_value = None # set in `step()`
self.model_cls = model_cls
self.model_config = model_config
@classmethod
def from_pretrained(cls, path, model_cls) -> "EMAModel":
_, ema_kwargs = model_cls.load_config(path, return_unused_kwargs=True)
model = model_cls.from_pretrained(path)
ema_model = cls(model.parameters(), model_cls=model_cls, model_config=model.config)
ema_model.load_state_dict(ema_kwargs)
return ema_model
def save_pretrained(self, path):
if self.model_cls is None:
raise ValueError("`save_pretrained` can only be used if `model_cls` was defined at __init__.")
if self.model_config is None:
raise ValueError("`save_pretrained` can only be used if `model_config` was defined at __init__.")
model = self.model_cls.from_config(self.model_config)
state_dict = self.state_dict()
state_dict.pop("shadow_params", None)
model.register_to_config(**state_dict)
self.copy_to(model.parameters())
model.save_pretrained(path)
def get_decay(self, optimization_step: int) -> float:
"""
Compute the decay factor for the exponential moving average.
"""
step = max(0, optimization_step - self.update_after_step - 1)
if step <= 0:
return 0.0
if self.use_ema_warmup:
cur_decay_value = 1 - (1 + step / self.inv_gamma) ** -self.power
else:
cur_decay_value = (1 + step) / (10 + step)
cur_decay_value = min(cur_decay_value, self.decay)
# make sure decay is not smaller than min_decay
cur_decay_value = max(cur_decay_value, self.min_decay)
return cur_decay_value
@torch.no_grad()
def step(self, parameters: Iterable[torch.nn.Parameter]):
if isinstance(parameters, torch.nn.Module):
deprecation_message = (
"Passing a `torch.nn.Module` to `ExponentialMovingAverage.step` is deprecated. "
"Please pass the parameters of the module instead."
)
deprecate(
"passing a `torch.nn.Module` to `ExponentialMovingAverage.step`",
"1.0.0",
deprecation_message,
standard_warn=False,
)
parameters = parameters.parameters()
parameters = list(parameters)
self.optimization_step += 1
# Compute the decay factor for the exponential moving average.
decay = self.get_decay(self.optimization_step)
self.cur_decay_value = decay
one_minus_decay = 1 - decay
for s_param, param in zip(self.shadow_params, parameters):
if param.requires_grad:
s_param.sub_(one_minus_decay * (s_param - param))
else:
s_param.copy_(param)
torch.cuda.empty_cache()
def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None:
"""
Copy current averaged parameters into given collection of parameters.
Args:
parameters: Iterable of `torch.nn.Parameter`; the parameters to be
updated with the stored moving averages. If `None`, the parameters with which this
`ExponentialMovingAverage` was initialized will be used.
"""
parameters = list(parameters)
for s_param, param in zip(self.shadow_params, parameters):
param.data.copy_(s_param.to(param.device).data)
def to(self, device=None, dtype=None) -> None:
r"""Move internal buffers of the ExponentialMovingAverage to `device`.
Args:
device: like `device` argument to `torch.Tensor.to`
"""
# .to() on the tensors handles None correctly
self.shadow_params = [
p.to(device=device, dtype=dtype) if p.is_floating_point() else p.to(device=device)
for p in self.shadow_params
]
def state_dict(self) -> dict:
r"""
Returns the state of the ExponentialMovingAverage as a dict. This method is used by accelerate during
checkpointing to save the ema state dict.
"""
# Following PyTorch conventions, references to tensors are returned:
# "returns a reference to the state and not its copy!" -
# https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict
return {
"decay": self.decay,
"min_decay": self.min_decay,
"optimization_step": self.optimization_step,
"update_after_step": self.update_after_step,
"use_ema_warmup": self.use_ema_warmup,
"inv_gamma": self.inv_gamma,
"power": self.power,
"shadow_params": self.shadow_params,
}
def store(self, parameters: Iterable[torch.nn.Parameter]) -> None:
r"""
Args:
Save the current parameters for restoring later.
parameters: Iterable of `torch.nn.Parameter`; the parameters to be
temporarily stored.
"""
self.temp_stored_params = [param.detach().cpu().clone() for param in parameters]
def restore(self, parameters: Iterable[torch.nn.Parameter]) -> None:
r"""
Args:
Restore the parameters stored with the `store` method. Useful to validate the model with EMA parameters without:
affecting the original optimization process. Store the parameters before the `copy_to()` method. After
validation (or model saving), use this to restore the former parameters.
parameters: Iterable of `torch.nn.Parameter`; the parameters to be
updated with the stored parameters. If `None`, the parameters with which this
`ExponentialMovingAverage` was initialized will be used.
"""
if self.temp_stored_params is None:
raise RuntimeError("This ExponentialMovingAverage has no `store()`ed weights " "to `restore()`")
for c_param, param in zip(self.temp_stored_params, parameters):
param.data.copy_(c_param.data)
# Better memory-wise.
self.temp_stored_params = None
def load_state_dict(self, state_dict: dict) -> None:
r"""
Args:
Loads the ExponentialMovingAverage state. This method is used by accelerate during checkpointing to save the
ema state dict.
state_dict (dict): EMA state. Should be an object returned
from a call to :meth:`state_dict`.
"""
# deepcopy, to be consistent with module API
state_dict = copy.deepcopy(state_dict)
self.decay = state_dict.get("decay", self.decay)
if self.decay < 0.0 or self.decay > 1.0:
raise ValueError("Decay must be between 0 and 1")
self.min_decay = state_dict.get("min_decay", self.min_decay)
if not isinstance(self.min_decay, float):
raise ValueError("Invalid min_decay")
self.optimization_step = state_dict.get("optimization_step", self.optimization_step)
if not isinstance(self.optimization_step, int):
raise ValueError("Invalid optimization_step")
self.update_after_step = state_dict.get("update_after_step", self.update_after_step)
if not isinstance(self.update_after_step, int):
raise ValueError("Invalid update_after_step")
self.use_ema_warmup = state_dict.get("use_ema_warmup", self.use_ema_warmup)
if not isinstance(self.use_ema_warmup, bool):
raise ValueError("Invalid use_ema_warmup")
self.inv_gamma = state_dict.get("inv_gamma", self.inv_gamma)
if not isinstance(self.inv_gamma, (float, int)):
raise ValueError("Invalid inv_gamma")
self.power = state_dict.get("power", self.power)
if not isinstance(self.power, (float, int)):
raise ValueError("Invalid power")
shadow_params = state_dict.get("shadow_params", None)
if shadow_params is not None:
self.shadow_params = shadow_params
if not isinstance(self.shadow_params, list):
raise ValueError("shadow_params must be a list")
if not all(isinstance(p, torch.Tensor) for p in self.shadow_params):
raise ValueError("shadow_params must all be Tensors")