forked from Nerogar/OneTrainer
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathBaseModelSetup.py
More file actions
166 lines (146 loc) · 4.48 KB
/
Copy pathBaseModelSetup.py
File metadata and controls
166 lines (146 loc) · 4.48 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
from abc import ABCMeta, abstractmethod
from typing import Iterable, Iterator
import torch
from torch import Tensor
from torch.nn import Parameter
from modules.model.BaseModel import BaseModel
from modules.util.TimedActionMixin import TimedActionMixin
from modules.util.TrainProgress import TrainProgress
from modules.util.config.TrainConfig import TrainConfig
from modules.util.enum.LearningRateScaler import LearningRateScaler
class BaseModelSetup(
TimedActionMixin,
metaclass=ABCMeta,
):
def __init__(
self,
train_device: torch.device,
temp_device: torch.device,
debug_mode: bool,
):
super(BaseModelSetup, self).__init__()
self.train_device = train_device
self.temp_device = temp_device
self.debug_mode = debug_mode
@abstractmethod
def create_parameters(
self,
model: BaseModel,
config: TrainConfig,
) -> Iterable[Parameter]:
pass
def create_parameters_for_optimizer(
self,
model: BaseModel,
config: TrainConfig,
) -> Iterable[Parameter] | list[dict]:
return self.create_parameters(model, config)
@abstractmethod
def setup_model(
self,
model: BaseModel,
config: TrainConfig,
):
pass
@abstractmethod
def setup_train_device(
self,
model: BaseModel,
config: TrainConfig,
):
pass
@abstractmethod
def predict(
self,
model: BaseModel,
batch: dict,
config: TrainConfig,
train_progress: TrainProgress,
*,
deterministic: bool = False,
) -> dict:
pass
@abstractmethod
def calculate_loss(
self,
model: BaseModel,
batch: dict,
data: dict,
config: TrainConfig,
) -> Tensor:
pass
@abstractmethod
def after_optimizer_step(
self,
model: BaseModel,
config: TrainConfig,
train_progress: TrainProgress,
):
pass
def create_param_groups(
self,
config: TrainConfig,
params: Iterator[Parameter] | list[Parameter],
lr_arg: float,
) -> dict:
batch_size_scale = 1 if config.learning_rate_scaler in [
LearningRateScaler.NONE,
LearningRateScaler.GRADIENT_ACCUMULATION,
] else config.batch_size
gradient_accumulation_steps_scale = 1 if config.learning_rate_scaler in [
LearningRateScaler.NONE,
LearningRateScaler.BATCH,
] else config.gradient_accumulation_steps
# Determine the learning rate
lr = lr_arg if lr_arg is not None else config.learning_rate
lr = lr * ((batch_size_scale * gradient_accumulation_steps_scale) ** 0.5)
# Create a parameter group for the text encoder
return {
'params': list(params),
'lr': lr,
'initial_lr': lr,
}
def stop_unet_training_elapsed(
self,
config: TrainConfig,
train_progress: TrainProgress,
):
return self.single_action_elapsed(
"stop_unet_training",
config.unet.stop_training_after,
config.unet.stop_training_after_unit,
train_progress,
)
def stop_prior_training_elapsed(
self,
config: TrainConfig,
train_progress: TrainProgress,
):
return self.single_action_elapsed(
"stop_prior_training",
config.prior.stop_training_after,
config.prior.stop_training_after_unit,
train_progress,
)
def stop_text_encoder_training_elapsed(
self,
config: TrainConfig,
train_progress: TrainProgress,
):
return self.single_action_elapsed(
"stop_text_encoder_training",
config.text_encoder.stop_training_after,
config.text_encoder.stop_training_after_unit,
train_progress,
)
def stop_text_encoder_2_training_elapsed(
self,
config: TrainConfig,
train_progress: TrainProgress,
):
return self.single_action_elapsed(
"stop_text_encoder_2_training",
config.text_encoder_2.stop_training_after,
config.text_encoder_2.stop_training_after_unit,
train_progress,
)