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
|
from contextlib import contextmanager, nullcontext
import torch
from slugify import slugify
from diffusers import UNet2DConditionModel
from transformers import CLIPTextModel
from trainer.base import TrainingStrategy, Checkpointer
from training.util import EMAModel
class TextualInversionCheckpointer(Checkpointer):
def __init__(
self,
ema_embeddings: EMAModel,
placeholder_tokens: list[str],
placeholder_token_ids: list[list[int]],
*args,
**kwargs,
):
super().__init__(*args, **kwargs)
self.ema_embeddings = ema_embeddings
self.placeholder_tokens = placeholder_tokens
self.placeholder_token_ids = placeholder_token_ids
@torch.no_grad()
def checkpoint(self, step, postfix):
print(f"Saving checkpoint for step {step}...")
checkpoints_path = self.output_dir.joinpath("checkpoints")
checkpoints_path.mkdir(parents=True, exist_ok=True)
text_encoder = self.accelerator.unwrap_model(self.text_encoder)
ema_context = self.ema_embeddings.apply_temporary(
text_encoder.text_model.embeddings.temp_token_embedding.parameters()
) if self.ema_embeddings is not None else nullcontext()
with ema_context:
for (token, ids) in zip(self.placeholder_tokens, self.placeholder_token_ids):
text_encoder.text_model.embeddings.save_embed(
ids,
checkpoints_path.joinpath(f"{slugify(token)}_{step}_{postfix}.bin")
)
@torch.no_grad()
def save_samples(self, step):
ema_context = self.ema_embeddings.apply_temporary(
self.text_encoder.text_model.embeddings.temp_token_embedding.parameters()
) if self.ema_embeddings is not None else nullcontext()
with ema_context:
super().save_samples(step)
class TextualInversionTrainingStrategy(TrainingStrategy):
def __init__(
self,
unet: UNet2DConditionModel,
text_encoder: CLIPTextModel,
placeholder_tokens: list[str],
placeholder_token_ids: list[list[int]],
learning_rate: float,
gradient_checkpointing: bool = False,
use_emb_decay: bool = False,
emb_decay_target: float = 0.4,
emb_decay_factor: float = 1,
emb_decay_start: float = 1e-4,
use_ema: bool = False,
ema_inv_gamma: float = 1.0,
ema_power: int = 1,
ema_max_decay: float = 0.9999,
*args,
**kwargs,
):
super().__init__(
unet=unet,
text_encoder=text_encoder,
*args,
**kwargs
)
self.text_encoder = text_encoder
self.unet = unet
self.placeholder_tokens = placeholder_tokens
self.placeholder_token_ids = placeholder_token_ids
self.gradient_checkpointing = gradient_checkpointing
self.learning_rate = learning_rate
self.use_emb_decay = use_emb_decay
self.emb_decay_target = emb_decay_target
self.emb_decay_factor = emb_decay_factor
self.emb_decay_start = emb_decay_start
self.text_encoder.text_model.embeddings.temp_token_embedding.requires_grad_(True)
self.ema_embeddings = None
if use_ema:
self.ema_embeddings = EMAModel(
self.text_encoder.text_model.embeddings.temp_token_embedding.parameters(),
inv_gamma=ema_inv_gamma,
power=ema_power,
max_value=ema_max_decay,
)
self.checkpointer = TextualInversionCheckpointer(
unet=unet,
text_encoder=text_encoder,
ema_embeddings=self.ema_embeddings,
*args,
**kwargs
)
@property
def main_model(self):
return self.text_encoder
@contextmanager
def on_train(self, epoch: int):
try:
if self.gradient_checkpointing:
self.unet.train()
with super().on_eval():
yield
finally:
pass
@contextmanager
def on_eval(self):
try:
if self.gradient_checkpointing:
self.unet.eval()
ema_context = self.ema_embeddings.apply_temporary(
self.text_encoder.text_model.embeddings.temp_token_embedding.parameters()
) if self.ema_embeddings is not None else nullcontext()
with ema_context, super().on_eval():
yield
finally:
pass
@torch.no_grad()
def on_after_optimize(self, lr: float):
if self.use_emb_decay:
self.text_encoder.text_model.embeddings.normalize(
self.emb_decay_target,
min(1.0, max(0.0, self.emb_decay_factor * ((lr - self.emb_decay_start) / (self.learning_rate - self.emb_decay_start))))
)
if self.ema_embeddings is not None:
self.ema_embeddings.step(self.text_encoder.text_model.embeddings.temp_token_embedding.parameters())
def on_log(self):
log = super().on_log()
added = {}
if self.ema_embeddings is not None:
added = {"ema_decay": self.ema_embeddings.decay}
return log.update(added)
|