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
|
from typing import Optional
from functools import partial
from contextlib import contextmanager
from pathlib import Path
import itertools
import json
import torch
from torch.utils.data import DataLoader
from accelerate import Accelerator
from transformers import CLIPTextModel
from diffusers import AutoencoderKL, UNet2DConditionModel, DPMSolverMultistepScheduler
from peft import get_peft_model_state_dict
from safetensors.torch import save_file
from slugify import slugify
from models.clip.tokenizer import MultiCLIPTokenizer
from training.functional import TrainingStrategy, TrainingCallbacks, save_samples
def lora_strategy_callbacks(
accelerator: Accelerator,
unet: UNet2DConditionModel,
text_encoder: CLIPTextModel,
tokenizer: MultiCLIPTokenizer,
vae: AutoencoderKL,
sample_scheduler: DPMSolverMultistepScheduler,
train_dataloader: DataLoader,
val_dataloader: Optional[DataLoader],
sample_output_dir: Path,
checkpoint_output_dir: Path,
seed: int,
placeholder_tokens: list[str],
placeholder_token_ids: list[list[int]],
pti_mode: bool = False,
use_emb_decay: bool = False,
emb_decay_target: float = 0.4,
emb_decay: float = 1e-2,
max_grad_norm: float = 1.0,
sample_batch_size: int = 1,
sample_num_batches: int = 1,
sample_num_steps: int = 20,
sample_guidance_scale: float = 7.5,
sample_image_size: Optional[int] = None,
):
sample_output_dir.mkdir(parents=True, exist_ok=True)
checkpoint_output_dir.mkdir(parents=True, exist_ok=True)
save_samples_ = partial(
save_samples,
accelerator=accelerator,
tokenizer=tokenizer,
vae=vae,
sample_scheduler=sample_scheduler,
train_dataloader=train_dataloader,
val_dataloader=val_dataloader,
output_dir=sample_output_dir,
seed=seed,
batch_size=sample_batch_size,
num_batches=sample_num_batches,
num_steps=sample_num_steps,
guidance_scale=sample_guidance_scale,
image_size=sample_image_size,
)
@contextmanager
def on_train(epoch: int):
unet.train()
text_encoder.train()
tokenizer.train()
yield
@contextmanager
def on_eval():
unet.eval()
text_encoder.eval()
tokenizer.eval()
yield
def on_before_optimize(epoch: int):
if not pti_mode:
accelerator.clip_grad_norm_(
itertools.chain(
unet.parameters(),
text_encoder.text_model.encoder.parameters(),
text_encoder.text_model.final_layer_norm.parameters(),
),
max_grad_norm
)
if use_emb_decay:
params = [
p
for p in text_encoder.text_model.embeddings.token_override_embedding.parameters()
if p.grad is not None
]
return torch.stack(params) if len(params) != 0 else None
@torch.no_grad()
def on_after_optimize(w, lrs: dict[str, float]):
if use_emb_decay and w is not None and "emb" in lrs:
lr = lrs["emb"]
lambda_ = emb_decay * lr
if lambda_ != 0:
norm = w[:, :].norm(dim=-1, keepdim=True)
w[:].add_((w[:] / norm.clamp_min(1e-12)) * lambda_ * (emb_decay_target - norm))
@torch.no_grad()
def on_checkpoint(step, postfix):
if postfix != "end":
return
print(f"Saving checkpoint for step {step}...")
unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=False)
text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=False)
for (token, ids) in zip(placeholder_tokens, placeholder_token_ids):
text_encoder_.text_model.embeddings.save_embed(
ids,
checkpoint_output_dir / f"{slugify(token)}_{step}_{postfix}.bin"
)
if not pti_mode:
lora_config = {}
state_dict = get_peft_model_state_dict(unet_, state_dict=accelerator.get_state_dict(unet_))
lora_config["peft_config"] = unet_.get_peft_config_as_dict(inference=True)
text_encoder_state_dict = get_peft_model_state_dict(
text_encoder_, state_dict=accelerator.get_state_dict(text_encoder_)
)
text_encoder_state_dict = {f"text_encoder_{k}": v for k, v in text_encoder_state_dict.items()}
state_dict.update(text_encoder_state_dict)
lora_config["text_encoder_peft_config"] = text_encoder_.get_peft_config_as_dict(inference=True)
save_file(state_dict, checkpoint_output_dir / f"{step}_{postfix}.safetensors")
with open(checkpoint_output_dir / "lora_config.json", "w") as f:
json.dump(lora_config, f)
del unet_, text_encoder_
if torch.cuda.is_available():
torch.cuda.empty_cache()
@torch.no_grad()
def on_sample(step):
unet_ = accelerator.unwrap_model(unet, keep_fp32_wrapper=True)
text_encoder_ = accelerator.unwrap_model(text_encoder, keep_fp32_wrapper=True)
save_samples_(step=step, unet=unet_, text_encoder=text_encoder_)
del unet_, text_encoder_
if torch.cuda.is_available():
torch.cuda.empty_cache()
return TrainingCallbacks(
on_train=on_train,
on_eval=on_eval,
on_before_optimize=on_before_optimize,
on_after_optimize=on_after_optimize,
on_checkpoint=on_checkpoint,
on_sample=on_sample,
)
def lora_prepare(
accelerator: Accelerator,
text_encoder: CLIPTextModel,
unet: UNet2DConditionModel,
optimizer: torch.optim.Optimizer,
train_dataloader: DataLoader,
val_dataloader: Optional[DataLoader],
lr_scheduler: torch.optim.lr_scheduler._LRScheduler,
**kwargs
):
text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler = accelerator.prepare(
text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler)
text_encoder.text_model.embeddings.token_override_embedding.params.requires_grad_(True)
return text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler
lora_strategy = TrainingStrategy(
callbacks=lora_strategy_callbacks,
prepare=lora_prepare,
)
|
