Upload lora-scripts/sd-scripts/library/sdxl_train_util.py with huggingface_hub
Browse files
lora-scripts/sd-scripts/library/sdxl_train_util.py
ADDED
@@ -0,0 +1,372 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import math
|
3 |
+
import os
|
4 |
+
from typing import Optional
|
5 |
+
|
6 |
+
import torch
|
7 |
+
from library.device_utils import init_ipex, clean_memory_on_device
|
8 |
+
init_ipex()
|
9 |
+
|
10 |
+
from accelerate import init_empty_weights
|
11 |
+
from tqdm import tqdm
|
12 |
+
from transformers import CLIPTokenizer
|
13 |
+
from library import model_util, sdxl_model_util, train_util, sdxl_original_unet
|
14 |
+
from library.sdxl_lpw_stable_diffusion import SdxlStableDiffusionLongPromptWeightingPipeline
|
15 |
+
from .utils import setup_logging
|
16 |
+
setup_logging()
|
17 |
+
import logging
|
18 |
+
logger = logging.getLogger(__name__)
|
19 |
+
|
20 |
+
TOKENIZER1_PATH = "openai/clip-vit-large-patch14"
|
21 |
+
TOKENIZER2_PATH = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
|
22 |
+
|
23 |
+
# DEFAULT_NOISE_OFFSET = 0.0357
|
24 |
+
|
25 |
+
|
26 |
+
def load_target_model(args, accelerator, model_version: str, weight_dtype):
|
27 |
+
model_dtype = match_mixed_precision(args, weight_dtype) # prepare fp16/bf16
|
28 |
+
for pi in range(accelerator.state.num_processes):
|
29 |
+
if pi == accelerator.state.local_process_index:
|
30 |
+
logger.info(f"loading model for process {accelerator.state.local_process_index}/{accelerator.state.num_processes}")
|
31 |
+
|
32 |
+
(
|
33 |
+
load_stable_diffusion_format,
|
34 |
+
text_encoder1,
|
35 |
+
text_encoder2,
|
36 |
+
vae,
|
37 |
+
unet,
|
38 |
+
logit_scale,
|
39 |
+
ckpt_info,
|
40 |
+
) = _load_target_model(
|
41 |
+
args.pretrained_model_name_or_path,
|
42 |
+
args.vae,
|
43 |
+
model_version,
|
44 |
+
weight_dtype,
|
45 |
+
accelerator.device if args.lowram else "cpu",
|
46 |
+
model_dtype,
|
47 |
+
)
|
48 |
+
|
49 |
+
# work on low-ram device
|
50 |
+
if args.lowram:
|
51 |
+
text_encoder1.to(accelerator.device)
|
52 |
+
text_encoder2.to(accelerator.device)
|
53 |
+
unet.to(accelerator.device)
|
54 |
+
vae.to(accelerator.device)
|
55 |
+
|
56 |
+
clean_memory_on_device(accelerator.device)
|
57 |
+
accelerator.wait_for_everyone()
|
58 |
+
|
59 |
+
return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info
|
60 |
+
|
61 |
+
|
62 |
+
def _load_target_model(
|
63 |
+
name_or_path: str, vae_path: Optional[str], model_version: str, weight_dtype, device="cpu", model_dtype=None
|
64 |
+
):
|
65 |
+
# model_dtype only work with full fp16/bf16
|
66 |
+
name_or_path = os.readlink(name_or_path) if os.path.islink(name_or_path) else name_or_path
|
67 |
+
load_stable_diffusion_format = os.path.isfile(name_or_path) # determine SD or Diffusers
|
68 |
+
|
69 |
+
if load_stable_diffusion_format:
|
70 |
+
logger.info(f"load StableDiffusion checkpoint: {name_or_path}")
|
71 |
+
(
|
72 |
+
text_encoder1,
|
73 |
+
text_encoder2,
|
74 |
+
vae,
|
75 |
+
unet,
|
76 |
+
logit_scale,
|
77 |
+
ckpt_info,
|
78 |
+
) = sdxl_model_util.load_models_from_sdxl_checkpoint(model_version, name_or_path, device, model_dtype)
|
79 |
+
else:
|
80 |
+
# Diffusers model is loaded to CPU
|
81 |
+
from diffusers import StableDiffusionXLPipeline
|
82 |
+
|
83 |
+
variant = "fp16" if weight_dtype == torch.float16 else None
|
84 |
+
logger.info(f"load Diffusers pretrained models: {name_or_path}, variant={variant}")
|
85 |
+
try:
|
86 |
+
try:
|
87 |
+
pipe = StableDiffusionXLPipeline.from_pretrained(
|
88 |
+
name_or_path, torch_dtype=model_dtype, variant=variant, tokenizer=None
|
89 |
+
)
|
90 |
+
except EnvironmentError as ex:
|
91 |
+
if variant is not None:
|
92 |
+
logger.info("try to load fp32 model")
|
93 |
+
pipe = StableDiffusionXLPipeline.from_pretrained(name_or_path, variant=None, tokenizer=None)
|
94 |
+
else:
|
95 |
+
raise ex
|
96 |
+
except EnvironmentError as ex:
|
97 |
+
logger.error(
|
98 |
+
f"model is not found as a file or in Hugging Face, perhaps file name is wrong? / 指定したモデル名のファイル、またはHugging Faceのモデルが見つかりません。ファイル名が誤っているかもしれません: {name_or_path}"
|
99 |
+
)
|
100 |
+
raise ex
|
101 |
+
|
102 |
+
text_encoder1 = pipe.text_encoder
|
103 |
+
text_encoder2 = pipe.text_encoder_2
|
104 |
+
|
105 |
+
# convert to fp32 for cache text_encoders outputs
|
106 |
+
if text_encoder1.dtype != torch.float32:
|
107 |
+
text_encoder1 = text_encoder1.to(dtype=torch.float32)
|
108 |
+
if text_encoder2.dtype != torch.float32:
|
109 |
+
text_encoder2 = text_encoder2.to(dtype=torch.float32)
|
110 |
+
|
111 |
+
vae = pipe.vae
|
112 |
+
unet = pipe.unet
|
113 |
+
del pipe
|
114 |
+
|
115 |
+
# Diffusers U-Net to original U-Net
|
116 |
+
state_dict = sdxl_model_util.convert_diffusers_unet_state_dict_to_sdxl(unet.state_dict())
|
117 |
+
with init_empty_weights():
|
118 |
+
unet = sdxl_original_unet.SdxlUNet2DConditionModel() # overwrite unet
|
119 |
+
sdxl_model_util._load_state_dict_on_device(unet, state_dict, device=device, dtype=model_dtype)
|
120 |
+
logger.info("U-Net converted to original U-Net")
|
121 |
+
|
122 |
+
logit_scale = None
|
123 |
+
ckpt_info = None
|
124 |
+
|
125 |
+
# VAEを読み込む
|
126 |
+
if vae_path is not None:
|
127 |
+
vae = model_util.load_vae(vae_path, weight_dtype)
|
128 |
+
logger.info("additional VAE loaded")
|
129 |
+
|
130 |
+
return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info
|
131 |
+
|
132 |
+
|
133 |
+
def load_tokenizers(args: argparse.Namespace):
|
134 |
+
logger.info("prepare tokenizers")
|
135 |
+
|
136 |
+
original_paths = [TOKENIZER1_PATH, TOKENIZER2_PATH]
|
137 |
+
tokeniers = []
|
138 |
+
for i, original_path in enumerate(original_paths):
|
139 |
+
tokenizer: CLIPTokenizer = None
|
140 |
+
if args.tokenizer_cache_dir:
|
141 |
+
local_tokenizer_path = os.path.join(args.tokenizer_cache_dir, original_path.replace("/", "_"))
|
142 |
+
if os.path.exists(local_tokenizer_path):
|
143 |
+
logger.info(f"load tokenizer from cache: {local_tokenizer_path}")
|
144 |
+
tokenizer = CLIPTokenizer.from_pretrained(local_tokenizer_path)
|
145 |
+
|
146 |
+
if tokenizer is None:
|
147 |
+
tokenizer = CLIPTokenizer.from_pretrained(original_path)
|
148 |
+
|
149 |
+
if args.tokenizer_cache_dir and not os.path.exists(local_tokenizer_path):
|
150 |
+
logger.info(f"save Tokenizer to cache: {local_tokenizer_path}")
|
151 |
+
tokenizer.save_pretrained(local_tokenizer_path)
|
152 |
+
|
153 |
+
if i == 1:
|
154 |
+
tokenizer.pad_token_id = 0 # fix pad token id to make same as open clip tokenizer
|
155 |
+
|
156 |
+
tokeniers.append(tokenizer)
|
157 |
+
|
158 |
+
if hasattr(args, "max_token_length") and args.max_token_length is not None:
|
159 |
+
logger.info(f"update token length: {args.max_token_length}")
|
160 |
+
|
161 |
+
return tokeniers
|
162 |
+
|
163 |
+
|
164 |
+
def match_mixed_precision(args, weight_dtype):
|
165 |
+
if args.full_fp16:
|
166 |
+
assert (
|
167 |
+
weight_dtype == torch.float16
|
168 |
+
), "full_fp16 requires mixed precision='fp16' / full_fp16を使う場合はmixed_precision='fp16'を指定してください。"
|
169 |
+
return weight_dtype
|
170 |
+
elif args.full_bf16:
|
171 |
+
assert (
|
172 |
+
weight_dtype == torch.bfloat16
|
173 |
+
), "full_bf16 requires mixed precision='bf16' / full_bf16を使う場合はmixed_precision='bf16'を指定してください。"
|
174 |
+
return weight_dtype
|
175 |
+
else:
|
176 |
+
return None
|
177 |
+
|
178 |
+
|
179 |
+
def timestep_embedding(timesteps, dim, max_period=10000):
|
180 |
+
"""
|
181 |
+
Create sinusoidal timestep embeddings.
|
182 |
+
:param timesteps: a 1-D Tensor of N indices, one per batch element.
|
183 |
+
These may be fractional.
|
184 |
+
:param dim: the dimension of the output.
|
185 |
+
:param max_period: controls the minimum frequency of the embeddings.
|
186 |
+
:return: an [N x dim] Tensor of positional embeddings.
|
187 |
+
"""
|
188 |
+
half = dim // 2
|
189 |
+
freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
|
190 |
+
device=timesteps.device
|
191 |
+
)
|
192 |
+
args = timesteps[:, None].float() * freqs[None]
|
193 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
194 |
+
if dim % 2:
|
195 |
+
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
196 |
+
return embedding
|
197 |
+
|
198 |
+
|
199 |
+
def get_timestep_embedding(x, outdim):
|
200 |
+
assert len(x.shape) == 2
|
201 |
+
b, dims = x.shape[0], x.shape[1]
|
202 |
+
x = torch.flatten(x)
|
203 |
+
emb = timestep_embedding(x, outdim)
|
204 |
+
emb = torch.reshape(emb, (b, dims * outdim))
|
205 |
+
return emb
|
206 |
+
|
207 |
+
|
208 |
+
def get_size_embeddings(orig_size, crop_size, target_size, device):
|
209 |
+
emb1 = get_timestep_embedding(orig_size, 256)
|
210 |
+
emb2 = get_timestep_embedding(crop_size, 256)
|
211 |
+
emb3 = get_timestep_embedding(target_size, 256)
|
212 |
+
vector = torch.cat([emb1, emb2, emb3], dim=1).to(device)
|
213 |
+
return vector
|
214 |
+
|
215 |
+
|
216 |
+
def save_sd_model_on_train_end(
|
217 |
+
args: argparse.Namespace,
|
218 |
+
src_path: str,
|
219 |
+
save_stable_diffusion_format: bool,
|
220 |
+
use_safetensors: bool,
|
221 |
+
save_dtype: torch.dtype,
|
222 |
+
epoch: int,
|
223 |
+
global_step: int,
|
224 |
+
text_encoder1,
|
225 |
+
text_encoder2,
|
226 |
+
unet,
|
227 |
+
vae,
|
228 |
+
logit_scale,
|
229 |
+
ckpt_info,
|
230 |
+
):
|
231 |
+
def sd_saver(ckpt_file, epoch_no, global_step):
|
232 |
+
sai_metadata = train_util.get_sai_model_spec(None, args, True, False, False, is_stable_diffusion_ckpt=True)
|
233 |
+
sdxl_model_util.save_stable_diffusion_checkpoint(
|
234 |
+
ckpt_file,
|
235 |
+
text_encoder1,
|
236 |
+
text_encoder2,
|
237 |
+
unet,
|
238 |
+
epoch_no,
|
239 |
+
global_step,
|
240 |
+
ckpt_info,
|
241 |
+
vae,
|
242 |
+
logit_scale,
|
243 |
+
sai_metadata,
|
244 |
+
save_dtype,
|
245 |
+
)
|
246 |
+
|
247 |
+
def diffusers_saver(out_dir):
|
248 |
+
sdxl_model_util.save_diffusers_checkpoint(
|
249 |
+
out_dir,
|
250 |
+
text_encoder1,
|
251 |
+
text_encoder2,
|
252 |
+
unet,
|
253 |
+
src_path,
|
254 |
+
vae,
|
255 |
+
use_safetensors=use_safetensors,
|
256 |
+
save_dtype=save_dtype,
|
257 |
+
)
|
258 |
+
|
259 |
+
train_util.save_sd_model_on_train_end_common(
|
260 |
+
args, save_stable_diffusion_format, use_safetensors, epoch, global_step, sd_saver, diffusers_saver
|
261 |
+
)
|
262 |
+
|
263 |
+
|
264 |
+
# epochとstepの保存、メタデータにepoch/stepが含まれ引数が同じになるため、統合している
|
265 |
+
# on_epoch_end: Trueならepoch終了時、Falseならstep経過時
|
266 |
+
def save_sd_model_on_epoch_end_or_stepwise(
|
267 |
+
args: argparse.Namespace,
|
268 |
+
on_epoch_end: bool,
|
269 |
+
accelerator,
|
270 |
+
src_path,
|
271 |
+
save_stable_diffusion_format: bool,
|
272 |
+
use_safetensors: bool,
|
273 |
+
save_dtype: torch.dtype,
|
274 |
+
epoch: int,
|
275 |
+
num_train_epochs: int,
|
276 |
+
global_step: int,
|
277 |
+
text_encoder1,
|
278 |
+
text_encoder2,
|
279 |
+
unet,
|
280 |
+
vae,
|
281 |
+
logit_scale,
|
282 |
+
ckpt_info,
|
283 |
+
):
|
284 |
+
def sd_saver(ckpt_file, epoch_no, global_step):
|
285 |
+
sai_metadata = train_util.get_sai_model_spec(None, args, True, False, False, is_stable_diffusion_ckpt=True)
|
286 |
+
sdxl_model_util.save_stable_diffusion_checkpoint(
|
287 |
+
ckpt_file,
|
288 |
+
text_encoder1,
|
289 |
+
text_encoder2,
|
290 |
+
unet,
|
291 |
+
epoch_no,
|
292 |
+
global_step,
|
293 |
+
ckpt_info,
|
294 |
+
vae,
|
295 |
+
logit_scale,
|
296 |
+
sai_metadata,
|
297 |
+
save_dtype,
|
298 |
+
)
|
299 |
+
|
300 |
+
def diffusers_saver(out_dir):
|
301 |
+
sdxl_model_util.save_diffusers_checkpoint(
|
302 |
+
out_dir,
|
303 |
+
text_encoder1,
|
304 |
+
text_encoder2,
|
305 |
+
unet,
|
306 |
+
src_path,
|
307 |
+
vae,
|
308 |
+
use_safetensors=use_safetensors,
|
309 |
+
save_dtype=save_dtype,
|
310 |
+
)
|
311 |
+
|
312 |
+
train_util.save_sd_model_on_epoch_end_or_stepwise_common(
|
313 |
+
args,
|
314 |
+
on_epoch_end,
|
315 |
+
accelerator,
|
316 |
+
save_stable_diffusion_format,
|
317 |
+
use_safetensors,
|
318 |
+
epoch,
|
319 |
+
num_train_epochs,
|
320 |
+
global_step,
|
321 |
+
sd_saver,
|
322 |
+
diffusers_saver,
|
323 |
+
)
|
324 |
+
|
325 |
+
|
326 |
+
def add_sdxl_training_arguments(parser: argparse.ArgumentParser):
|
327 |
+
parser.add_argument(
|
328 |
+
"--cache_text_encoder_outputs", action="store_true", help="cache text encoder outputs / text encoderの出力をキャッシュする"
|
329 |
+
)
|
330 |
+
parser.add_argument(
|
331 |
+
"--cache_text_encoder_outputs_to_disk",
|
332 |
+
action="store_true",
|
333 |
+
help="cache text encoder outputs to disk / text encoderの出力をディスクにキャッシュする",
|
334 |
+
)
|
335 |
+
|
336 |
+
|
337 |
+
def verify_sdxl_training_args(args: argparse.Namespace, supportTextEncoderCaching: bool = True):
|
338 |
+
assert not args.v2, "v2 cannot be enabled in SDXL training / SDXL学習ではv2を有効にすることはできません"
|
339 |
+
if args.v_parameterization:
|
340 |
+
logger.warning("v_parameterization will be unexpected / SDXL学習ではv_parameterizationは想定外の動作になります")
|
341 |
+
|
342 |
+
if args.clip_skip is not None:
|
343 |
+
logger.warning("clip_skip will be unexpected / SDXL学習ではclip_skipは動作しません")
|
344 |
+
|
345 |
+
# if args.multires_noise_iterations:
|
346 |
+
# logger.info(
|
347 |
+
# f"Warning: SDXL has been trained with noise_offset={DEFAULT_NOISE_OFFSET}, but noise_offset is disabled due to multires_noise_iterations / SDXLはnoise_offset={DEFAULT_NOISE_OFFSET}で学習されていますが、multires_noise_iterationsが有効になっているためnoise_offsetは無効になります"
|
348 |
+
# )
|
349 |
+
# else:
|
350 |
+
# if args.noise_offset is None:
|
351 |
+
# args.noise_offset = DEFAULT_NOISE_OFFSET
|
352 |
+
# elif args.noise_offset != DEFAULT_NOISE_OFFSET:
|
353 |
+
# logger.info(
|
354 |
+
# f"Warning: SDXL has been trained with noise_offset={DEFAULT_NOISE_OFFSET} / SDXLはnoise_offset={DEFAULT_NOISE_OFFSET}で学習されています"
|
355 |
+
# )
|
356 |
+
# logger.info(f"noise_offset is set to {args.noise_offset} / noise_offsetが{args.noise_offset}に設定されました")
|
357 |
+
|
358 |
+
assert (
|
359 |
+
not hasattr(args, "weighted_captions") or not args.weighted_captions
|
360 |
+
), "weighted_captions cannot be enabled in SDXL training currently / SDXL学習では今のところweighted_captionsを有効にすることはできません"
|
361 |
+
|
362 |
+
if supportTextEncoderCaching:
|
363 |
+
if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs:
|
364 |
+
args.cache_text_encoder_outputs = True
|
365 |
+
logger.warning(
|
366 |
+
"cache_text_encoder_outputs is enabled because cache_text_encoder_outputs_to_disk is enabled / "
|
367 |
+
+ "cache_text_encoder_outputs_to_diskが有効になっているためcache_text_encoder_outputsが有効になりました"
|
368 |
+
)
|
369 |
+
|
370 |
+
|
371 |
+
def sample_images(*args, **kwargs):
|
372 |
+
return train_util.sample_images_common(SdxlStableDiffusionLongPromptWeightingPipeline, *args, **kwargs)
|