Commit
•
81c7085
1
Parent(s):
4e4c632
Upload 3 files
Browse files- sdxl_train.py +797 -0
- sdxl_train_util.py +391 -0
- train_util.py +0 -0
sdxl_train.py
ADDED
@@ -0,0 +1,797 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# training with captions
|
2 |
+
|
3 |
+
import argparse
|
4 |
+
import gc
|
5 |
+
import math
|
6 |
+
import os
|
7 |
+
from multiprocessing import Value
|
8 |
+
from typing import List
|
9 |
+
import toml
|
10 |
+
|
11 |
+
from tqdm import tqdm
|
12 |
+
import torch
|
13 |
+
|
14 |
+
try:
|
15 |
+
import intel_extension_for_pytorch as ipex
|
16 |
+
|
17 |
+
if torch.xpu.is_available():
|
18 |
+
from library.ipex import ipex_init
|
19 |
+
|
20 |
+
ipex_init()
|
21 |
+
except Exception:
|
22 |
+
pass
|
23 |
+
from accelerate.utils import set_seed
|
24 |
+
from diffusers import DDPMScheduler
|
25 |
+
from library import sdxl_model_util
|
26 |
+
|
27 |
+
import library.train_util as train_util
|
28 |
+
import library.config_util as config_util
|
29 |
+
import library.sdxl_train_util as sdxl_train_util
|
30 |
+
from library.config_util import (
|
31 |
+
ConfigSanitizer,
|
32 |
+
BlueprintGenerator,
|
33 |
+
)
|
34 |
+
import library.custom_train_functions as custom_train_functions
|
35 |
+
from library.custom_train_functions import (
|
36 |
+
apply_snr_weight,
|
37 |
+
prepare_scheduler_for_custom_training,
|
38 |
+
scale_v_prediction_loss_like_noise_prediction,
|
39 |
+
add_v_prediction_like_loss,
|
40 |
+
)
|
41 |
+
from library.sdxl_original_unet import SdxlUNet2DConditionModel
|
42 |
+
from library.train_util import EMAModel
|
43 |
+
|
44 |
+
|
45 |
+
UNET_NUM_BLOCKS_FOR_BLOCK_LR = 23
|
46 |
+
|
47 |
+
|
48 |
+
def get_block_params_to_optimize(unet: SdxlUNet2DConditionModel, block_lrs: List[float]) -> List[dict]:
|
49 |
+
block_params = [[] for _ in range(len(block_lrs))]
|
50 |
+
|
51 |
+
for i, (name, param) in enumerate(unet.named_parameters()):
|
52 |
+
if name.startswith("time_embed.") or name.startswith("label_emb."):
|
53 |
+
block_index = 0 # 0
|
54 |
+
elif name.startswith("input_blocks."): # 1-9
|
55 |
+
block_index = 1 + int(name.split(".")[1])
|
56 |
+
elif name.startswith("middle_block."): # 10-12
|
57 |
+
block_index = 10 + int(name.split(".")[1])
|
58 |
+
elif name.startswith("output_blocks."): # 13-21
|
59 |
+
block_index = 13 + int(name.split(".")[1])
|
60 |
+
elif name.startswith("out."): # 22
|
61 |
+
block_index = 22
|
62 |
+
else:
|
63 |
+
raise ValueError(f"unexpected parameter name: {name}")
|
64 |
+
|
65 |
+
block_params[block_index].append(param)
|
66 |
+
|
67 |
+
params_to_optimize = []
|
68 |
+
for i, params in enumerate(block_params):
|
69 |
+
if block_lrs[i] == 0: # 0のときは学習しない do not optimize when lr is 0
|
70 |
+
continue
|
71 |
+
params_to_optimize.append({"params": params, "lr": block_lrs[i]})
|
72 |
+
|
73 |
+
return params_to_optimize
|
74 |
+
|
75 |
+
|
76 |
+
def append_block_lr_to_logs(block_lrs, logs, lr_scheduler, optimizer_type):
|
77 |
+
lrs = lr_scheduler.get_last_lr()
|
78 |
+
|
79 |
+
lr_index = 0
|
80 |
+
block_index = 0
|
81 |
+
while lr_index < len(lrs):
|
82 |
+
if block_index < UNET_NUM_BLOCKS_FOR_BLOCK_LR:
|
83 |
+
name = f"block{block_index}"
|
84 |
+
if block_lrs[block_index] == 0:
|
85 |
+
block_index += 1
|
86 |
+
continue
|
87 |
+
|
88 |
+
elif block_index == UNET_NUM_BLOCKS_FOR_BLOCK_LR:
|
89 |
+
name = "text_encoder1"
|
90 |
+
elif block_index == UNET_NUM_BLOCKS_FOR_BLOCK_LR + 1:
|
91 |
+
name = "text_encoder2"
|
92 |
+
else:
|
93 |
+
raise ValueError(f"unexpected block_index: {block_index}")
|
94 |
+
|
95 |
+
block_index += 1
|
96 |
+
|
97 |
+
logs["lr/" + name] = float(lrs[lr_index])
|
98 |
+
|
99 |
+
if optimizer_type.lower().startswith("DAdapt".lower()) or optimizer_type.lower() == "Prodigy".lower():
|
100 |
+
logs["lr/d*lr/" + name] = (
|
101 |
+
lr_scheduler.optimizers[-1].param_groups[lr_index]["d"] * lr_scheduler.optimizers[-1].param_groups[lr_index]["lr"]
|
102 |
+
)
|
103 |
+
|
104 |
+
lr_index += 1
|
105 |
+
|
106 |
+
|
107 |
+
def train(args):
|
108 |
+
train_util.verify_training_args(args)
|
109 |
+
train_util.prepare_dataset_args(args, True)
|
110 |
+
sdxl_train_util.verify_sdxl_training_args(args)
|
111 |
+
|
112 |
+
assert not args.weighted_captions, "weighted_captions is not supported currently / weighted_captionsは現在サポートされていません"
|
113 |
+
assert (
|
114 |
+
not args.train_text_encoder or not args.cache_text_encoder_outputs
|
115 |
+
), "cache_text_encoder_outputs is not supported when training text encoder / text encoderを学習するときはcache_text_encoder_outputsはサポートされていません"
|
116 |
+
|
117 |
+
if args.block_lr:
|
118 |
+
block_lrs = [float(lr) for lr in args.block_lr.split(",")]
|
119 |
+
assert (
|
120 |
+
len(block_lrs) == UNET_NUM_BLOCKS_FOR_BLOCK_LR
|
121 |
+
), f"block_lr must have {UNET_NUM_BLOCKS_FOR_BLOCK_LR} values / block_lrは{UNET_NUM_BLOCKS_FOR_BLOCK_LR}個の値を指定してください"
|
122 |
+
else:
|
123 |
+
block_lrs = None
|
124 |
+
|
125 |
+
cache_latents = args.cache_latents
|
126 |
+
use_dreambooth_method = args.in_json is None
|
127 |
+
|
128 |
+
if args.seed is not None:
|
129 |
+
set_seed(args.seed) # 乱数系列を初期化する
|
130 |
+
|
131 |
+
tokenizer1, tokenizer2 = sdxl_train_util.load_tokenizers(args)
|
132 |
+
|
133 |
+
# データセットを準備する
|
134 |
+
if args.dataset_class is None:
|
135 |
+
blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, False, True))
|
136 |
+
if args.dataset_config is not None:
|
137 |
+
print(f"Load dataset config from {args.dataset_config}")
|
138 |
+
user_config = config_util.load_user_config(args.dataset_config)
|
139 |
+
ignored = ["train_data_dir", "in_json"]
|
140 |
+
if any(getattr(args, attr) is not None for attr in ignored):
|
141 |
+
print(
|
142 |
+
"ignore following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format(
|
143 |
+
", ".join(ignored)
|
144 |
+
)
|
145 |
+
)
|
146 |
+
else:
|
147 |
+
if use_dreambooth_method:
|
148 |
+
print("Using DreamBooth method.")
|
149 |
+
user_config = {
|
150 |
+
"datasets": [
|
151 |
+
{
|
152 |
+
"subsets": config_util.generate_dreambooth_subsets_config_by_subdirs(
|
153 |
+
args.train_data_dir, args.reg_data_dir
|
154 |
+
)
|
155 |
+
}
|
156 |
+
]
|
157 |
+
}
|
158 |
+
else:
|
159 |
+
print("Training with captions.")
|
160 |
+
user_config = {
|
161 |
+
"datasets": [
|
162 |
+
{
|
163 |
+
"subsets": [
|
164 |
+
{
|
165 |
+
"image_dir": args.train_data_dir,
|
166 |
+
"metadata_file": args.in_json,
|
167 |
+
}
|
168 |
+
]
|
169 |
+
}
|
170 |
+
]
|
171 |
+
}
|
172 |
+
|
173 |
+
blueprint = blueprint_generator.generate(user_config, args, tokenizer=[tokenizer1, tokenizer2])
|
174 |
+
train_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group)
|
175 |
+
else:
|
176 |
+
train_dataset_group = train_util.load_arbitrary_dataset(args, [tokenizer1, tokenizer2])
|
177 |
+
|
178 |
+
current_epoch = Value("i", 0)
|
179 |
+
current_step = Value("i", 0)
|
180 |
+
ds_for_collator = train_dataset_group if args.max_data_loader_n_workers == 0 else None
|
181 |
+
collator = train_util.collator_class(current_epoch, current_step, ds_for_collator)
|
182 |
+
|
183 |
+
train_dataset_group.verify_bucket_reso_steps(32)
|
184 |
+
|
185 |
+
if args.debug_dataset:
|
186 |
+
train_util.debug_dataset(train_dataset_group, True)
|
187 |
+
return
|
188 |
+
if len(train_dataset_group) == 0:
|
189 |
+
print(
|
190 |
+
"No data found. Please verify the metadata file and train_data_dir option. / 画像がありません。メタデータおよびtrain_data_dirオプションを確認してください。"
|
191 |
+
)
|
192 |
+
return
|
193 |
+
|
194 |
+
if cache_latents:
|
195 |
+
assert (
|
196 |
+
train_dataset_group.is_latent_cacheable()
|
197 |
+
), "when caching latents, either color_aug or random_crop cannot be used / latentをキャッシュするときはcolor_augとrandom_cropは使えません"
|
198 |
+
|
199 |
+
if args.cache_text_encoder_outputs:
|
200 |
+
assert (
|
201 |
+
train_dataset_group.is_text_encoder_output_cacheable()
|
202 |
+
), "when caching text encoder output, either caption_dropout_rate, shuffle_caption, token_warmup_step or caption_tag_dropout_rate cannot be used / text encoderの出力をキャッシュするときはcaption_dropout_rate, shuffle_caption, token_warmup_step, caption_tag_dropout_rateは使えません"
|
203 |
+
|
204 |
+
# acceleratorを準備する
|
205 |
+
print("prepare accelerator")
|
206 |
+
accelerator = train_util.prepare_accelerator(args)
|
207 |
+
|
208 |
+
# mixed precisionに対応した型を用意しておき適宜castする
|
209 |
+
weight_dtype, save_dtype = train_util.prepare_dtype(args)
|
210 |
+
vae_dtype = torch.float32 if args.no_half_vae else weight_dtype
|
211 |
+
|
212 |
+
# モデルを読み込む
|
213 |
+
(
|
214 |
+
load_stable_diffusion_format,
|
215 |
+
text_encoder1,
|
216 |
+
text_encoder2,
|
217 |
+
vae,
|
218 |
+
unet,
|
219 |
+
logit_scale,
|
220 |
+
ckpt_info,
|
221 |
+
) = sdxl_train_util.load_target_model(args, accelerator, "sdxl", weight_dtype)
|
222 |
+
# logit_scale = logit_scale.to(accelerator.device, dtype=weight_dtype)
|
223 |
+
|
224 |
+
# verify load/save model formats
|
225 |
+
if load_stable_diffusion_format:
|
226 |
+
src_stable_diffusion_ckpt = args.pretrained_model_name_or_path
|
227 |
+
src_diffusers_model_path = None
|
228 |
+
else:
|
229 |
+
src_stable_diffusion_ckpt = None
|
230 |
+
src_diffusers_model_path = args.pretrained_model_name_or_path
|
231 |
+
|
232 |
+
if args.save_model_as is None:
|
233 |
+
save_stable_diffusion_format = load_stable_diffusion_format
|
234 |
+
use_safetensors = args.use_safetensors
|
235 |
+
else:
|
236 |
+
save_stable_diffusion_format = args.save_model_as.lower() == "ckpt" or args.save_model_as.lower() == "safetensors"
|
237 |
+
use_safetensors = args.use_safetensors or ("safetensors" in args.save_model_as.lower())
|
238 |
+
# assert save_stable_diffusion_format, "save_model_as must be ckpt or safetensors / save_model_asはckptかsafetensorsである必要があります"
|
239 |
+
|
240 |
+
# Diffusers版のxformers使用フラグを設定する関数
|
241 |
+
def set_diffusers_xformers_flag(model, valid):
|
242 |
+
def fn_recursive_set_mem_eff(module: torch.nn.Module):
|
243 |
+
if hasattr(module, "set_use_memory_efficient_attention_xformers"):
|
244 |
+
module.set_use_memory_efficient_attention_xformers(valid)
|
245 |
+
|
246 |
+
for child in module.children():
|
247 |
+
fn_recursive_set_mem_eff(child)
|
248 |
+
|
249 |
+
fn_recursive_set_mem_eff(model)
|
250 |
+
|
251 |
+
# モデルに xformers とか memory efficient attention を組み込む
|
252 |
+
if args.diffusers_xformers:
|
253 |
+
# もうU-Netを独自にしたので動かないけどVAEのxformersは動くはず
|
254 |
+
accelerator.print("Use xformers by Diffusers")
|
255 |
+
# set_diffusers_xformers_flag(unet, True)
|
256 |
+
set_diffusers_xformers_flag(vae, True)
|
257 |
+
else:
|
258 |
+
# Windows版のxformersはfloatで学習できなかったりするのでxformersを使わない設定も可能にしておく必要がある
|
259 |
+
accelerator.print("Disable Diffusers' xformers")
|
260 |
+
train_util.replace_unet_modules(unet, args.mem_eff_attn, args.xformers, args.sdpa)
|
261 |
+
if torch.__version__ >= "2.0.0": # PyTorch 2.0.0 以上対応のxformersなら以下が使える
|
262 |
+
vae.set_use_memory_efficient_attention_xformers(args.xformers)
|
263 |
+
|
264 |
+
# 学習を準備する
|
265 |
+
if cache_latents:
|
266 |
+
vae.to(accelerator.device, dtype=vae_dtype)
|
267 |
+
vae.requires_grad_(False)
|
268 |
+
vae.eval()
|
269 |
+
with torch.no_grad():
|
270 |
+
train_dataset_group.cache_latents(vae, args.vae_batch_size, args.cache_latents_to_disk, accelerator.is_main_process)
|
271 |
+
vae.to("cpu")
|
272 |
+
if torch.cuda.is_available():
|
273 |
+
torch.cuda.empty_cache()
|
274 |
+
gc.collect()
|
275 |
+
|
276 |
+
accelerator.wait_for_everyone()
|
277 |
+
|
278 |
+
# 学習を準備する:モデルを適切な状態にする
|
279 |
+
training_models = []
|
280 |
+
if args.gradient_checkpointing:
|
281 |
+
unet.enable_gradient_checkpointing()
|
282 |
+
training_models.append(unet)
|
283 |
+
if args.train_text_encoder:
|
284 |
+
# TODO each option for two text encoders?
|
285 |
+
accelerator.print("enable text encoder training")
|
286 |
+
if args.gradient_checkpointing:
|
287 |
+
text_encoder1.gradient_checkpointing_enable()
|
288 |
+
text_encoder2.gradient_checkpointing_enable()
|
289 |
+
training_models.append(text_encoder1)
|
290 |
+
training_models.append(text_encoder2)
|
291 |
+
# set require_grad=True later
|
292 |
+
else:
|
293 |
+
text_encoder1.requires_grad_(False)
|
294 |
+
text_encoder2.requires_grad_(False)
|
295 |
+
text_encoder1.eval()
|
296 |
+
text_encoder2.eval()
|
297 |
+
|
298 |
+
# TextEncoderの出力をキャッシュする
|
299 |
+
if args.cache_text_encoder_outputs:
|
300 |
+
# Text Encodes are eval and no grad
|
301 |
+
with torch.no_grad():
|
302 |
+
train_dataset_group.cache_text_encoder_outputs(
|
303 |
+
(tokenizer1, tokenizer2),
|
304 |
+
(text_encoder1, text_encoder2),
|
305 |
+
accelerator.device,
|
306 |
+
None,
|
307 |
+
args.cache_text_encoder_outputs_to_disk,
|
308 |
+
accelerator.is_main_process,
|
309 |
+
)
|
310 |
+
accelerator.wait_for_everyone()
|
311 |
+
|
312 |
+
if not cache_latents:
|
313 |
+
vae.requires_grad_(False)
|
314 |
+
vae.eval()
|
315 |
+
vae.to(accelerator.device, dtype=vae_dtype)
|
316 |
+
|
317 |
+
for m in training_models:
|
318 |
+
m.requires_grad_(True)
|
319 |
+
|
320 |
+
if block_lrs is None:
|
321 |
+
params = []
|
322 |
+
for m in training_models:
|
323 |
+
params.extend(m.parameters())
|
324 |
+
params_to_optimize = params
|
325 |
+
|
326 |
+
# calculate number of trainable parameters
|
327 |
+
n_params = 0
|
328 |
+
for p in params:
|
329 |
+
n_params += p.numel()
|
330 |
+
else:
|
331 |
+
params_to_optimize = get_block_params_to_optimize(training_models[0], block_lrs) # U-Net
|
332 |
+
for m in training_models[1:]: # Text Encoders if exists
|
333 |
+
params_to_optimize.append({"params": m.parameters(), "lr": args.learning_rate})
|
334 |
+
|
335 |
+
# calculate number of trainable parameters
|
336 |
+
n_params = 0
|
337 |
+
for params in params_to_optimize:
|
338 |
+
for p in params["params"]:
|
339 |
+
n_params += p.numel()
|
340 |
+
|
341 |
+
accelerator.print(f"number of models: {len(training_models)}")
|
342 |
+
accelerator.print(f"number of trainable parameters: {n_params}")
|
343 |
+
|
344 |
+
# 学習に必要なクラスを準備する
|
345 |
+
accelerator.print("prepare optimizer, data loader etc.")
|
346 |
+
_, _, optimizer = train_util.get_optimizer(args, trainable_params=params_to_optimize)
|
347 |
+
|
348 |
+
# dataloaderを準備する
|
349 |
+
# DataLoaderのプロセス数:0はメインプロセスになる
|
350 |
+
n_workers = min(args.max_data_loader_n_workers, os.cpu_count() - 1) # cpu_count-1 ただし最大で指定された数まで
|
351 |
+
train_dataloader = torch.utils.data.DataLoader(
|
352 |
+
train_dataset_group,
|
353 |
+
batch_size=1,
|
354 |
+
shuffle=True,
|
355 |
+
collate_fn=collator,
|
356 |
+
num_workers=n_workers,
|
357 |
+
persistent_workers=args.persistent_data_loader_workers,
|
358 |
+
)
|
359 |
+
|
360 |
+
# 学習ステップ数を計算する
|
361 |
+
if args.max_train_epochs is not None:
|
362 |
+
args.max_train_steps = args.max_train_epochs * math.ceil(
|
363 |
+
len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
|
364 |
+
)
|
365 |
+
accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
|
366 |
+
|
367 |
+
# データセット側にも学習ステップを送信
|
368 |
+
train_dataset_group.set_max_train_steps(args.max_train_steps)
|
369 |
+
|
370 |
+
# lr schedulerを用意する
|
371 |
+
lr_scheduler = train_util.get_scheduler_fix(args, optimizer, accelerator.num_processes)
|
372 |
+
|
373 |
+
# 実験的機能:勾配も含めたfp16/bf16学習を行う モデル全体をfp16/bf16にする
|
374 |
+
if args.full_fp16:
|
375 |
+
assert (
|
376 |
+
args.mixed_precision == "fp16"
|
377 |
+
), "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。"
|
378 |
+
accelerator.print("enable full fp16 training.")
|
379 |
+
unet.to(weight_dtype)
|
380 |
+
text_encoder1.to(weight_dtype)
|
381 |
+
text_encoder2.to(weight_dtype)
|
382 |
+
elif args.full_bf16:
|
383 |
+
assert (
|
384 |
+
args.mixed_precision == "bf16"
|
385 |
+
), "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。"
|
386 |
+
accelerator.print("enable full bf16 training.")
|
387 |
+
unet.to(weight_dtype)
|
388 |
+
text_encoder1.to(weight_dtype)
|
389 |
+
text_encoder2.to(weight_dtype)
|
390 |
+
|
391 |
+
if args.enable_ema:
|
392 |
+
#ema_dtype = weight_dtype if (args.full_bf16 or args.full_fp16) else torch.float
|
393 |
+
ema = EMAModel(params_to_optimize, decay=args.ema_decay, beta=args.ema_exp_beta, max_train_steps=args.max_train_steps)
|
394 |
+
ema.to(accelerator.device, dtype=weight_dtype)
|
395 |
+
ema = accelerator.prepare(ema)
|
396 |
+
else:
|
397 |
+
ema = None
|
398 |
+
# acceleratorがなんかよろしくやってくれるらしい
|
399 |
+
if args.train_text_encoder:
|
400 |
+
unet, text_encoder1, text_encoder2, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
401 |
+
unet, text_encoder1, text_encoder2, optimizer, train_dataloader, lr_scheduler
|
402 |
+
)
|
403 |
+
|
404 |
+
# transform DDP after prepare
|
405 |
+
text_encoder1, text_encoder2, unet = train_util.transform_models_if_DDP([text_encoder1, text_encoder2, unet])
|
406 |
+
else:
|
407 |
+
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
|
408 |
+
(unet,) = train_util.transform_models_if_DDP([unet])
|
409 |
+
text_encoder1.to(weight_dtype)
|
410 |
+
text_encoder2.to(weight_dtype)
|
411 |
+
|
412 |
+
# TextEncoderの出力をキャッシュするときにはCPUへ移動する
|
413 |
+
if args.cache_text_encoder_outputs:
|
414 |
+
# move Text Encoders for sampling images. Text Encoder doesn't work on CPU with fp16
|
415 |
+
text_encoder1.to("cpu", dtype=torch.float32)
|
416 |
+
text_encoder2.to("cpu", dtype=torch.float32)
|
417 |
+
if torch.cuda.is_available():
|
418 |
+
torch.cuda.empty_cache()
|
419 |
+
else:
|
420 |
+
# make sure Text Encoders are on GPU
|
421 |
+
text_encoder1.to(accelerator.device)
|
422 |
+
text_encoder2.to(accelerator.device)
|
423 |
+
|
424 |
+
# 実験的機能:勾配も含めたfp16学習を行う PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
|
425 |
+
if args.full_fp16:
|
426 |
+
train_util.patch_accelerator_for_fp16_training(accelerator)
|
427 |
+
|
428 |
+
# resumeする
|
429 |
+
train_util.resume_from_local_or_hf_if_specified(accelerator, args)
|
430 |
+
|
431 |
+
# epoch数を計算する
|
432 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
433 |
+
num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
|
434 |
+
if (args.save_n_epoch_ratio is not None) and (args.save_n_epoch_ratio > 0):
|
435 |
+
args.save_every_n_epochs = math.floor(num_train_epochs / args.save_n_epoch_ratio) or 1
|
436 |
+
|
437 |
+
# 学習する
|
438 |
+
# total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
|
439 |
+
accelerator.print("running training / 学習開始")
|
440 |
+
accelerator.print(f" num examples / サンプル数: {train_dataset_group.num_train_images}")
|
441 |
+
accelerator.print(f" num batches per epoch / 1epochのバッチ数: {len(train_dataloader)}")
|
442 |
+
accelerator.print(f" num epochs / epoch数: {num_train_epochs}")
|
443 |
+
accelerator.print(f" batch size per device / バッチサイズ: {', '.join([str(d.batch_size) for d in train_dataset_group.datasets])}")
|
444 |
+
# accelerator.print(
|
445 |
+
# f" total train batch size (with parallel & distributed & accumulation) / 総バッチサイズ(並列学習、勾配合計含む): {total_batch_size}"
|
446 |
+
# )
|
447 |
+
accelerator.print(f" gradient accumulation steps / 勾配を合計するステップ数 = {args.gradient_accumulation_steps}")
|
448 |
+
accelerator.print(f" total optimization steps / 学習ステップ数: {args.max_train_steps}")
|
449 |
+
|
450 |
+
progress_bar = tqdm(range(args.max_train_steps), smoothing=0, disable=not accelerator.is_local_main_process, desc="steps")
|
451 |
+
global_step = 0
|
452 |
+
|
453 |
+
noise_scheduler = DDPMScheduler(
|
454 |
+
beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000, clip_sample=False
|
455 |
+
)
|
456 |
+
prepare_scheduler_for_custom_training(noise_scheduler, accelerator.device)
|
457 |
+
if args.zero_terminal_snr:
|
458 |
+
custom_train_functions.fix_noise_scheduler_betas_for_zero_terminal_snr(noise_scheduler)
|
459 |
+
|
460 |
+
if accelerator.is_main_process:
|
461 |
+
init_kwargs = {}
|
462 |
+
if args.log_tracker_config is not None:
|
463 |
+
init_kwargs = toml.load(args.log_tracker_config)
|
464 |
+
accelerator.init_trackers("finetuning" if args.log_tracker_name is None else args.log_tracker_name, init_kwargs=init_kwargs)
|
465 |
+
|
466 |
+
for epoch in range(num_train_epochs):
|
467 |
+
accelerator.print(f"\nepoch {epoch+1}/{num_train_epochs}")
|
468 |
+
current_epoch.value = epoch + 1
|
469 |
+
|
470 |
+
for m in training_models:
|
471 |
+
m.train()
|
472 |
+
|
473 |
+
loss_total = 0
|
474 |
+
for step, batch in enumerate(train_dataloader):
|
475 |
+
current_step.value = global_step
|
476 |
+
with accelerator.accumulate(training_models[0]): # 複数モデルに対応していない模様だがとりあえずこうしておく
|
477 |
+
if "latents" in batch and batch["latents"] is not None:
|
478 |
+
latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
|
479 |
+
else:
|
480 |
+
with torch.no_grad():
|
481 |
+
# latentに変換
|
482 |
+
latents = vae.encode(batch["images"].to(vae_dtype)).latent_dist.sample().to(weight_dtype)
|
483 |
+
|
484 |
+
# NaNが含まれていれば警告を表示し0に置き換える
|
485 |
+
if torch.any(torch.isnan(latents)):
|
486 |
+
accelerator.print("NaN found in latents, replacing with zeros")
|
487 |
+
latents = torch.where(torch.isnan(latents), torch.zeros_like(latents), latents)
|
488 |
+
latents = latents * sdxl_model_util.VAE_SCALE_FACTOR
|
489 |
+
|
490 |
+
if "text_encoder_outputs1_list" not in batch or batch["text_encoder_outputs1_list"] is None:
|
491 |
+
input_ids1 = batch["input_ids"]
|
492 |
+
input_ids2 = batch["input_ids2"]
|
493 |
+
with torch.set_grad_enabled(args.train_text_encoder):
|
494 |
+
# Get the text embedding for conditioning
|
495 |
+
# TODO support weighted captions
|
496 |
+
# if args.weighted_captions:
|
497 |
+
# encoder_hidden_states = get_weighted_text_embeddings(
|
498 |
+
# tokenizer,
|
499 |
+
# text_encoder,
|
500 |
+
# batch["captions"],
|
501 |
+
# accelerator.device,
|
502 |
+
# args.max_token_length // 75 if args.max_token_length else 1,
|
503 |
+
# clip_skip=args.clip_skip,
|
504 |
+
# )
|
505 |
+
# else:
|
506 |
+
input_ids1 = input_ids1.to(accelerator.device)
|
507 |
+
input_ids2 = input_ids2.to(accelerator.device)
|
508 |
+
encoder_hidden_states1, encoder_hidden_states2, pool2 = train_util.get_hidden_states_sdxl(
|
509 |
+
args.max_token_length,
|
510 |
+
input_ids1,
|
511 |
+
input_ids2,
|
512 |
+
tokenizer1,
|
513 |
+
tokenizer2,
|
514 |
+
text_encoder1,
|
515 |
+
text_encoder2,
|
516 |
+
None if not args.full_fp16 else weight_dtype,
|
517 |
+
)
|
518 |
+
else:
|
519 |
+
encoder_hidden_states1 = batch["text_encoder_outputs1_list"].to(accelerator.device).to(weight_dtype)
|
520 |
+
encoder_hidden_states2 = batch["text_encoder_outputs2_list"].to(accelerator.device).to(weight_dtype)
|
521 |
+
pool2 = batch["text_encoder_pool2_list"].to(accelerator.device).to(weight_dtype)
|
522 |
+
|
523 |
+
# # verify that the text encoder outputs are correct
|
524 |
+
# ehs1, ehs2, p2 = train_util.get_hidden_states_sdxl(
|
525 |
+
# args.max_token_length,
|
526 |
+
# batch["input_ids"].to(text_encoder1.device),
|
527 |
+
# batch["input_ids2"].to(text_encoder1.device),
|
528 |
+
# tokenizer1,
|
529 |
+
# tokenizer2,
|
530 |
+
# text_encoder1,
|
531 |
+
# text_encoder2,
|
532 |
+
# None if not args.full_fp16 else weight_dtype,
|
533 |
+
# )
|
534 |
+
# b_size = encoder_hidden_states1.shape[0]
|
535 |
+
# assert ((encoder_hidden_states1.to("cpu") - ehs1.to(dtype=weight_dtype)).abs().max() > 1e-2).sum() <= b_size * 2
|
536 |
+
# assert ((encoder_hidden_states2.to("cpu") - ehs2.to(dtype=weight_dtype)).abs().max() > 1e-2).sum() <= b_size * 2
|
537 |
+
# assert ((pool2.to("cpu") - p2.to(dtype=weight_dtype)).abs().max() > 1e-2).sum() <= b_size * 2
|
538 |
+
# print("text encoder outputs verified")
|
539 |
+
|
540 |
+
# get size embeddings
|
541 |
+
orig_size = batch["original_sizes_hw"]
|
542 |
+
crop_size = batch["crop_top_lefts"]
|
543 |
+
target_size = batch["target_sizes_hw"]
|
544 |
+
embs = sdxl_train_util.get_size_embeddings(orig_size, crop_size, target_size, accelerator.device).to(weight_dtype)
|
545 |
+
|
546 |
+
# concat embeddings
|
547 |
+
vector_embedding = torch.cat([pool2, embs], dim=1).to(weight_dtype)
|
548 |
+
text_embedding = torch.cat([encoder_hidden_states1, encoder_hidden_states2], dim=2).to(weight_dtype)
|
549 |
+
|
550 |
+
# Sample noise, sample a random timestep for each image, and add noise to the latents,
|
551 |
+
# with noise offset and/or multires noise if specified
|
552 |
+
noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
|
553 |
+
|
554 |
+
noisy_latents = noisy_latents.to(weight_dtype) # TODO check why noisy_latents is not weight_dtype
|
555 |
+
|
556 |
+
# Predict the noise residual
|
557 |
+
with accelerator.autocast():
|
558 |
+
noise_pred = unet(noisy_latents, timesteps, text_embedding, vector_embedding)
|
559 |
+
|
560 |
+
target = noise
|
561 |
+
|
562 |
+
if args.min_snr_gamma or args.scale_v_pred_loss_like_noise_pred or args.v_pred_like_loss:
|
563 |
+
|
564 |
+
# do not mean over batch dimension for snr weight or scale v-pred loss
|
565 |
+
loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
|
566 |
+
loss = loss.mean([1, 2, 3])
|
567 |
+
|
568 |
+
if args.min_snr_gamma:
|
569 |
+
loss = apply_snr_weight(loss, timesteps, noise_scheduler, args.min_snr_gamma)
|
570 |
+
if args.scale_v_pred_loss_like_noise_pred:
|
571 |
+
loss = scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler)
|
572 |
+
if args.v_pred_like_loss:
|
573 |
+
loss = add_v_prediction_like_loss(loss, timesteps, noise_scheduler, args.v_pred_like_loss)
|
574 |
+
|
575 |
+
loss = loss.mean() # mean over batch dimension
|
576 |
+
else:
|
577 |
+
loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="mean")
|
578 |
+
|
579 |
+
accelerator.backward(loss)
|
580 |
+
if accelerator.sync_gradients and args.max_grad_norm != 0.0:
|
581 |
+
params_to_clip = []
|
582 |
+
for m in training_models:
|
583 |
+
params_to_clip.extend(m.parameters())
|
584 |
+
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
|
585 |
+
|
586 |
+
optimizer.step()
|
587 |
+
lr_scheduler.step()
|
588 |
+
optimizer.zero_grad(set_to_none=True)
|
589 |
+
if args.enable_ema:
|
590 |
+
with torch.no_grad(), accelerator.autocast():
|
591 |
+
ema.step(params_to_optimize)
|
592 |
+
|
593 |
+
# Checks if the accelerator has performed an optimization step behind the scenes
|
594 |
+
if accelerator.sync_gradients:
|
595 |
+
progress_bar.update(1)
|
596 |
+
global_step += 1
|
597 |
+
|
598 |
+
sdxl_train_util.sample_images(
|
599 |
+
accelerator,
|
600 |
+
args,
|
601 |
+
None,
|
602 |
+
global_step,
|
603 |
+
accelerator.device,
|
604 |
+
vae,
|
605 |
+
[tokenizer1, tokenizer2],
|
606 |
+
[text_encoder1, text_encoder2],
|
607 |
+
unet,
|
608 |
+
)
|
609 |
+
|
610 |
+
# 指定ステップごとにモデルを保存
|
611 |
+
if args.save_every_n_steps is not None and global_step % args.save_every_n_steps == 0:
|
612 |
+
accelerator.wait_for_everyone()
|
613 |
+
if accelerator.is_main_process:
|
614 |
+
src_path = src_stable_diffusion_ckpt if save_stable_diffusion_format else src_diffusers_model_path
|
615 |
+
sdxl_train_util.save_sd_model_on_epoch_end_or_stepwise(
|
616 |
+
args,
|
617 |
+
False,
|
618 |
+
accelerator,
|
619 |
+
src_path,
|
620 |
+
save_stable_diffusion_format,
|
621 |
+
use_safetensors,
|
622 |
+
save_dtype,
|
623 |
+
epoch,
|
624 |
+
num_train_epochs,
|
625 |
+
global_step,
|
626 |
+
accelerator.unwrap_model(text_encoder1),
|
627 |
+
accelerator.unwrap_model(text_encoder2),
|
628 |
+
accelerator.unwrap_model(unet),
|
629 |
+
vae,
|
630 |
+
logit_scale,
|
631 |
+
ckpt_info,
|
632 |
+
ema=ema,
|
633 |
+
params_to_replace=params_to_optimize,
|
634 |
+
)
|
635 |
+
|
636 |
+
current_loss = loss.detach().item() # 平均なのでbatch sizeは関係ないはず
|
637 |
+
if args.logging_dir is not None:
|
638 |
+
logs = {"loss": current_loss}
|
639 |
+
if block_lrs is None:
|
640 |
+
logs["lr"] = float(lr_scheduler.get_last_lr()[0])
|
641 |
+
if (
|
642 |
+
args.optimizer_type.lower().startswith("DAdapt".lower()) or args.optimizer_type.lower() == "Prodigy".lower()
|
643 |
+
): # tracking d*lr value
|
644 |
+
logs["lr/d*lr"] = (
|
645 |
+
lr_scheduler.optimizers[0].param_groups[0]["d"] * lr_scheduler.optimizers[0].param_groups[0]["lr"]
|
646 |
+
)
|
647 |
+
else:
|
648 |
+
append_block_lr_to_logs(block_lrs, logs, lr_scheduler, args.optimizer_type)
|
649 |
+
|
650 |
+
accelerator.log(logs, step=global_step)
|
651 |
+
|
652 |
+
# TODO moving averageにする
|
653 |
+
loss_total += current_loss
|
654 |
+
avr_loss = loss_total / (step + 1)
|
655 |
+
logs = {"loss": avr_loss} # , "lr": lr_scheduler.get_last_lr()[0]}
|
656 |
+
progress_bar.set_postfix(**logs)
|
657 |
+
|
658 |
+
if global_step >= args.max_train_steps:
|
659 |
+
break
|
660 |
+
|
661 |
+
if args.logging_dir is not None:
|
662 |
+
logs = {"loss/epoch": loss_total / len(train_dataloader)}
|
663 |
+
accelerator.log(logs, step=epoch + 1)
|
664 |
+
|
665 |
+
accelerator.wait_for_everyone()
|
666 |
+
|
667 |
+
if args.save_every_n_epochs is not None:
|
668 |
+
if accelerator.is_main_process:
|
669 |
+
src_path = src_stable_diffusion_ckpt if save_stable_diffusion_format else src_diffusers_model_path
|
670 |
+
sdxl_train_util.save_sd_model_on_epoch_end_or_stepwise(
|
671 |
+
args,
|
672 |
+
True,
|
673 |
+
accelerator,
|
674 |
+
src_path,
|
675 |
+
save_stable_diffusion_format,
|
676 |
+
use_safetensors,
|
677 |
+
save_dtype,
|
678 |
+
epoch,
|
679 |
+
num_train_epochs,
|
680 |
+
global_step,
|
681 |
+
accelerator.unwrap_model(text_encoder1),
|
682 |
+
accelerator.unwrap_model(text_encoder2),
|
683 |
+
accelerator.unwrap_model(unet),
|
684 |
+
vae,
|
685 |
+
logit_scale,
|
686 |
+
ckpt_info,
|
687 |
+
ema=ema,
|
688 |
+
params_to_replace=params_to_optimize,
|
689 |
+
)
|
690 |
+
|
691 |
+
sdxl_train_util.sample_images(
|
692 |
+
accelerator,
|
693 |
+
args,
|
694 |
+
epoch + 1,
|
695 |
+
global_step,
|
696 |
+
accelerator.device,
|
697 |
+
vae,
|
698 |
+
[tokenizer1, tokenizer2],
|
699 |
+
[text_encoder1, text_encoder2],
|
700 |
+
unet,
|
701 |
+
)
|
702 |
+
|
703 |
+
is_main_process = accelerator.is_main_process
|
704 |
+
# if is_main_process:
|
705 |
+
unet = accelerator.unwrap_model(unet)
|
706 |
+
text_encoder1 = accelerator.unwrap_model(text_encoder1)
|
707 |
+
text_encoder2 = accelerator.unwrap_model(text_encoder2)
|
708 |
+
if args.enable_ema:
|
709 |
+
ema = accelerator.unwrap_model(ema)
|
710 |
+
|
711 |
+
accelerator.end_training()
|
712 |
+
|
713 |
+
if args.save_state: # and is_main_process:
|
714 |
+
train_util.save_state_on_train_end(args, accelerator)
|
715 |
+
|
716 |
+
del accelerator # この後メモリを使うのでこれは消す
|
717 |
+
|
718 |
+
if is_main_process:
|
719 |
+
src_path = src_stable_diffusion_ckpt if save_stable_diffusion_format else src_diffusers_model_path
|
720 |
+
if args.enable_ema and not args.ema_save_only_ema_weights:
|
721 |
+
temp_name = args.output_name
|
722 |
+
args.output_name = args.output_name + "-non-EMA"
|
723 |
+
sdxl_train_util.save_sd_model_on_train_end(
|
724 |
+
args,
|
725 |
+
src_path,
|
726 |
+
save_stable_diffusion_format,
|
727 |
+
use_safetensors,
|
728 |
+
save_dtype,
|
729 |
+
epoch,
|
730 |
+
global_step,
|
731 |
+
text_encoder1,
|
732 |
+
text_encoder2,
|
733 |
+
unet,
|
734 |
+
vae,
|
735 |
+
logit_scale,
|
736 |
+
ckpt_info,
|
737 |
+
)
|
738 |
+
args.output_name = temp_name
|
739 |
+
if args.enable_ema:
|
740 |
+
print("Saving EMA:")
|
741 |
+
ema.copy_to(params_to_optimize)
|
742 |
+
|
743 |
+
sdxl_train_util.save_sd_model_on_train_end(
|
744 |
+
args,
|
745 |
+
src_path,
|
746 |
+
save_stable_diffusion_format,
|
747 |
+
use_safetensors,
|
748 |
+
save_dtype,
|
749 |
+
epoch,
|
750 |
+
global_step,
|
751 |
+
text_encoder1,
|
752 |
+
text_encoder2,
|
753 |
+
unet,
|
754 |
+
vae,
|
755 |
+
logit_scale,
|
756 |
+
ckpt_info,
|
757 |
+
)
|
758 |
+
print("model saved.")
|
759 |
+
|
760 |
+
|
761 |
+
def setup_parser() -> argparse.ArgumentParser:
|
762 |
+
parser = argparse.ArgumentParser()
|
763 |
+
|
764 |
+
train_util.add_sd_models_arguments(parser)
|
765 |
+
train_util.add_dataset_arguments(parser, True, True, True)
|
766 |
+
train_util.add_training_arguments(parser, False)
|
767 |
+
train_util.add_sd_saving_arguments(parser)
|
768 |
+
train_util.add_optimizer_arguments(parser)
|
769 |
+
config_util.add_config_arguments(parser)
|
770 |
+
custom_train_functions.add_custom_train_arguments(parser)
|
771 |
+
sdxl_train_util.add_sdxl_training_arguments(parser)
|
772 |
+
|
773 |
+
parser.add_argument("--diffusers_xformers", action="store_true", help="use xformers by diffusers / Diffusersでxformersを使用する")
|
774 |
+
parser.add_argument("--train_text_encoder", action="store_true", help="train text encoder / text encoderも学習する")
|
775 |
+
parser.add_argument(
|
776 |
+
"--no_half_vae",
|
777 |
+
action="store_true",
|
778 |
+
help="do not use fp16/bf16 VAE in mixed precision (use float VAE) / mixed precisionでも fp16/bf16 VAEを使わずfloat VAEを使う",
|
779 |
+
)
|
780 |
+
parser.add_argument(
|
781 |
+
"--block_lr",
|
782 |
+
type=str,
|
783 |
+
default=None,
|
784 |
+
help=f"learning rates for each block of U-Net, comma-separated, {UNET_NUM_BLOCKS_FOR_BLOCK_LR} values / "
|
785 |
+
+ f"U-Netの各ブロックの学習率、カンマ区切り、{UNET_NUM_BLOCKS_FOR_BLOCK_LR}個の値",
|
786 |
+
)
|
787 |
+
|
788 |
+
return parser
|
789 |
+
|
790 |
+
|
791 |
+
if __name__ == "__main__":
|
792 |
+
parser = setup_parser()
|
793 |
+
|
794 |
+
args = parser.parse_args()
|
795 |
+
args = train_util.read_config_from_file(args, parser)
|
796 |
+
|
797 |
+
train(args)
|
sdxl_train_util.py
ADDED
@@ -0,0 +1,391 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import gc
|
3 |
+
import math
|
4 |
+
import os
|
5 |
+
from typing import Optional
|
6 |
+
import torch
|
7 |
+
from accelerate import init_empty_weights
|
8 |
+
from tqdm import tqdm
|
9 |
+
from transformers import CLIPTokenizer
|
10 |
+
from library import model_util, sdxl_model_util, train_util, sdxl_original_unet
|
11 |
+
from library.sdxl_lpw_stable_diffusion import SdxlStableDiffusionLongPromptWeightingPipeline
|
12 |
+
|
13 |
+
TOKENIZER1_PATH = "openai/clip-vit-large-patch14"
|
14 |
+
TOKENIZER2_PATH = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
|
15 |
+
|
16 |
+
# DEFAULT_NOISE_OFFSET = 0.0357
|
17 |
+
|
18 |
+
|
19 |
+
def load_target_model(args, accelerator, model_version: str, weight_dtype):
|
20 |
+
# load models for each process
|
21 |
+
model_dtype = match_mixed_precision(args, weight_dtype) # prepare fp16/bf16
|
22 |
+
for pi in range(accelerator.state.num_processes):
|
23 |
+
if pi == accelerator.state.local_process_index:
|
24 |
+
print(f"loading model for process {accelerator.state.local_process_index}/{accelerator.state.num_processes}")
|
25 |
+
|
26 |
+
(
|
27 |
+
load_stable_diffusion_format,
|
28 |
+
text_encoder1,
|
29 |
+
text_encoder2,
|
30 |
+
vae,
|
31 |
+
unet,
|
32 |
+
logit_scale,
|
33 |
+
ckpt_info,
|
34 |
+
) = _load_target_model(
|
35 |
+
args.pretrained_model_name_or_path,
|
36 |
+
args.vae,
|
37 |
+
model_version,
|
38 |
+
weight_dtype,
|
39 |
+
accelerator.device if args.lowram else "cpu",
|
40 |
+
model_dtype,
|
41 |
+
)
|
42 |
+
|
43 |
+
# work on low-ram device
|
44 |
+
if args.lowram:
|
45 |
+
text_encoder1.to(accelerator.device)
|
46 |
+
text_encoder2.to(accelerator.device)
|
47 |
+
unet.to(accelerator.device)
|
48 |
+
vae.to(accelerator.device)
|
49 |
+
|
50 |
+
gc.collect()
|
51 |
+
torch.cuda.empty_cache()
|
52 |
+
accelerator.wait_for_everyone()
|
53 |
+
|
54 |
+
text_encoder1, text_encoder2, unet = train_util.transform_models_if_DDP([text_encoder1, text_encoder2, unet])
|
55 |
+
|
56 |
+
return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info
|
57 |
+
|
58 |
+
|
59 |
+
def _load_target_model(
|
60 |
+
name_or_path: str, vae_path: Optional[str], model_version: str, weight_dtype, device="cpu", model_dtype=None
|
61 |
+
):
|
62 |
+
# model_dtype only work with full fp16/bf16
|
63 |
+
name_or_path = os.readlink(name_or_path) if os.path.islink(name_or_path) else name_or_path
|
64 |
+
load_stable_diffusion_format = os.path.isfile(name_or_path) # determine SD or Diffusers
|
65 |
+
|
66 |
+
if load_stable_diffusion_format:
|
67 |
+
print(f"load StableDiffusion checkpoint: {name_or_path}")
|
68 |
+
(
|
69 |
+
text_encoder1,
|
70 |
+
text_encoder2,
|
71 |
+
vae,
|
72 |
+
unet,
|
73 |
+
logit_scale,
|
74 |
+
ckpt_info,
|
75 |
+
) = sdxl_model_util.load_models_from_sdxl_checkpoint(model_version, name_or_path, device, model_dtype)
|
76 |
+
else:
|
77 |
+
# Diffusers model is loaded to CPU
|
78 |
+
from diffusers import StableDiffusionXLPipeline
|
79 |
+
|
80 |
+
variant = "fp16" if weight_dtype == torch.float16 else None
|
81 |
+
print(f"load Diffusers pretrained models: {name_or_path}, variant={variant}")
|
82 |
+
try:
|
83 |
+
try:
|
84 |
+
pipe = StableDiffusionXLPipeline.from_pretrained(
|
85 |
+
name_or_path, torch_dtype=model_dtype, variant=variant, tokenizer=None
|
86 |
+
)
|
87 |
+
except EnvironmentError as ex:
|
88 |
+
if variant is not None:
|
89 |
+
print("try to load fp32 model")
|
90 |
+
pipe = StableDiffusionXLPipeline.from_pretrained(name_or_path, variant=None, tokenizer=None)
|
91 |
+
else:
|
92 |
+
raise ex
|
93 |
+
except EnvironmentError as ex:
|
94 |
+
print(
|
95 |
+
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}"
|
96 |
+
)
|
97 |
+
raise ex
|
98 |
+
|
99 |
+
text_encoder1 = pipe.text_encoder
|
100 |
+
text_encoder2 = pipe.text_encoder_2
|
101 |
+
|
102 |
+
# convert to fp32 for cache text_encoders outputs
|
103 |
+
if text_encoder1.dtype != torch.float32:
|
104 |
+
text_encoder1 = text_encoder1.to(dtype=torch.float32)
|
105 |
+
if text_encoder2.dtype != torch.float32:
|
106 |
+
text_encoder2 = text_encoder2.to(dtype=torch.float32)
|
107 |
+
|
108 |
+
vae = pipe.vae
|
109 |
+
unet = pipe.unet
|
110 |
+
del pipe
|
111 |
+
|
112 |
+
# Diffusers U-Net to original U-Net
|
113 |
+
state_dict = sdxl_model_util.convert_diffusers_unet_state_dict_to_sdxl(unet.state_dict())
|
114 |
+
with init_empty_weights():
|
115 |
+
unet = sdxl_original_unet.SdxlUNet2DConditionModel() # overwrite unet
|
116 |
+
sdxl_model_util._load_state_dict_on_device(unet, state_dict, device=device, dtype=model_dtype)
|
117 |
+
print("U-Net converted to original U-Net")
|
118 |
+
|
119 |
+
logit_scale = None
|
120 |
+
ckpt_info = None
|
121 |
+
|
122 |
+
# VAEを読み込む
|
123 |
+
if vae_path is not None:
|
124 |
+
vae = model_util.load_vae(vae_path, weight_dtype)
|
125 |
+
print("additional VAE loaded")
|
126 |
+
|
127 |
+
return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info
|
128 |
+
|
129 |
+
|
130 |
+
def load_tokenizers(args: argparse.Namespace):
|
131 |
+
print("prepare tokenizers")
|
132 |
+
|
133 |
+
original_paths = [TOKENIZER1_PATH, TOKENIZER2_PATH]
|
134 |
+
tokeniers = []
|
135 |
+
for i, original_path in enumerate(original_paths):
|
136 |
+
tokenizer: CLIPTokenizer = None
|
137 |
+
if args.tokenizer_cache_dir:
|
138 |
+
local_tokenizer_path = os.path.join(args.tokenizer_cache_dir, original_path.replace("/", "_"))
|
139 |
+
if os.path.exists(local_tokenizer_path):
|
140 |
+
print(f"load tokenizer from cache: {local_tokenizer_path}")
|
141 |
+
tokenizer = CLIPTokenizer.from_pretrained(local_tokenizer_path)
|
142 |
+
|
143 |
+
if tokenizer is None:
|
144 |
+
tokenizer = CLIPTokenizer.from_pretrained(original_path)
|
145 |
+
|
146 |
+
if args.tokenizer_cache_dir and not os.path.exists(local_tokenizer_path):
|
147 |
+
print(f"save Tokenizer to cache: {local_tokenizer_path}")
|
148 |
+
tokenizer.save_pretrained(local_tokenizer_path)
|
149 |
+
|
150 |
+
if i == 1:
|
151 |
+
tokenizer.pad_token_id = 0 # fix pad token id to make same as open clip tokenizer
|
152 |
+
|
153 |
+
tokeniers.append(tokenizer)
|
154 |
+
|
155 |
+
if hasattr(args, "max_token_length") and args.max_token_length is not None:
|
156 |
+
print(f"update token length: {args.max_token_length}")
|
157 |
+
|
158 |
+
return tokeniers
|
159 |
+
|
160 |
+
|
161 |
+
def match_mixed_precision(args, weight_dtype):
|
162 |
+
if args.full_fp16:
|
163 |
+
assert (
|
164 |
+
weight_dtype == torch.float16
|
165 |
+
), "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。"
|
166 |
+
return weight_dtype
|
167 |
+
elif args.full_bf16:
|
168 |
+
assert (
|
169 |
+
weight_dtype == torch.bfloat16
|
170 |
+
), "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。"
|
171 |
+
return weight_dtype
|
172 |
+
else:
|
173 |
+
return None
|
174 |
+
|
175 |
+
|
176 |
+
def timestep_embedding(timesteps, dim, max_period=10000):
|
177 |
+
"""
|
178 |
+
Create sinusoidal timestep embeddings.
|
179 |
+
:param timesteps: a 1-D Tensor of N indices, one per batch element.
|
180 |
+
These may be fractional.
|
181 |
+
:param dim: the dimension of the output.
|
182 |
+
:param max_period: controls the minimum frequency of the embeddings.
|
183 |
+
:return: an [N x dim] Tensor of positional embeddings.
|
184 |
+
"""
|
185 |
+
half = dim // 2
|
186 |
+
freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
|
187 |
+
device=timesteps.device
|
188 |
+
)
|
189 |
+
args = timesteps[:, None].float() * freqs[None]
|
190 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
191 |
+
if dim % 2:
|
192 |
+
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
193 |
+
return embedding
|
194 |
+
|
195 |
+
|
196 |
+
def get_timestep_embedding(x, outdim):
|
197 |
+
assert len(x.shape) == 2
|
198 |
+
b, dims = x.shape[0], x.shape[1]
|
199 |
+
x = torch.flatten(x)
|
200 |
+
emb = timestep_embedding(x, outdim)
|
201 |
+
emb = torch.reshape(emb, (b, dims * outdim))
|
202 |
+
return emb
|
203 |
+
|
204 |
+
|
205 |
+
def get_size_embeddings(orig_size, crop_size, target_size, device):
|
206 |
+
emb1 = get_timestep_embedding(orig_size, 256)
|
207 |
+
emb2 = get_timestep_embedding(crop_size, 256)
|
208 |
+
emb3 = get_timestep_embedding(target_size, 256)
|
209 |
+
vector = torch.cat([emb1, emb2, emb3], dim=1).to(device)
|
210 |
+
return vector
|
211 |
+
|
212 |
+
|
213 |
+
def save_sd_model_on_train_end(
|
214 |
+
args: argparse.Namespace,
|
215 |
+
src_path: str,
|
216 |
+
save_stable_diffusion_format: bool,
|
217 |
+
use_safetensors: bool,
|
218 |
+
save_dtype: torch.dtype,
|
219 |
+
epoch: int,
|
220 |
+
global_step: int,
|
221 |
+
text_encoder1,
|
222 |
+
text_encoder2,
|
223 |
+
unet,
|
224 |
+
vae,
|
225 |
+
logit_scale,
|
226 |
+
ckpt_info,
|
227 |
+
):
|
228 |
+
def sd_saver(ckpt_file, epoch_no, global_step):
|
229 |
+
sai_metadata = train_util.get_sai_model_spec(None, args, True, False, False, is_stable_diffusion_ckpt=True)
|
230 |
+
sdxl_model_util.save_stable_diffusion_checkpoint(
|
231 |
+
ckpt_file,
|
232 |
+
text_encoder1,
|
233 |
+
text_encoder2,
|
234 |
+
unet,
|
235 |
+
epoch_no,
|
236 |
+
global_step,
|
237 |
+
ckpt_info,
|
238 |
+
vae,
|
239 |
+
logit_scale,
|
240 |
+
sai_metadata,
|
241 |
+
save_dtype,
|
242 |
+
)
|
243 |
+
|
244 |
+
def diffusers_saver(out_dir):
|
245 |
+
sdxl_model_util.save_diffusers_checkpoint(
|
246 |
+
out_dir,
|
247 |
+
text_encoder1,
|
248 |
+
text_encoder2,
|
249 |
+
unet,
|
250 |
+
src_path,
|
251 |
+
vae,
|
252 |
+
use_safetensors=use_safetensors,
|
253 |
+
save_dtype=save_dtype,
|
254 |
+
)
|
255 |
+
|
256 |
+
train_util.save_sd_model_on_train_end_common(
|
257 |
+
args, save_stable_diffusion_format, use_safetensors, epoch, global_step, sd_saver, diffusers_saver
|
258 |
+
)
|
259 |
+
|
260 |
+
|
261 |
+
# 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
|
262 |
+
# on_epoch_end: Trueならepoch終了時、Falseならstep経過時
|
263 |
+
def save_sd_model_on_epoch_end_or_stepwise(
|
264 |
+
args: argparse.Namespace,
|
265 |
+
on_epoch_end: bool,
|
266 |
+
accelerator,
|
267 |
+
src_path,
|
268 |
+
save_stable_diffusion_format: bool,
|
269 |
+
use_safetensors: bool,
|
270 |
+
save_dtype: torch.dtype,
|
271 |
+
epoch: int,
|
272 |
+
num_train_epochs: int,
|
273 |
+
global_step: int,
|
274 |
+
text_encoder1,
|
275 |
+
text_encoder2,
|
276 |
+
unet,
|
277 |
+
vae,
|
278 |
+
logit_scale,
|
279 |
+
ckpt_info,
|
280 |
+
ema = None,
|
281 |
+
params_to_replace = None,
|
282 |
+
):
|
283 |
+
def sd_saver(ckpt_file, epoch_no, global_step):
|
284 |
+
sai_metadata = train_util.get_sai_model_spec(None, args, True, False, False, is_stable_diffusion_ckpt=True)
|
285 |
+
sdxl_model_util.save_stable_diffusion_checkpoint(
|
286 |
+
ckpt_file,
|
287 |
+
text_encoder1,
|
288 |
+
text_encoder2,
|
289 |
+
unet,
|
290 |
+
epoch_no,
|
291 |
+
global_step,
|
292 |
+
ckpt_info,
|
293 |
+
vae,
|
294 |
+
logit_scale,
|
295 |
+
sai_metadata,
|
296 |
+
save_dtype,
|
297 |
+
)
|
298 |
+
|
299 |
+
def diffusers_saver(out_dir):
|
300 |
+
sdxl_model_util.save_diffusers_checkpoint(
|
301 |
+
out_dir,
|
302 |
+
text_encoder1,
|
303 |
+
text_encoder2,
|
304 |
+
unet,
|
305 |
+
src_path,
|
306 |
+
vae,
|
307 |
+
use_safetensors=use_safetensors,
|
308 |
+
save_dtype=save_dtype,
|
309 |
+
)
|
310 |
+
|
311 |
+
if args.enable_ema and not args.ema_save_only_ema_weights and ema:
|
312 |
+
temp_name = args.output_name
|
313 |
+
args.output_name = args.output_name + "-non-EMA"
|
314 |
+
|
315 |
+
train_util.save_sd_model_on_epoch_end_or_stepwise_common(
|
316 |
+
args,
|
317 |
+
on_epoch_end,
|
318 |
+
accelerator,
|
319 |
+
save_stable_diffusion_format,
|
320 |
+
use_safetensors,
|
321 |
+
epoch,
|
322 |
+
num_train_epochs,
|
323 |
+
global_step,
|
324 |
+
sd_saver,
|
325 |
+
diffusers_saver,
|
326 |
+
)
|
327 |
+
args.output_name = temp_name if temp_name else args.output_name
|
328 |
+
if args.enable_ema and ema:
|
329 |
+
with ema.ema_parameters(params_to_replace):
|
330 |
+
print("Saving EMA:")
|
331 |
+
train_util.save_sd_model_on_epoch_end_or_stepwise_common(
|
332 |
+
args,
|
333 |
+
on_epoch_end,
|
334 |
+
accelerator,
|
335 |
+
save_stable_diffusion_format,
|
336 |
+
use_safetensors,
|
337 |
+
epoch,
|
338 |
+
num_train_epochs,
|
339 |
+
global_step,
|
340 |
+
sd_saver,
|
341 |
+
diffusers_saver,
|
342 |
+
)
|
343 |
+
|
344 |
+
|
345 |
+
def add_sdxl_training_arguments(parser: argparse.ArgumentParser):
|
346 |
+
parser.add_argument(
|
347 |
+
"--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"
|
348 |
+
)
|
349 |
+
parser.add_argument(
|
350 |
+
"--cache_text_encoder_outputs_to_disk",
|
351 |
+
action="store_true",
|
352 |
+
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",
|
353 |
+
)
|
354 |
+
|
355 |
+
|
356 |
+
def verify_sdxl_training_args(args: argparse.Namespace, supportTextEncoderCaching: bool = True):
|
357 |
+
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"
|
358 |
+
if args.v_parameterization:
|
359 |
+
print("v_parameterization will be unexpected / SDXL学習ではv_parameterizationは想定外の動作になります")
|
360 |
+
|
361 |
+
if args.clip_skip is not None:
|
362 |
+
print("clip_skip will be unexpected / SDXL学習ではclip_skipは動作しません")
|
363 |
+
|
364 |
+
# if args.multires_noise_iterations:
|
365 |
+
# print(
|
366 |
+
# 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"
|
367 |
+
# )
|
368 |
+
# else:
|
369 |
+
# if args.noise_offset is None:
|
370 |
+
# args.noise_offset = DEFAULT_NOISE_OFFSET
|
371 |
+
# elif args.noise_offset != DEFAULT_NOISE_OFFSET:
|
372 |
+
# print(
|
373 |
+
# 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"
|
374 |
+
# )
|
375 |
+
# print(f"noise_offset is set to {args.noise_offset} / noise_offsetが{args.noise_offset}に設定されました")
|
376 |
+
|
377 |
+
assert (
|
378 |
+
not hasattr(args, "weighted_captions") or not args.weighted_captions
|
379 |
+
), "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"
|
380 |
+
|
381 |
+
if supportTextEncoderCaching:
|
382 |
+
if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs:
|
383 |
+
args.cache_text_encoder_outputs = True
|
384 |
+
print(
|
385 |
+
"cache_text_encoder_outputs is enabled because cache_text_encoder_outputs_to_disk is enabled / "
|
386 |
+
+ "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"
|
387 |
+
)
|
388 |
+
|
389 |
+
|
390 |
+
def sample_images(*args, **kwargs):
|
391 |
+
return train_util.sample_images_common(SdxlStableDiffusionLongPromptWeightingPipeline, *args, **kwargs)
|
train_util.py
ADDED
The diff for this file is too large to render.
See raw diff
|
|