Spaces:
Runtime error
Runtime error
RohitGandikota
commited on
Commit
โข
47a88ae
1
Parent(s):
6491cdf
pushing training code
Browse files- __init__.py +2 -1
- app.py +27 -9
- trainscripts/textsliders/data/config-xl.yaml +1 -1
- trainscripts/textsliders/data/prompts-xl.yaml +27 -18
- trainscripts/textsliders/demotrain.py +434 -0
- trainscripts/textsliders/prompt_util.py +10 -1
__init__.py
CHANGED
@@ -1 +1,2 @@
|
|
1 |
-
from trainscripts.textsliders import lora
|
|
|
|
1 |
+
from trainscripts.textsliders import lora
|
2 |
+
from trainscripts.textsliders import demotrain
|
app.py
CHANGED
@@ -6,7 +6,7 @@ from diffusers.pipelines import StableDiffusionXLPipeline
|
|
6 |
StableDiffusionXLPipeline.__call__ = call
|
7 |
import os
|
8 |
from trainscripts.textsliders.lora import LoRANetwork, DEFAULT_TARGET_REPLACE, UNET_TARGET_REPLACE_MODULE_CONV
|
9 |
-
|
10 |
|
11 |
os.environ['CURL_CA_BUNDLE'] = ''
|
12 |
model_map = {'Age' : 'models/age.pt',
|
@@ -204,10 +204,26 @@ class Demo:
|
|
204 |
)
|
205 |
|
206 |
def train(self, target_concept,positive_prompt, negative_prompt, rank, iterations_input, lr_input, train_method, neg_guidance, iterations, lr, pbar = gr.Progress(track_tqdm=True)):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
207 |
|
208 |
-
|
209 |
-
# return [gr.update(interactive=True, value='Train'), gr.update(value='Someone else is training... Try again soon'), None, gr.update()]
|
210 |
|
|
|
|
|
|
|
|
|
211 |
# if train_method == 'ESD-x':
|
212 |
|
213 |
# modules = ".*attn2$"
|
@@ -223,7 +239,7 @@ class Demo:
|
|
223 |
# modules = ".*attn1$"
|
224 |
# frozen = []
|
225 |
|
226 |
-
#
|
227 |
|
228 |
# save_path = f"models/{randn}_{prompt.lower().replace(' ', '')}.pt"
|
229 |
|
@@ -237,7 +253,7 @@ class Demo:
|
|
237 |
|
238 |
# model_map['Custom'] = save_path
|
239 |
|
240 |
-
#
|
241 |
return [None, None, None, None]
|
242 |
|
243 |
def inference(self, prompt, seed, start_noise, scale, model_name, pbar = gr.Progress(track_tqdm=True)):
|
@@ -267,10 +283,12 @@ class Demo:
|
|
267 |
name = os.path.basename(model_path)
|
268 |
rank = 4
|
269 |
alpha = 1
|
270 |
-
if
|
271 |
-
rank =
|
272 |
-
|
273 |
-
|
|
|
|
|
274 |
if 'alpha1' in model_path:
|
275 |
alpha = 1.0
|
276 |
network = LoRANetwork(
|
|
|
6 |
StableDiffusionXLPipeline.__call__ = call
|
7 |
import os
|
8 |
from trainscripts.textsliders.lora import LoRANetwork, DEFAULT_TARGET_REPLACE, UNET_TARGET_REPLACE_MODULE_CONV
|
9 |
+
from trainscripts.textsliders.demotrain import train_xl
|
10 |
|
11 |
os.environ['CURL_CA_BUNDLE'] = ''
|
12 |
model_map = {'Age' : 'models/age.pt',
|
|
|
204 |
)
|
205 |
|
206 |
def train(self, target_concept,positive_prompt, negative_prompt, rank, iterations_input, lr_input, train_method, neg_guidance, iterations, lr, pbar = gr.Progress(track_tqdm=True)):
|
207 |
+
|
208 |
+
|
209 |
+
randn = torch.randint(1, 10000000, (1,)).item()
|
210 |
+
save_name = f'{randn}_{target_concept.replace(',','').replace(' ','').replace('.','')[:10]}_{positive_prompt.replace(',','').replace(' ','').replace('.','')[:10]}'
|
211 |
+
save_name += f'_alpha-{1}'
|
212 |
+
save_name += f'_noxattn'
|
213 |
+
save_name += f'_rank_{rank}.pt'
|
214 |
+
|
215 |
+
if self.training:
|
216 |
+
return [gr.update(interactive=True, value='Train'), gr.update(value='Someone else is training... Try again soon'), None, gr.update()]
|
217 |
+
|
218 |
+
self.training = True
|
219 |
+
train_xl(target, postive, negative, lr, iterations, config_file, rank, device, attributes)
|
220 |
|
221 |
+
self.training = False
|
|
|
222 |
|
223 |
+
torch.cuda.empty_cache()
|
224 |
+
model_map['Custom Slider'] = f'models/{save_name}'
|
225 |
+
|
226 |
+
return [gr.update(interactive=True, value='Train'), gr.update(value='Done Training! \n Try your custom slider in the "Test" tab'), save_path, gr.Dropdown.update(choices=list(model_map.keys()), value='Custom Slider')]
|
227 |
# if train_method == 'ESD-x':
|
228 |
|
229 |
# modules = ".*attn2$"
|
|
|
239 |
# modules = ".*attn1$"
|
240 |
# frozen = []
|
241 |
|
242 |
+
#
|
243 |
|
244 |
# save_path = f"models/{randn}_{prompt.lower().replace(' ', '')}.pt"
|
245 |
|
|
|
253 |
|
254 |
# model_map['Custom'] = save_path
|
255 |
|
256 |
+
#
|
257 |
return [None, None, None, None]
|
258 |
|
259 |
def inference(self, prompt, seed, start_noise, scale, model_name, pbar = gr.Progress(track_tqdm=True)):
|
|
|
283 |
name = os.path.basename(model_path)
|
284 |
rank = 4
|
285 |
alpha = 1
|
286 |
+
if rank in model_path:
|
287 |
+
rank = int(model_path.split('_')[-1].replace('.pt',''))
|
288 |
+
# if 'rank4' in model_path:
|
289 |
+
# rank = 4
|
290 |
+
# if 'rank8' in model_path:
|
291 |
+
# rank = 8
|
292 |
if 'alpha1' in model_path:
|
293 |
alpha = 1.0
|
294 |
network = LoRANetwork(
|
trainscripts/textsliders/data/config-xl.yaml
CHANGED
@@ -19,7 +19,7 @@ train:
|
|
19 |
save:
|
20 |
name: "temp"
|
21 |
path: "./models"
|
22 |
-
per_steps:
|
23 |
precision: "bfloat16"
|
24 |
logging:
|
25 |
use_wandb: false
|
|
|
19 |
save:
|
20 |
name: "temp"
|
21 |
path: "./models"
|
22 |
+
per_steps: 5000000
|
23 |
precision: "bfloat16"
|
24 |
logging:
|
25 |
use_wandb: false
|
trainscripts/textsliders/data/prompts-xl.yaml
CHANGED
@@ -1,3 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
####################################################################################################### AGE SLIDER
|
2 |
# - target: "male person" # what word for erasing the positive concept from
|
3 |
# positive: "male person, very old" # concept to erase
|
@@ -257,24 +266,24 @@
|
|
257 |
# dynamic_resolution: false
|
258 |
# batch_size: 1
|
259 |
####################################################################################################### SCULPTURE SLIDER
|
260 |
-
- target: "male person" # what word for erasing the positive concept from
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
- target: "female person" # what word for erasing the positive concept from
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
####################################################################################################### METAL SLIDER
|
279 |
# - target: "" # what word for erasing the positive concept from
|
280 |
# positive: "made out of metal, metallic style, iron, copper, platinum metal," # concept to erase
|
|
|
1 |
+
- target: "" # what word for erasing the positive concept from
|
2 |
+
positive: "" # concept to erase
|
3 |
+
unconditional: "" # word to take the difference from the positive concept
|
4 |
+
neutral: "" # starting point for conditioning the target
|
5 |
+
action: "enhance" # erase or enhance
|
6 |
+
guidance_scale: 4
|
7 |
+
resolution: 512
|
8 |
+
dynamic_resolution: false
|
9 |
+
batch_size: 1
|
10 |
####################################################################################################### AGE SLIDER
|
11 |
# - target: "male person" # what word for erasing the positive concept from
|
12 |
# positive: "male person, very old" # concept to erase
|
|
|
266 |
# dynamic_resolution: false
|
267 |
# batch_size: 1
|
268 |
####################################################################################################### SCULPTURE SLIDER
|
269 |
+
# - target: "male person" # what word for erasing the positive concept from
|
270 |
+
# positive: "male person, cement sculpture, cement greek statue style" # concept to erase
|
271 |
+
# unconditional: "male person, realistic, hyper realistic" # word to take the difference from the positive concept
|
272 |
+
# neutral: "male person" # starting point for conditioning the target
|
273 |
+
# action: "enhance" # erase or enhance
|
274 |
+
# guidance_scale: 4
|
275 |
+
# resolution: 512
|
276 |
+
# dynamic_resolution: false
|
277 |
+
# batch_size: 1
|
278 |
+
# - target: "female person" # what word for erasing the positive concept from
|
279 |
+
# positive: "female person, cement sculpture, cement greek statue style" # concept to erase
|
280 |
+
# unconditional: "female person, realistic, hyper realistic" # word to take the difference from the positive concept
|
281 |
+
# neutral: "female person" # starting point for conditioning the target
|
282 |
+
# action: "enhance" # erase or enhance
|
283 |
+
# guidance_scale: 4
|
284 |
+
# resolution: 512
|
285 |
+
# dynamic_resolution: false
|
286 |
+
# batch_size: 1
|
287 |
####################################################################################################### METAL SLIDER
|
288 |
# - target: "" # what word for erasing the positive concept from
|
289 |
# positive: "made out of metal, metallic style, iron, copper, platinum metal," # concept to erase
|
trainscripts/textsliders/demotrain.py
ADDED
@@ -0,0 +1,434 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# ref:
|
2 |
+
# - https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py#L566
|
3 |
+
# - https://huggingface.co/spaces/baulab/Erasing-Concepts-In-Diffusion/blob/main/train.py
|
4 |
+
|
5 |
+
from typing import List, Optional
|
6 |
+
import argparse
|
7 |
+
import ast
|
8 |
+
from pathlib import Path
|
9 |
+
import gc
|
10 |
+
|
11 |
+
import torch
|
12 |
+
from tqdm import tqdm
|
13 |
+
|
14 |
+
|
15 |
+
from lora import LoRANetwork, DEFAULT_TARGET_REPLACE, UNET_TARGET_REPLACE_MODULE_CONV
|
16 |
+
import train_util
|
17 |
+
import model_util
|
18 |
+
import prompt_util
|
19 |
+
from prompt_util import (
|
20 |
+
PromptEmbedsCache,
|
21 |
+
PromptEmbedsPair,
|
22 |
+
PromptSettings,
|
23 |
+
PromptEmbedsXL,
|
24 |
+
)
|
25 |
+
import debug_util
|
26 |
+
import config_util
|
27 |
+
from config_util import RootConfig
|
28 |
+
|
29 |
+
import wandb
|
30 |
+
|
31 |
+
NUM_IMAGES_PER_PROMPT = 1
|
32 |
+
|
33 |
+
|
34 |
+
def flush():
|
35 |
+
torch.cuda.empty_cache()
|
36 |
+
gc.collect()
|
37 |
+
|
38 |
+
|
39 |
+
def train(
|
40 |
+
config: RootConfig,
|
41 |
+
prompts: list[PromptSettings],
|
42 |
+
device,
|
43 |
+
):
|
44 |
+
metadata = {
|
45 |
+
"prompts": ",".join([prompt.json() for prompt in prompts]),
|
46 |
+
"config": config.json(),
|
47 |
+
}
|
48 |
+
save_path = Path(config.save.path)
|
49 |
+
|
50 |
+
modules = DEFAULT_TARGET_REPLACE
|
51 |
+
if config.network.type == "c3lier":
|
52 |
+
modules += UNET_TARGET_REPLACE_MODULE_CONV
|
53 |
+
|
54 |
+
if config.logging.verbose:
|
55 |
+
print(metadata)
|
56 |
+
|
57 |
+
if config.logging.use_wandb:
|
58 |
+
wandb.init(project=f"LECO_{config.save.name}", config=metadata)
|
59 |
+
|
60 |
+
weight_dtype = config_util.parse_precision(config.train.precision)
|
61 |
+
save_weight_dtype = config_util.parse_precision(config.train.precision)
|
62 |
+
|
63 |
+
(
|
64 |
+
tokenizers,
|
65 |
+
text_encoders,
|
66 |
+
unet,
|
67 |
+
noise_scheduler,
|
68 |
+
) = model_util.load_models_xl(
|
69 |
+
config.pretrained_model.name_or_path,
|
70 |
+
scheduler_name=config.train.noise_scheduler,
|
71 |
+
)
|
72 |
+
|
73 |
+
for text_encoder in text_encoders:
|
74 |
+
text_encoder.to(device, dtype=weight_dtype)
|
75 |
+
text_encoder.requires_grad_(False)
|
76 |
+
text_encoder.eval()
|
77 |
+
|
78 |
+
unet.to(device, dtype=weight_dtype)
|
79 |
+
if config.other.use_xformers:
|
80 |
+
unet.enable_xformers_memory_efficient_attention()
|
81 |
+
unet.requires_grad_(False)
|
82 |
+
unet.eval()
|
83 |
+
|
84 |
+
network = LoRANetwork(
|
85 |
+
unet,
|
86 |
+
rank=config.network.rank,
|
87 |
+
multiplier=1.0,
|
88 |
+
alpha=config.network.alpha,
|
89 |
+
train_method=config.network.training_method,
|
90 |
+
).to(device, dtype=weight_dtype)
|
91 |
+
|
92 |
+
optimizer_module = train_util.get_optimizer(config.train.optimizer)
|
93 |
+
#optimizer_args
|
94 |
+
optimizer_kwargs = {}
|
95 |
+
if config.train.optimizer_args is not None and len(config.train.optimizer_args) > 0:
|
96 |
+
for arg in config.train.optimizer_args.split(" "):
|
97 |
+
key, value = arg.split("=")
|
98 |
+
value = ast.literal_eval(value)
|
99 |
+
optimizer_kwargs[key] = value
|
100 |
+
|
101 |
+
optimizer = optimizer_module(network.prepare_optimizer_params(), lr=config.train.lr, **optimizer_kwargs)
|
102 |
+
lr_scheduler = train_util.get_lr_scheduler(
|
103 |
+
config.train.lr_scheduler,
|
104 |
+
optimizer,
|
105 |
+
max_iterations=config.train.iterations,
|
106 |
+
lr_min=config.train.lr / 100,
|
107 |
+
)
|
108 |
+
criteria = torch.nn.MSELoss()
|
109 |
+
|
110 |
+
print("Prompts")
|
111 |
+
for settings in prompts:
|
112 |
+
print(settings)
|
113 |
+
|
114 |
+
# debug
|
115 |
+
debug_util.check_requires_grad(network)
|
116 |
+
debug_util.check_training_mode(network)
|
117 |
+
|
118 |
+
cache = PromptEmbedsCache()
|
119 |
+
prompt_pairs: list[PromptEmbedsPair] = []
|
120 |
+
|
121 |
+
with torch.no_grad():
|
122 |
+
for settings in prompts:
|
123 |
+
print(settings)
|
124 |
+
for prompt in [
|
125 |
+
settings.target,
|
126 |
+
settings.positive,
|
127 |
+
settings.neutral,
|
128 |
+
settings.unconditional,
|
129 |
+
]:
|
130 |
+
if cache[prompt] == None:
|
131 |
+
tex_embs, pool_embs = train_util.encode_prompts_xl(
|
132 |
+
tokenizers,
|
133 |
+
text_encoders,
|
134 |
+
[prompt],
|
135 |
+
num_images_per_prompt=NUM_IMAGES_PER_PROMPT,
|
136 |
+
)
|
137 |
+
cache[prompt] = PromptEmbedsXL(
|
138 |
+
tex_embs,
|
139 |
+
pool_embs
|
140 |
+
)
|
141 |
+
|
142 |
+
prompt_pairs.append(
|
143 |
+
PromptEmbedsPair(
|
144 |
+
criteria,
|
145 |
+
cache[settings.target],
|
146 |
+
cache[settings.positive],
|
147 |
+
cache[settings.unconditional],
|
148 |
+
cache[settings.neutral],
|
149 |
+
settings,
|
150 |
+
)
|
151 |
+
)
|
152 |
+
|
153 |
+
for tokenizer, text_encoder in zip(tokenizers, text_encoders):
|
154 |
+
del tokenizer, text_encoder
|
155 |
+
|
156 |
+
flush()
|
157 |
+
|
158 |
+
pbar = tqdm(range(config.train.iterations))
|
159 |
+
|
160 |
+
loss = None
|
161 |
+
|
162 |
+
for i in pbar:
|
163 |
+
with torch.no_grad():
|
164 |
+
noise_scheduler.set_timesteps(
|
165 |
+
config.train.max_denoising_steps, device=device
|
166 |
+
)
|
167 |
+
|
168 |
+
optimizer.zero_grad()
|
169 |
+
|
170 |
+
prompt_pair: PromptEmbedsPair = prompt_pairs[
|
171 |
+
torch.randint(0, len(prompt_pairs), (1,)).item()
|
172 |
+
]
|
173 |
+
|
174 |
+
# 1 ~ 49 ใใใฉใณใใ
|
175 |
+
timesteps_to = torch.randint(
|
176 |
+
1, config.train.max_denoising_steps, (1,)
|
177 |
+
).item()
|
178 |
+
|
179 |
+
height, width = prompt_pair.resolution, prompt_pair.resolution
|
180 |
+
if prompt_pair.dynamic_resolution:
|
181 |
+
height, width = train_util.get_random_resolution_in_bucket(
|
182 |
+
prompt_pair.resolution
|
183 |
+
)
|
184 |
+
|
185 |
+
if config.logging.verbose:
|
186 |
+
print("gudance_scale:", prompt_pair.guidance_scale)
|
187 |
+
print("resolution:", prompt_pair.resolution)
|
188 |
+
print("dynamic_resolution:", prompt_pair.dynamic_resolution)
|
189 |
+
if prompt_pair.dynamic_resolution:
|
190 |
+
print("bucketed resolution:", (height, width))
|
191 |
+
print("batch_size:", prompt_pair.batch_size)
|
192 |
+
print("dynamic_crops:", prompt_pair.dynamic_crops)
|
193 |
+
|
194 |
+
latents = train_util.get_initial_latents(
|
195 |
+
noise_scheduler, prompt_pair.batch_size, height, width, 1
|
196 |
+
).to(device, dtype=weight_dtype)
|
197 |
+
|
198 |
+
add_time_ids = train_util.get_add_time_ids(
|
199 |
+
height,
|
200 |
+
width,
|
201 |
+
dynamic_crops=prompt_pair.dynamic_crops,
|
202 |
+
dtype=weight_dtype,
|
203 |
+
).to(device, dtype=weight_dtype)
|
204 |
+
|
205 |
+
with network:
|
206 |
+
# ใกใใฃใจใใใคใบใใใใใใฎใ่ฟใ
|
207 |
+
denoised_latents = train_util.diffusion_xl(
|
208 |
+
unet,
|
209 |
+
noise_scheduler,
|
210 |
+
latents, # ๅ็ดใชใใคใบใฎlatentsใๆธกใ
|
211 |
+
text_embeddings=train_util.concat_embeddings(
|
212 |
+
prompt_pair.unconditional.text_embeds,
|
213 |
+
prompt_pair.target.text_embeds,
|
214 |
+
prompt_pair.batch_size,
|
215 |
+
),
|
216 |
+
add_text_embeddings=train_util.concat_embeddings(
|
217 |
+
prompt_pair.unconditional.pooled_embeds,
|
218 |
+
prompt_pair.target.pooled_embeds,
|
219 |
+
prompt_pair.batch_size,
|
220 |
+
),
|
221 |
+
add_time_ids=train_util.concat_embeddings(
|
222 |
+
add_time_ids, add_time_ids, prompt_pair.batch_size
|
223 |
+
),
|
224 |
+
start_timesteps=0,
|
225 |
+
total_timesteps=timesteps_to,
|
226 |
+
guidance_scale=3,
|
227 |
+
)
|
228 |
+
|
229 |
+
noise_scheduler.set_timesteps(1000)
|
230 |
+
|
231 |
+
current_timestep = noise_scheduler.timesteps[
|
232 |
+
int(timesteps_to * 1000 / config.train.max_denoising_steps)
|
233 |
+
]
|
234 |
+
|
235 |
+
# with network: ใฎๅคใงใฏ็ฉบใฎLoRAใฎใฟใๆๅนใซใชใ
|
236 |
+
positive_latents = train_util.predict_noise_xl(
|
237 |
+
unet,
|
238 |
+
noise_scheduler,
|
239 |
+
current_timestep,
|
240 |
+
denoised_latents,
|
241 |
+
text_embeddings=train_util.concat_embeddings(
|
242 |
+
prompt_pair.unconditional.text_embeds,
|
243 |
+
prompt_pair.positive.text_embeds,
|
244 |
+
prompt_pair.batch_size,
|
245 |
+
),
|
246 |
+
add_text_embeddings=train_util.concat_embeddings(
|
247 |
+
prompt_pair.unconditional.pooled_embeds,
|
248 |
+
prompt_pair.positive.pooled_embeds,
|
249 |
+
prompt_pair.batch_size,
|
250 |
+
),
|
251 |
+
add_time_ids=train_util.concat_embeddings(
|
252 |
+
add_time_ids, add_time_ids, prompt_pair.batch_size
|
253 |
+
),
|
254 |
+
guidance_scale=1,
|
255 |
+
).to(device, dtype=weight_dtype)
|
256 |
+
neutral_latents = train_util.predict_noise_xl(
|
257 |
+
unet,
|
258 |
+
noise_scheduler,
|
259 |
+
current_timestep,
|
260 |
+
denoised_latents,
|
261 |
+
text_embeddings=train_util.concat_embeddings(
|
262 |
+
prompt_pair.unconditional.text_embeds,
|
263 |
+
prompt_pair.neutral.text_embeds,
|
264 |
+
prompt_pair.batch_size,
|
265 |
+
),
|
266 |
+
add_text_embeddings=train_util.concat_embeddings(
|
267 |
+
prompt_pair.unconditional.pooled_embeds,
|
268 |
+
prompt_pair.neutral.pooled_embeds,
|
269 |
+
prompt_pair.batch_size,
|
270 |
+
),
|
271 |
+
add_time_ids=train_util.concat_embeddings(
|
272 |
+
add_time_ids, add_time_ids, prompt_pair.batch_size
|
273 |
+
),
|
274 |
+
guidance_scale=1,
|
275 |
+
).to(device, dtype=weight_dtype)
|
276 |
+
unconditional_latents = train_util.predict_noise_xl(
|
277 |
+
unet,
|
278 |
+
noise_scheduler,
|
279 |
+
current_timestep,
|
280 |
+
denoised_latents,
|
281 |
+
text_embeddings=train_util.concat_embeddings(
|
282 |
+
prompt_pair.unconditional.text_embeds,
|
283 |
+
prompt_pair.unconditional.text_embeds,
|
284 |
+
prompt_pair.batch_size,
|
285 |
+
),
|
286 |
+
add_text_embeddings=train_util.concat_embeddings(
|
287 |
+
prompt_pair.unconditional.pooled_embeds,
|
288 |
+
prompt_pair.unconditional.pooled_embeds,
|
289 |
+
prompt_pair.batch_size,
|
290 |
+
),
|
291 |
+
add_time_ids=train_util.concat_embeddings(
|
292 |
+
add_time_ids, add_time_ids, prompt_pair.batch_size
|
293 |
+
),
|
294 |
+
guidance_scale=1,
|
295 |
+
).to(device, dtype=weight_dtype)
|
296 |
+
|
297 |
+
if config.logging.verbose:
|
298 |
+
print("positive_latents:", positive_latents[0, 0, :5, :5])
|
299 |
+
print("neutral_latents:", neutral_latents[0, 0, :5, :5])
|
300 |
+
print("unconditional_latents:", unconditional_latents[0, 0, :5, :5])
|
301 |
+
|
302 |
+
with network:
|
303 |
+
target_latents = train_util.predict_noise_xl(
|
304 |
+
unet,
|
305 |
+
noise_scheduler,
|
306 |
+
current_timestep,
|
307 |
+
denoised_latents,
|
308 |
+
text_embeddings=train_util.concat_embeddings(
|
309 |
+
prompt_pair.unconditional.text_embeds,
|
310 |
+
prompt_pair.target.text_embeds,
|
311 |
+
prompt_pair.batch_size,
|
312 |
+
),
|
313 |
+
add_text_embeddings=train_util.concat_embeddings(
|
314 |
+
prompt_pair.unconditional.pooled_embeds,
|
315 |
+
prompt_pair.target.pooled_embeds,
|
316 |
+
prompt_pair.batch_size,
|
317 |
+
),
|
318 |
+
add_time_ids=train_util.concat_embeddings(
|
319 |
+
add_time_ids, add_time_ids, prompt_pair.batch_size
|
320 |
+
),
|
321 |
+
guidance_scale=1,
|
322 |
+
).to(device, dtype=weight_dtype)
|
323 |
+
|
324 |
+
if config.logging.verbose:
|
325 |
+
print("target_latents:", target_latents[0, 0, :5, :5])
|
326 |
+
|
327 |
+
positive_latents.requires_grad = False
|
328 |
+
neutral_latents.requires_grad = False
|
329 |
+
unconditional_latents.requires_grad = False
|
330 |
+
|
331 |
+
loss = prompt_pair.loss(
|
332 |
+
target_latents=target_latents,
|
333 |
+
positive_latents=positive_latents,
|
334 |
+
neutral_latents=neutral_latents,
|
335 |
+
unconditional_latents=unconditional_latents,
|
336 |
+
)
|
337 |
+
|
338 |
+
# 1000ๅใใชใใจใใฃใจ0.000...ใซใชใฃใฆใใพใฃใฆ่ฆใ็ฎ็ใซ้ข็ฝใใชใ
|
339 |
+
pbar.set_description(f"Loss*1k: {loss.item()*1000:.4f}")
|
340 |
+
if config.logging.use_wandb:
|
341 |
+
wandb.log(
|
342 |
+
{"loss": loss, "iteration": i, "lr": lr_scheduler.get_last_lr()[0]}
|
343 |
+
)
|
344 |
+
|
345 |
+
loss.backward()
|
346 |
+
optimizer.step()
|
347 |
+
lr_scheduler.step()
|
348 |
+
|
349 |
+
del (
|
350 |
+
positive_latents,
|
351 |
+
neutral_latents,
|
352 |
+
unconditional_latents,
|
353 |
+
target_latents,
|
354 |
+
latents,
|
355 |
+
)
|
356 |
+
flush()
|
357 |
+
|
358 |
+
# if (
|
359 |
+
# i % config.save.per_steps == 0
|
360 |
+
# and i != 0
|
361 |
+
# and i != config.train.iterations - 1
|
362 |
+
# ):
|
363 |
+
# print("Saving...")
|
364 |
+
# save_path.mkdir(parents=True, exist_ok=True)
|
365 |
+
# network.save_weights(
|
366 |
+
# save_path / f"{config.save.name}_{i}steps.pt",
|
367 |
+
# dtype=save_weight_dtype,
|
368 |
+
# )
|
369 |
+
|
370 |
+
print("Saving...")
|
371 |
+
save_path.mkdir(parents=True, exist_ok=True)
|
372 |
+
network.save_weights(
|
373 |
+
save_path / f"{config.save.name}",
|
374 |
+
dtype=save_weight_dtype,
|
375 |
+
)
|
376 |
+
|
377 |
+
del (
|
378 |
+
unet,
|
379 |
+
noise_scheduler,
|
380 |
+
loss,
|
381 |
+
optimizer,
|
382 |
+
network,
|
383 |
+
)
|
384 |
+
|
385 |
+
flush()
|
386 |
+
|
387 |
+
print("Done.")
|
388 |
+
|
389 |
+
|
390 |
+
# def main(args):
|
391 |
+
# config_file = args.config_file
|
392 |
+
|
393 |
+
# config = config_util.load_config_from_yaml(config_file)
|
394 |
+
# if args.name is not None:
|
395 |
+
# config.save.name = args.name
|
396 |
+
# attributes = []
|
397 |
+
# if args.attributes is not None:
|
398 |
+
# attributes = args.attributes.split(',')
|
399 |
+
# attributes = [a.strip() for a in attributes]
|
400 |
+
|
401 |
+
# config.network.alpha = args.alpha
|
402 |
+
# config.network.rank = args.rank
|
403 |
+
# config.save.name += f'_alpha{args.alpha}'
|
404 |
+
# config.save.name += f'_rank{config.network.rank }'
|
405 |
+
# config.save.name += f'_{config.network.training_method}'
|
406 |
+
# config.save.path += f'/{config.save.name}'
|
407 |
+
|
408 |
+
# prompts = prompt_util.load_prompts_from_yaml(config.prompts_file, attributes)
|
409 |
+
|
410 |
+
# device = torch.device(f"cuda:{args.device}")
|
411 |
+
# train(config, prompts, device)
|
412 |
+
|
413 |
+
|
414 |
+
def train_xl(target, postive, negative, lr, iterations, config_file, rank, device, attributes,save_name):
|
415 |
+
|
416 |
+
config = config_util.load_config_from_yaml(config_file)
|
417 |
+
randn = torch.randint(1, 10000000, (1,)).item()
|
418 |
+
config.save.name = save_name
|
419 |
+
|
420 |
+
config.train.lr = float(lr)
|
421 |
+
config.train.iterations=int(iterations)
|
422 |
+
|
423 |
+
if attributes is not None:
|
424 |
+
attributes = attributes.split(',')
|
425 |
+
attributes = [a.strip() for a in attributes]
|
426 |
+
config.network.alpha = 1.0
|
427 |
+
config.network.rank = rank
|
428 |
+
|
429 |
+
config.save.path += f'/{config.save.name}'
|
430 |
+
|
431 |
+
prompts = prompt_util.load_prompts_from_yaml(path=config.prompts_file, target=target, positive=positive, negative=negative, attributes=attributes)
|
432 |
+
|
433 |
+
device = torch.device(f"cuda:{device}")
|
434 |
+
train(config, prompts, device)
|
trainscripts/textsliders/prompt_util.py
CHANGED
@@ -148,9 +148,18 @@ class PromptEmbedsPair:
|
|
148 |
raise ValueError("action must be erase or enhance")
|
149 |
|
150 |
|
151 |
-
def load_prompts_from_yaml(path, attributes = []):
|
152 |
with open(path, "r") as f:
|
153 |
prompts = yaml.safe_load(f)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
print(prompts)
|
155 |
if len(prompts) == 0:
|
156 |
raise ValueError("prompts file is empty")
|
|
|
148 |
raise ValueError("action must be erase or enhance")
|
149 |
|
150 |
|
151 |
+
def load_prompts_from_yaml(path, target, positive, negative, attributes = []):
|
152 |
with open(path, "r") as f:
|
153 |
prompts = yaml.safe_load(f)
|
154 |
+
new = []
|
155 |
+
for prompt in prompts:
|
156 |
+
copy_ = copy.deepcopy(prompt)
|
157 |
+
copy_['target'] = target
|
158 |
+
copy_['positive'] = positive
|
159 |
+
copy_['neutral'] = target
|
160 |
+
copy_['unconditional'] = negative
|
161 |
+
new.append(copy_)
|
162 |
+
prompts = new
|
163 |
print(prompts)
|
164 |
if len(prompts) == 0:
|
165 |
raise ValueError("prompts file is empty")
|