Upload lora-scripts/sd-scripts/library/custom_train_functions.py with huggingface_hub
Browse files
lora-scripts/sd-scripts/library/custom_train_functions.py
ADDED
@@ -0,0 +1,548 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import argparse
|
3 |
+
import random
|
4 |
+
import re
|
5 |
+
from typing import List, Optional, Union
|
6 |
+
from .utils import setup_logging
|
7 |
+
|
8 |
+
setup_logging()
|
9 |
+
import logging
|
10 |
+
|
11 |
+
logger = logging.getLogger(__name__)
|
12 |
+
|
13 |
+
|
14 |
+
def prepare_scheduler_for_custom_training(noise_scheduler, device):
|
15 |
+
if hasattr(noise_scheduler, "all_snr"):
|
16 |
+
return
|
17 |
+
|
18 |
+
alphas_cumprod = noise_scheduler.alphas_cumprod
|
19 |
+
sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod)
|
20 |
+
sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - alphas_cumprod)
|
21 |
+
alpha = sqrt_alphas_cumprod
|
22 |
+
sigma = sqrt_one_minus_alphas_cumprod
|
23 |
+
all_snr = (alpha / sigma) ** 2
|
24 |
+
|
25 |
+
noise_scheduler.all_snr = all_snr.to(device)
|
26 |
+
|
27 |
+
|
28 |
+
def fix_noise_scheduler_betas_for_zero_terminal_snr(noise_scheduler):
|
29 |
+
# fix beta: zero terminal SNR
|
30 |
+
logger.info(f"fix noise scheduler betas: https://arxiv.org/abs/2305.08891")
|
31 |
+
|
32 |
+
def enforce_zero_terminal_snr(betas):
|
33 |
+
# Convert betas to alphas_bar_sqrt
|
34 |
+
alphas = 1 - betas
|
35 |
+
alphas_bar = alphas.cumprod(0)
|
36 |
+
alphas_bar_sqrt = alphas_bar.sqrt()
|
37 |
+
|
38 |
+
# Store old values.
|
39 |
+
alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
|
40 |
+
alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
|
41 |
+
# Shift so last timestep is zero.
|
42 |
+
alphas_bar_sqrt -= alphas_bar_sqrt_T
|
43 |
+
# Scale so first timestep is back to old value.
|
44 |
+
alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
|
45 |
+
|
46 |
+
# Convert alphas_bar_sqrt to betas
|
47 |
+
alphas_bar = alphas_bar_sqrt**2
|
48 |
+
alphas = alphas_bar[1:] / alphas_bar[:-1]
|
49 |
+
alphas = torch.cat([alphas_bar[0:1], alphas])
|
50 |
+
betas = 1 - alphas
|
51 |
+
return betas
|
52 |
+
|
53 |
+
betas = noise_scheduler.betas
|
54 |
+
betas = enforce_zero_terminal_snr(betas)
|
55 |
+
alphas = 1.0 - betas
|
56 |
+
alphas_cumprod = torch.cumprod(alphas, dim=0)
|
57 |
+
|
58 |
+
# logger.info(f"original: {noise_scheduler.betas}")
|
59 |
+
# logger.info(f"fixed: {betas}")
|
60 |
+
|
61 |
+
noise_scheduler.betas = betas
|
62 |
+
noise_scheduler.alphas = alphas
|
63 |
+
noise_scheduler.alphas_cumprod = alphas_cumprod
|
64 |
+
|
65 |
+
|
66 |
+
def apply_snr_weight(loss, timesteps, noise_scheduler, gamma, v_prediction=False):
|
67 |
+
snr = torch.stack([noise_scheduler.all_snr[t] for t in timesteps])
|
68 |
+
min_snr_gamma = torch.minimum(snr, torch.full_like(snr, gamma))
|
69 |
+
if v_prediction:
|
70 |
+
snr_weight = torch.div(min_snr_gamma, snr + 1).float().to(loss.device)
|
71 |
+
else:
|
72 |
+
snr_weight = torch.div(min_snr_gamma, snr).float().to(loss.device)
|
73 |
+
loss = loss * snr_weight
|
74 |
+
return loss
|
75 |
+
|
76 |
+
|
77 |
+
def scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler):
|
78 |
+
scale = get_snr_scale(timesteps, noise_scheduler)
|
79 |
+
loss = loss * scale
|
80 |
+
return loss
|
81 |
+
|
82 |
+
|
83 |
+
def get_snr_scale(timesteps, noise_scheduler):
|
84 |
+
snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps]) # batch_size
|
85 |
+
snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000) # if timestep is 0, snr_t is inf, so limit it to 1000
|
86 |
+
scale = snr_t / (snr_t + 1)
|
87 |
+
# # show debug info
|
88 |
+
# logger.info(f"timesteps: {timesteps}, snr_t: {snr_t}, scale: {scale}")
|
89 |
+
return scale
|
90 |
+
|
91 |
+
|
92 |
+
def add_v_prediction_like_loss(loss, timesteps, noise_scheduler, v_pred_like_loss):
|
93 |
+
scale = get_snr_scale(timesteps, noise_scheduler)
|
94 |
+
# logger.info(f"add v-prediction like loss: {v_pred_like_loss}, scale: {scale}, loss: {loss}, time: {timesteps}")
|
95 |
+
loss = loss + loss / scale * v_pred_like_loss
|
96 |
+
return loss
|
97 |
+
|
98 |
+
|
99 |
+
def apply_debiased_estimation(loss, timesteps, noise_scheduler):
|
100 |
+
snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps]) # batch_size
|
101 |
+
snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000) # if timestep is 0, snr_t is inf, so limit it to 1000
|
102 |
+
weight = 1 / torch.sqrt(snr_t)
|
103 |
+
loss = weight * loss
|
104 |
+
return loss
|
105 |
+
|
106 |
+
|
107 |
+
# TODO train_utilと分散しているのでどちらかに寄せる
|
108 |
+
|
109 |
+
|
110 |
+
def add_custom_train_arguments(parser: argparse.ArgumentParser, support_weighted_captions: bool = True):
|
111 |
+
parser.add_argument(
|
112 |
+
"--min_snr_gamma",
|
113 |
+
type=float,
|
114 |
+
default=None,
|
115 |
+
help="gamma for reducing the weight of high loss timesteps. Lower numbers have stronger effect. 5 is recommended by paper. / 低いタイムステップでの高いlossに対して重みを減らすためのgamma値、低いほど効果が強く、論文では5が推奨",
|
116 |
+
)
|
117 |
+
parser.add_argument(
|
118 |
+
"--scale_v_pred_loss_like_noise_pred",
|
119 |
+
action="store_true",
|
120 |
+
help="scale v-prediction loss like noise prediction loss / v-prediction lossをnoise prediction lossと同じようにスケーリングする",
|
121 |
+
)
|
122 |
+
parser.add_argument(
|
123 |
+
"--v_pred_like_loss",
|
124 |
+
type=float,
|
125 |
+
default=None,
|
126 |
+
help="add v-prediction like loss multiplied by this value / v-prediction lossをこの値をかけたものをlossに加算する",
|
127 |
+
)
|
128 |
+
parser.add_argument(
|
129 |
+
"--debiased_estimation_loss",
|
130 |
+
action="store_true",
|
131 |
+
help="debiased estimation loss / debiased estimation loss",
|
132 |
+
)
|
133 |
+
if support_weighted_captions:
|
134 |
+
parser.add_argument(
|
135 |
+
"--weighted_captions",
|
136 |
+
action="store_true",
|
137 |
+
default=False,
|
138 |
+
help="Enable weighted captions in the standard style (token:1.3). No commas inside parens, or shuffle/dropout may break the decoder. / 「[token]」、「(token)」「(token:1.3)」のような重み付きキャプションを有効にする。カンマを括弧内に入れるとシャッフルやdropoutで重みづけがおかしくなるので注意",
|
139 |
+
)
|
140 |
+
|
141 |
+
|
142 |
+
re_attention = re.compile(
|
143 |
+
r"""
|
144 |
+
\\\(|
|
145 |
+
\\\)|
|
146 |
+
\\\[|
|
147 |
+
\\]|
|
148 |
+
\\\\|
|
149 |
+
\\|
|
150 |
+
\(|
|
151 |
+
\[|
|
152 |
+
:([+-]?[.\d]+)\)|
|
153 |
+
\)|
|
154 |
+
]|
|
155 |
+
[^\\()\[\]:]+|
|
156 |
+
:
|
157 |
+
""",
|
158 |
+
re.X,
|
159 |
+
)
|
160 |
+
|
161 |
+
|
162 |
+
def parse_prompt_attention(text):
|
163 |
+
"""
|
164 |
+
Parses a string with attention tokens and returns a list of pairs: text and its associated weight.
|
165 |
+
Accepted tokens are:
|
166 |
+
(abc) - increases attention to abc by a multiplier of 1.1
|
167 |
+
(abc:3.12) - increases attention to abc by a multiplier of 3.12
|
168 |
+
[abc] - decreases attention to abc by a multiplier of 1.1
|
169 |
+
\( - literal character '('
|
170 |
+
\[ - literal character '['
|
171 |
+
\) - literal character ')'
|
172 |
+
\] - literal character ']'
|
173 |
+
\\ - literal character '\'
|
174 |
+
anything else - just text
|
175 |
+
>>> parse_prompt_attention('normal text')
|
176 |
+
[['normal text', 1.0]]
|
177 |
+
>>> parse_prompt_attention('an (important) word')
|
178 |
+
[['an ', 1.0], ['important', 1.1], [' word', 1.0]]
|
179 |
+
>>> parse_prompt_attention('(unbalanced')
|
180 |
+
[['unbalanced', 1.1]]
|
181 |
+
>>> parse_prompt_attention('\(literal\]')
|
182 |
+
[['(literal]', 1.0]]
|
183 |
+
>>> parse_prompt_attention('(unnecessary)(parens)')
|
184 |
+
[['unnecessaryparens', 1.1]]
|
185 |
+
>>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).')
|
186 |
+
[['a ', 1.0],
|
187 |
+
['house', 1.5730000000000004],
|
188 |
+
[' ', 1.1],
|
189 |
+
['on', 1.0],
|
190 |
+
[' a ', 1.1],
|
191 |
+
['hill', 0.55],
|
192 |
+
[', sun, ', 1.1],
|
193 |
+
['sky', 1.4641000000000006],
|
194 |
+
['.', 1.1]]
|
195 |
+
"""
|
196 |
+
|
197 |
+
res = []
|
198 |
+
round_brackets = []
|
199 |
+
square_brackets = []
|
200 |
+
|
201 |
+
round_bracket_multiplier = 1.1
|
202 |
+
square_bracket_multiplier = 1 / 1.1
|
203 |
+
|
204 |
+
def multiply_range(start_position, multiplier):
|
205 |
+
for p in range(start_position, len(res)):
|
206 |
+
res[p][1] *= multiplier
|
207 |
+
|
208 |
+
for m in re_attention.finditer(text):
|
209 |
+
text = m.group(0)
|
210 |
+
weight = m.group(1)
|
211 |
+
|
212 |
+
if text.startswith("\\"):
|
213 |
+
res.append([text[1:], 1.0])
|
214 |
+
elif text == "(":
|
215 |
+
round_brackets.append(len(res))
|
216 |
+
elif text == "[":
|
217 |
+
square_brackets.append(len(res))
|
218 |
+
elif weight is not None and len(round_brackets) > 0:
|
219 |
+
multiply_range(round_brackets.pop(), float(weight))
|
220 |
+
elif text == ")" and len(round_brackets) > 0:
|
221 |
+
multiply_range(round_brackets.pop(), round_bracket_multiplier)
|
222 |
+
elif text == "]" and len(square_brackets) > 0:
|
223 |
+
multiply_range(square_brackets.pop(), square_bracket_multiplier)
|
224 |
+
else:
|
225 |
+
res.append([text, 1.0])
|
226 |
+
|
227 |
+
for pos in round_brackets:
|
228 |
+
multiply_range(pos, round_bracket_multiplier)
|
229 |
+
|
230 |
+
for pos in square_brackets:
|
231 |
+
multiply_range(pos, square_bracket_multiplier)
|
232 |
+
|
233 |
+
if len(res) == 0:
|
234 |
+
res = [["", 1.0]]
|
235 |
+
|
236 |
+
# merge runs of identical weights
|
237 |
+
i = 0
|
238 |
+
while i + 1 < len(res):
|
239 |
+
if res[i][1] == res[i + 1][1]:
|
240 |
+
res[i][0] += res[i + 1][0]
|
241 |
+
res.pop(i + 1)
|
242 |
+
else:
|
243 |
+
i += 1
|
244 |
+
|
245 |
+
return res
|
246 |
+
|
247 |
+
|
248 |
+
def get_prompts_with_weights(tokenizer, prompt: List[str], max_length: int):
|
249 |
+
r"""
|
250 |
+
Tokenize a list of prompts and return its tokens with weights of each token.
|
251 |
+
|
252 |
+
No padding, starting or ending token is included.
|
253 |
+
"""
|
254 |
+
tokens = []
|
255 |
+
weights = []
|
256 |
+
truncated = False
|
257 |
+
for text in prompt:
|
258 |
+
texts_and_weights = parse_prompt_attention(text)
|
259 |
+
text_token = []
|
260 |
+
text_weight = []
|
261 |
+
for word, weight in texts_and_weights:
|
262 |
+
# tokenize and discard the starting and the ending token
|
263 |
+
token = tokenizer(word).input_ids[1:-1]
|
264 |
+
text_token += token
|
265 |
+
# copy the weight by length of token
|
266 |
+
text_weight += [weight] * len(token)
|
267 |
+
# stop if the text is too long (longer than truncation limit)
|
268 |
+
if len(text_token) > max_length:
|
269 |
+
truncated = True
|
270 |
+
break
|
271 |
+
# truncate
|
272 |
+
if len(text_token) > max_length:
|
273 |
+
truncated = True
|
274 |
+
text_token = text_token[:max_length]
|
275 |
+
text_weight = text_weight[:max_length]
|
276 |
+
tokens.append(text_token)
|
277 |
+
weights.append(text_weight)
|
278 |
+
if truncated:
|
279 |
+
logger.warning("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples")
|
280 |
+
return tokens, weights
|
281 |
+
|
282 |
+
|
283 |
+
def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, no_boseos_middle=True, chunk_length=77):
|
284 |
+
r"""
|
285 |
+
Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length.
|
286 |
+
"""
|
287 |
+
max_embeddings_multiples = (max_length - 2) // (chunk_length - 2)
|
288 |
+
weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length
|
289 |
+
for i in range(len(tokens)):
|
290 |
+
tokens[i] = [bos] + tokens[i] + [eos] * (max_length - 1 - len(tokens[i]))
|
291 |
+
if no_boseos_middle:
|
292 |
+
weights[i] = [1.0] + weights[i] + [1.0] * (max_length - 1 - len(weights[i]))
|
293 |
+
else:
|
294 |
+
w = []
|
295 |
+
if len(weights[i]) == 0:
|
296 |
+
w = [1.0] * weights_length
|
297 |
+
else:
|
298 |
+
for j in range(max_embeddings_multiples):
|
299 |
+
w.append(1.0) # weight for starting token in this chunk
|
300 |
+
w += weights[i][j * (chunk_length - 2) : min(len(weights[i]), (j + 1) * (chunk_length - 2))]
|
301 |
+
w.append(1.0) # weight for ending token in this chunk
|
302 |
+
w += [1.0] * (weights_length - len(w))
|
303 |
+
weights[i] = w[:]
|
304 |
+
|
305 |
+
return tokens, weights
|
306 |
+
|
307 |
+
|
308 |
+
def get_unweighted_text_embeddings(
|
309 |
+
tokenizer,
|
310 |
+
text_encoder,
|
311 |
+
text_input: torch.Tensor,
|
312 |
+
chunk_length: int,
|
313 |
+
clip_skip: int,
|
314 |
+
eos: int,
|
315 |
+
pad: int,
|
316 |
+
no_boseos_middle: Optional[bool] = True,
|
317 |
+
):
|
318 |
+
"""
|
319 |
+
When the length of tokens is a multiple of the capacity of the text encoder,
|
320 |
+
it should be split into chunks and sent to the text encoder individually.
|
321 |
+
"""
|
322 |
+
max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2)
|
323 |
+
if max_embeddings_multiples > 1:
|
324 |
+
text_embeddings = []
|
325 |
+
for i in range(max_embeddings_multiples):
|
326 |
+
# extract the i-th chunk
|
327 |
+
text_input_chunk = text_input[:, i * (chunk_length - 2) : (i + 1) * (chunk_length - 2) + 2].clone()
|
328 |
+
|
329 |
+
# cover the head and the tail by the starting and the ending tokens
|
330 |
+
text_input_chunk[:, 0] = text_input[0, 0]
|
331 |
+
if pad == eos: # v1
|
332 |
+
text_input_chunk[:, -1] = text_input[0, -1]
|
333 |
+
else: # v2
|
334 |
+
for j in range(len(text_input_chunk)):
|
335 |
+
if text_input_chunk[j, -1] != eos and text_input_chunk[j, -1] != pad: # 最後に普通の文字がある
|
336 |
+
text_input_chunk[j, -1] = eos
|
337 |
+
if text_input_chunk[j, 1] == pad: # BOSだけであとはPAD
|
338 |
+
text_input_chunk[j, 1] = eos
|
339 |
+
|
340 |
+
if clip_skip is None or clip_skip == 1:
|
341 |
+
text_embedding = text_encoder(text_input_chunk)[0]
|
342 |
+
else:
|
343 |
+
enc_out = text_encoder(text_input_chunk, output_hidden_states=True, return_dict=True)
|
344 |
+
text_embedding = enc_out["hidden_states"][-clip_skip]
|
345 |
+
text_embedding = text_encoder.text_model.final_layer_norm(text_embedding)
|
346 |
+
|
347 |
+
if no_boseos_middle:
|
348 |
+
if i == 0:
|
349 |
+
# discard the ending token
|
350 |
+
text_embedding = text_embedding[:, :-1]
|
351 |
+
elif i == max_embeddings_multiples - 1:
|
352 |
+
# discard the starting token
|
353 |
+
text_embedding = text_embedding[:, 1:]
|
354 |
+
else:
|
355 |
+
# discard both starting and ending tokens
|
356 |
+
text_embedding = text_embedding[:, 1:-1]
|
357 |
+
|
358 |
+
text_embeddings.append(text_embedding)
|
359 |
+
text_embeddings = torch.concat(text_embeddings, axis=1)
|
360 |
+
else:
|
361 |
+
if clip_skip is None or clip_skip == 1:
|
362 |
+
text_embeddings = text_encoder(text_input)[0]
|
363 |
+
else:
|
364 |
+
enc_out = text_encoder(text_input, output_hidden_states=True, return_dict=True)
|
365 |
+
text_embeddings = enc_out["hidden_states"][-clip_skip]
|
366 |
+
text_embeddings = text_encoder.text_model.final_layer_norm(text_embeddings)
|
367 |
+
return text_embeddings
|
368 |
+
|
369 |
+
|
370 |
+
def get_weighted_text_embeddings(
|
371 |
+
tokenizer,
|
372 |
+
text_encoder,
|
373 |
+
prompt: Union[str, List[str]],
|
374 |
+
device,
|
375 |
+
max_embeddings_multiples: Optional[int] = 3,
|
376 |
+
no_boseos_middle: Optional[bool] = False,
|
377 |
+
clip_skip=None,
|
378 |
+
):
|
379 |
+
r"""
|
380 |
+
Prompts can be assigned with local weights using brackets. For example,
|
381 |
+
prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful',
|
382 |
+
and the embedding tokens corresponding to the words get multiplied by a constant, 1.1.
|
383 |
+
|
384 |
+
Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean.
|
385 |
+
|
386 |
+
Args:
|
387 |
+
prompt (`str` or `List[str]`):
|
388 |
+
The prompt or prompts to guide the image generation.
|
389 |
+
max_embeddings_multiples (`int`, *optional*, defaults to `3`):
|
390 |
+
The max multiple length of prompt embeddings compared to the max output length of text encoder.
|
391 |
+
no_boseos_middle (`bool`, *optional*, defaults to `False`):
|
392 |
+
If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and
|
393 |
+
ending token in each of the chunk in the middle.
|
394 |
+
skip_parsing (`bool`, *optional*, defaults to `False`):
|
395 |
+
Skip the parsing of brackets.
|
396 |
+
skip_weighting (`bool`, *optional*, defaults to `False`):
|
397 |
+
Skip the weighting. When the parsing is skipped, it is forced True.
|
398 |
+
"""
|
399 |
+
max_length = (tokenizer.model_max_length - 2) * max_embeddings_multiples + 2
|
400 |
+
if isinstance(prompt, str):
|
401 |
+
prompt = [prompt]
|
402 |
+
|
403 |
+
prompt_tokens, prompt_weights = get_prompts_with_weights(tokenizer, prompt, max_length - 2)
|
404 |
+
|
405 |
+
# round up the longest length of tokens to a multiple of (model_max_length - 2)
|
406 |
+
max_length = max([len(token) for token in prompt_tokens])
|
407 |
+
|
408 |
+
max_embeddings_multiples = min(
|
409 |
+
max_embeddings_multiples,
|
410 |
+
(max_length - 1) // (tokenizer.model_max_length - 2) + 1,
|
411 |
+
)
|
412 |
+
max_embeddings_multiples = max(1, max_embeddings_multiples)
|
413 |
+
max_length = (tokenizer.model_max_length - 2) * max_embeddings_multiples + 2
|
414 |
+
|
415 |
+
# pad the length of tokens and weights
|
416 |
+
bos = tokenizer.bos_token_id
|
417 |
+
eos = tokenizer.eos_token_id
|
418 |
+
pad = tokenizer.pad_token_id
|
419 |
+
prompt_tokens, prompt_weights = pad_tokens_and_weights(
|
420 |
+
prompt_tokens,
|
421 |
+
prompt_weights,
|
422 |
+
max_length,
|
423 |
+
bos,
|
424 |
+
eos,
|
425 |
+
no_boseos_middle=no_boseos_middle,
|
426 |
+
chunk_length=tokenizer.model_max_length,
|
427 |
+
)
|
428 |
+
prompt_tokens = torch.tensor(prompt_tokens, dtype=torch.long, device=device)
|
429 |
+
|
430 |
+
# get the embeddings
|
431 |
+
text_embeddings = get_unweighted_text_embeddings(
|
432 |
+
tokenizer,
|
433 |
+
text_encoder,
|
434 |
+
prompt_tokens,
|
435 |
+
tokenizer.model_max_length,
|
436 |
+
clip_skip,
|
437 |
+
eos,
|
438 |
+
pad,
|
439 |
+
no_boseos_middle=no_boseos_middle,
|
440 |
+
)
|
441 |
+
prompt_weights = torch.tensor(prompt_weights, dtype=text_embeddings.dtype, device=device)
|
442 |
+
|
443 |
+
# assign weights to the prompts and normalize in the sense of mean
|
444 |
+
previous_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype)
|
445 |
+
text_embeddings = text_embeddings * prompt_weights.unsqueeze(-1)
|
446 |
+
current_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype)
|
447 |
+
text_embeddings = text_embeddings * (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1)
|
448 |
+
|
449 |
+
return text_embeddings
|
450 |
+
|
451 |
+
|
452 |
+
# https://wandb.ai/johnowhitaker/multires_noise/reports/Multi-Resolution-Noise-for-Diffusion-Model-Training--VmlldzozNjYyOTU2
|
453 |
+
def pyramid_noise_like(noise, device, iterations=6, discount=0.4):
|
454 |
+
b, c, w, h = noise.shape # EDIT: w and h get over-written, rename for a different variant!
|
455 |
+
u = torch.nn.Upsample(size=(w, h), mode="bilinear").to(device)
|
456 |
+
for i in range(iterations):
|
457 |
+
r = random.random() * 2 + 2 # Rather than always going 2x,
|
458 |
+
wn, hn = max(1, int(w / (r**i))), max(1, int(h / (r**i)))
|
459 |
+
noise += u(torch.randn(b, c, wn, hn).to(device)) * discount**i
|
460 |
+
if wn == 1 or hn == 1:
|
461 |
+
break # Lowest resolution is 1x1
|
462 |
+
return noise / noise.std() # Scaled back to roughly unit variance
|
463 |
+
|
464 |
+
|
465 |
+
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
|
466 |
+
def apply_noise_offset(latents, noise, noise_offset, adaptive_noise_scale):
|
467 |
+
if noise_offset is None:
|
468 |
+
return noise
|
469 |
+
if adaptive_noise_scale is not None:
|
470 |
+
# latent shape: (batch_size, channels, height, width)
|
471 |
+
# abs mean value for each channel
|
472 |
+
latent_mean = torch.abs(latents.mean(dim=(2, 3), keepdim=True))
|
473 |
+
|
474 |
+
# multiply adaptive noise scale to the mean value and add it to the noise offset
|
475 |
+
noise_offset = noise_offset + adaptive_noise_scale * latent_mean
|
476 |
+
noise_offset = torch.clamp(noise_offset, 0.0, None) # in case of adaptive noise scale is negative
|
477 |
+
|
478 |
+
noise = noise + noise_offset * torch.randn((latents.shape[0], latents.shape[1], 1, 1), device=latents.device)
|
479 |
+
return noise
|
480 |
+
|
481 |
+
|
482 |
+
def apply_masked_loss(loss, batch):
|
483 |
+
# mask image is -1 to 1. we need to convert it to 0 to 1
|
484 |
+
mask_image = batch["conditioning_images"].to(dtype=loss.dtype)[:, 0].unsqueeze(1) # use R channel
|
485 |
+
|
486 |
+
# resize to the same size as the loss
|
487 |
+
mask_image = torch.nn.functional.interpolate(mask_image, size=loss.shape[2:], mode="area")
|
488 |
+
mask_image = mask_image / 2 + 0.5
|
489 |
+
loss = loss * mask_image
|
490 |
+
return loss
|
491 |
+
|
492 |
+
|
493 |
+
"""
|
494 |
+
##########################################
|
495 |
+
# Perlin Noise
|
496 |
+
def rand_perlin_2d(device, shape, res, fade=lambda t: 6 * t**5 - 15 * t**4 + 10 * t**3):
|
497 |
+
delta = (res[0] / shape[0], res[1] / shape[1])
|
498 |
+
d = (shape[0] // res[0], shape[1] // res[1])
|
499 |
+
|
500 |
+
grid = (
|
501 |
+
torch.stack(
|
502 |
+
torch.meshgrid(torch.arange(0, res[0], delta[0], device=device), torch.arange(0, res[1], delta[1], device=device)),
|
503 |
+
dim=-1,
|
504 |
+
)
|
505 |
+
% 1
|
506 |
+
)
|
507 |
+
angles = 2 * torch.pi * torch.rand(res[0] + 1, res[1] + 1, device=device)
|
508 |
+
gradients = torch.stack((torch.cos(angles), torch.sin(angles)), dim=-1)
|
509 |
+
|
510 |
+
tile_grads = (
|
511 |
+
lambda slice1, slice2: gradients[slice1[0] : slice1[1], slice2[0] : slice2[1]]
|
512 |
+
.repeat_interleave(d[0], 0)
|
513 |
+
.repeat_interleave(d[1], 1)
|
514 |
+
)
|
515 |
+
dot = lambda grad, shift: (
|
516 |
+
torch.stack((grid[: shape[0], : shape[1], 0] + shift[0], grid[: shape[0], : shape[1], 1] + shift[1]), dim=-1)
|
517 |
+
* grad[: shape[0], : shape[1]]
|
518 |
+
).sum(dim=-1)
|
519 |
+
|
520 |
+
n00 = dot(tile_grads([0, -1], [0, -1]), [0, 0])
|
521 |
+
n10 = dot(tile_grads([1, None], [0, -1]), [-1, 0])
|
522 |
+
n01 = dot(tile_grads([0, -1], [1, None]), [0, -1])
|
523 |
+
n11 = dot(tile_grads([1, None], [1, None]), [-1, -1])
|
524 |
+
t = fade(grid[: shape[0], : shape[1]])
|
525 |
+
return 1.414 * torch.lerp(torch.lerp(n00, n10, t[..., 0]), torch.lerp(n01, n11, t[..., 0]), t[..., 1])
|
526 |
+
|
527 |
+
|
528 |
+
def rand_perlin_2d_octaves(device, shape, res, octaves=1, persistence=0.5):
|
529 |
+
noise = torch.zeros(shape, device=device)
|
530 |
+
frequency = 1
|
531 |
+
amplitude = 1
|
532 |
+
for _ in range(octaves):
|
533 |
+
noise += amplitude * rand_perlin_2d(device, shape, (frequency * res[0], frequency * res[1]))
|
534 |
+
frequency *= 2
|
535 |
+
amplitude *= persistence
|
536 |
+
return noise
|
537 |
+
|
538 |
+
|
539 |
+
def perlin_noise(noise, device, octaves):
|
540 |
+
_, c, w, h = noise.shape
|
541 |
+
perlin = lambda: rand_perlin_2d_octaves(device, (w, h), (4, 4), octaves)
|
542 |
+
noise_perlin = []
|
543 |
+
for _ in range(c):
|
544 |
+
noise_perlin.append(perlin())
|
545 |
+
noise_perlin = torch.stack(noise_perlin).unsqueeze(0) # (1, c, w, h)
|
546 |
+
noise += noise_perlin # broadcast for each batch
|
547 |
+
return noise / noise.std() # Scaled back to roughly unit variance
|
548 |
+
"""
|