Upload pipeline.py
Browse files- pipeline.py +145 -70
pipeline.py
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
import random
|
2 |
-
from typing import Callable, Dict
|
3 |
|
4 |
import torch
|
5 |
from diffusers import DiffusionPipeline
|
@@ -11,11 +11,10 @@ from tqdm import tqdm
|
|
11 |
|
12 |
|
13 |
def get_scaled_coeffs():
|
14 |
-
"""get_scaled_coeffs.
|
15 |
-
"""
|
16 |
beta_min = 0.85
|
17 |
beta_max = 12.0
|
18 |
-
return beta_min**0.5, beta_max**0.5-beta_min**0.5
|
19 |
|
20 |
|
21 |
def beta(t):
|
@@ -27,7 +26,7 @@ def beta(t):
|
|
27 |
t
|
28 |
"""
|
29 |
a, b = get_scaled_coeffs()
|
30 |
-
return (a+t*b)**2
|
31 |
|
32 |
|
33 |
def int_beta(t):
|
@@ -39,7 +38,7 @@ def int_beta(t):
|
|
39 |
t
|
40 |
"""
|
41 |
a, b = get_scaled_coeffs()
|
42 |
-
return ((a+b*t)**3-a**3)/(3*b)
|
43 |
|
44 |
|
45 |
def sigma(t):
|
@@ -50,7 +49,7 @@ def sigma(t):
|
|
50 |
t :
|
51 |
t
|
52 |
"""
|
53 |
-
return torch.expm1(int_beta(t))**0.5
|
54 |
|
55 |
|
56 |
def sigma_orig(t):
|
@@ -61,13 +60,21 @@ def sigma_orig(t):
|
|
61 |
t :
|
62 |
t
|
63 |
"""
|
64 |
-
return (-torch.expm1(-int_beta(t)))**0.5
|
65 |
|
66 |
|
67 |
class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
68 |
"""SuperDiffSDXLPipeline."""
|
69 |
|
70 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
"""__init__.
|
72 |
|
73 |
Parameters
|
@@ -99,13 +106,14 @@ class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
|
99 |
text_encoder.to(device)
|
100 |
text_encoder_2.to(device)
|
101 |
|
102 |
-
self.register_modules(
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
|
|
109 |
|
110 |
def prepare_prompt_input(self, prompt_o, prompt_b, batch_size, height, width):
|
111 |
"""prepare_prompt_input.
|
@@ -123,44 +131,82 @@ class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
|
123 |
width :
|
124 |
width
|
125 |
"""
|
126 |
-
text_input = self.tokenizer(
|
127 |
-
|
128 |
-
|
129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
with torch.no_grad():
|
131 |
text_embeddings = self.text_encoder(
|
132 |
-
text_input.input_ids.to(self.device), output_hidden_states=True
|
|
|
133 |
text_embeddings_2 = self.text_encoder_2(
|
134 |
-
text_input_2.input_ids.to(self.device), output_hidden_states=True
|
|
|
135 |
prompt_embeds_o = torch.concat(
|
136 |
-
(text_embeddings.hidden_states[-2],
|
|
|
|
|
|
|
137 |
pooled_prompt_embeds_o = text_embeddings_2[0]
|
138 |
negative_prompt_embeds = torch.zeros_like(prompt_embeds_o)
|
139 |
negative_pooled_prompt_embeds = torch.zeros_like(
|
140 |
pooled_prompt_embeds_o)
|
141 |
|
142 |
-
text_input = self.tokenizer(
|
143 |
-
|
144 |
-
|
145 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
146 |
with torch.no_grad():
|
147 |
text_embeddings = self.text_encoder(
|
148 |
-
text_input.input_ids.to(self.device), output_hidden_states=True
|
|
|
149 |
text_embeddings_2 = self.text_encoder_2(
|
150 |
-
text_input_2.input_ids.to(self.device), output_hidden_states=True
|
|
|
151 |
prompt_embeds_b = torch.concat(
|
152 |
-
(text_embeddings.hidden_states[-2],
|
|
|
|
|
|
|
153 |
pooled_prompt_embeds_b = text_embeddings_2[0]
|
154 |
add_time_ids_o = torch.tensor([(height, width, 0, 0, height, width)])
|
155 |
add_time_ids_b = torch.tensor([(height, width, 0, 0, height, width)])
|
156 |
negative_add_time_ids = torch.tensor(
|
157 |
[(height, width, 0, 0, height, width)])
|
158 |
prompt_embeds = torch.cat(
|
159 |
-
[negative_prompt_embeds, prompt_embeds_o, prompt_embeds_b], dim=0
|
|
|
160 |
add_text_embeds = torch.cat(
|
161 |
-
[
|
|
|
|
|
|
|
|
|
|
|
|
|
162 |
add_time_ids = torch.cat(
|
163 |
-
[negative_add_time_ids, add_time_ids_o, add_time_ids_b], dim=0
|
|
|
164 |
|
165 |
prompt_embeds = prompt_embeds.to(self.device)
|
166 |
add_text_embeds = add_text_embeds.to(self.device)
|
@@ -234,16 +280,8 @@ class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
|
234 |
embeddings : Callable
|
235 |
embeddings
|
236 |
"""
|
237 |
-
def v(_x, _e): return self.model(
|
238 |
-
"""v.
|
239 |
|
240 |
-
|
241 |
-
----------
|
242 |
-
_x :
|
243 |
-
_x
|
244 |
-
_e :
|
245 |
-
_e
|
246 |
-
"""
|
247 |
"""v.
|
248 |
|
249 |
Parameters
|
@@ -253,8 +291,10 @@ class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
|
253 |
_e :
|
254 |
_e
|
255 |
"""
|
256 |
-
|
257 |
-
|
|
|
|
|
258 |
embeds = torch.cat(embeddings)
|
259 |
latent_input = latents
|
260 |
vel = v(latent_input, embeds)
|
@@ -309,10 +349,15 @@ class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
|
309 |
self.seed
|
310 |
) # Seed generator to create the initial latent noise
|
311 |
|
312 |
-
latents = torch.randn(
|
313 |
-
|
|
|
|
|
|
|
|
|
314 |
prompt_embeds, added_cond_kwargs = self.prepare_prompt_input(
|
315 |
-
prompt_1, prompt_2, batch_size, height, width
|
|
|
316 |
|
317 |
return {
|
318 |
"latents": latents,
|
@@ -338,38 +383,68 @@ class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
|
338 |
added_cond_kwargs = model_inputs["added_cond_kwargs"]
|
339 |
|
340 |
t = torch.tensor(1.0)
|
341 |
-
dt = 1.0/self.num_inference_steps
|
342 |
train_number_steps = 1000
|
343 |
-
latents = latents * (sigma(t)**2+1)**0.5
|
344 |
with torch.no_grad():
|
345 |
for i in tqdm(range(self.num_inference_steps)):
|
346 |
latent_model_input = torch.cat([latents] * 3)
|
347 |
sigma_t = sigma(t)
|
348 |
-
dsigma = sigma(t-dt) - sigma_t
|
349 |
-
latent_model_input /= (sigma_t**2+1)**0.5
|
350 |
with torch.no_grad():
|
351 |
-
noise_pred = self.unet(
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
356 |
|
357 |
# noise = torch.sqrt(2*torch.abs(dsigma)*sigma_t)*torch.randn_like(latents)
|
358 |
-
noise = torch.sqrt(2*torch.abs(dsigma)*sigma_t)*torch.empty_like(
|
359 |
-
latents, device=self.device
|
360 |
-
|
361 |
-
|
362 |
-
|
363 |
-
|
364 |
-
|
365 |
-
|
366 |
-
|
367 |
-
|
368 |
-
|
369 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
370 |
|
371 |
if i < self.num_inference_steps - 1:
|
372 |
-
latents += 2*dsigma * noise_pred + noise
|
373 |
else:
|
374 |
latents += dsigma * noise_pred
|
375 |
|
@@ -389,7 +464,7 @@ class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
|
389 |
Callable
|
390 |
|
391 |
"""
|
392 |
-
latents = latents/self.vae.config.scaling_factor
|
393 |
latents = latents.to(torch.float32)
|
394 |
with torch.no_grad():
|
395 |
image = self.vae.decode(latents, return_dict=False)[0]
|
|
|
1 |
import random
|
2 |
+
from typing import Callable, Dict
|
3 |
|
4 |
import torch
|
5 |
from diffusers import DiffusionPipeline
|
|
|
11 |
|
12 |
|
13 |
def get_scaled_coeffs():
|
14 |
+
"""get_scaled_coeffs."""
|
|
|
15 |
beta_min = 0.85
|
16 |
beta_max = 12.0
|
17 |
+
return beta_min**0.5, beta_max**0.5 - beta_min**0.5
|
18 |
|
19 |
|
20 |
def beta(t):
|
|
|
26 |
t
|
27 |
"""
|
28 |
a, b = get_scaled_coeffs()
|
29 |
+
return (a + t * b) ** 2
|
30 |
|
31 |
|
32 |
def int_beta(t):
|
|
|
38 |
t
|
39 |
"""
|
40 |
a, b = get_scaled_coeffs()
|
41 |
+
return ((a + b * t) ** 3 - a**3) / (3 * b)
|
42 |
|
43 |
|
44 |
def sigma(t):
|
|
|
49 |
t :
|
50 |
t
|
51 |
"""
|
52 |
+
return torch.expm1(int_beta(t)) ** 0.5
|
53 |
|
54 |
|
55 |
def sigma_orig(t):
|
|
|
60 |
t :
|
61 |
t
|
62 |
"""
|
63 |
+
return (-torch.expm1(-int_beta(t))) ** 0.5
|
64 |
|
65 |
|
66 |
class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
67 |
"""SuperDiffSDXLPipeline."""
|
68 |
|
69 |
+
def __init__(
|
70 |
+
self,
|
71 |
+
unet: Callable,
|
72 |
+
vae: Callable,
|
73 |
+
text_encoder: Callable,
|
74 |
+
text_encoder_2: Callable,
|
75 |
+
tokenizer: Callable,
|
76 |
+
tokenizer_2: Callable,
|
77 |
+
) -> None:
|
78 |
"""__init__.
|
79 |
|
80 |
Parameters
|
|
|
106 |
text_encoder.to(device)
|
107 |
text_encoder_2.to(device)
|
108 |
|
109 |
+
self.register_modules(
|
110 |
+
unet=unet,
|
111 |
+
vae=vae,
|
112 |
+
text_encoder=text_encoder,
|
113 |
+
text_encoder_2=text_encoder_2,
|
114 |
+
tokenizer=tokenizer,
|
115 |
+
tokenizer_2=tokenizer_2,
|
116 |
+
)
|
117 |
|
118 |
def prepare_prompt_input(self, prompt_o, prompt_b, batch_size, height, width):
|
119 |
"""prepare_prompt_input.
|
|
|
131 |
width :
|
132 |
width
|
133 |
"""
|
134 |
+
text_input = self.tokenizer(
|
135 |
+
prompt_o * batch_size,
|
136 |
+
padding="max_length",
|
137 |
+
max_length=self.tokenizer.model_max_length,
|
138 |
+
truncation=True,
|
139 |
+
return_tensors="pt",
|
140 |
+
)
|
141 |
+
text_input_2 = self.tokenizer_2(
|
142 |
+
prompt_o * batch_size,
|
143 |
+
padding="max_length",
|
144 |
+
max_length=self.tokenizer_2.model_max_length,
|
145 |
+
truncation=True,
|
146 |
+
return_tensors="pt",
|
147 |
+
)
|
148 |
with torch.no_grad():
|
149 |
text_embeddings = self.text_encoder(
|
150 |
+
text_input.input_ids.to(self.device), output_hidden_states=True
|
151 |
+
)
|
152 |
text_embeddings_2 = self.text_encoder_2(
|
153 |
+
text_input_2.input_ids.to(self.device), output_hidden_states=True
|
154 |
+
)
|
155 |
prompt_embeds_o = torch.concat(
|
156 |
+
(text_embeddings.hidden_states[-2],
|
157 |
+
text_embeddings_2.hidden_states[-2]),
|
158 |
+
dim=-1,
|
159 |
+
)
|
160 |
pooled_prompt_embeds_o = text_embeddings_2[0]
|
161 |
negative_prompt_embeds = torch.zeros_like(prompt_embeds_o)
|
162 |
negative_pooled_prompt_embeds = torch.zeros_like(
|
163 |
pooled_prompt_embeds_o)
|
164 |
|
165 |
+
text_input = self.tokenizer(
|
166 |
+
prompt_b * batch_size,
|
167 |
+
padding="max_length",
|
168 |
+
max_length=self.tokenizer.model_max_length,
|
169 |
+
truncation=True,
|
170 |
+
return_tensors="pt",
|
171 |
+
)
|
172 |
+
text_input_2 = self.tokenizer_2(
|
173 |
+
prompt_b * batch_size,
|
174 |
+
padding="max_length",
|
175 |
+
max_length=self.tokenizer_2.model_max_length,
|
176 |
+
truncation=True,
|
177 |
+
return_tensors="pt",
|
178 |
+
)
|
179 |
with torch.no_grad():
|
180 |
text_embeddings = self.text_encoder(
|
181 |
+
text_input.input_ids.to(self.device), output_hidden_states=True
|
182 |
+
)
|
183 |
text_embeddings_2 = self.text_encoder_2(
|
184 |
+
text_input_2.input_ids.to(self.device), output_hidden_states=True
|
185 |
+
)
|
186 |
prompt_embeds_b = torch.concat(
|
187 |
+
(text_embeddings.hidden_states[-2],
|
188 |
+
text_embeddings_2.hidden_states[-2]),
|
189 |
+
dim=-1,
|
190 |
+
)
|
191 |
pooled_prompt_embeds_b = text_embeddings_2[0]
|
192 |
add_time_ids_o = torch.tensor([(height, width, 0, 0, height, width)])
|
193 |
add_time_ids_b = torch.tensor([(height, width, 0, 0, height, width)])
|
194 |
negative_add_time_ids = torch.tensor(
|
195 |
[(height, width, 0, 0, height, width)])
|
196 |
prompt_embeds = torch.cat(
|
197 |
+
[negative_prompt_embeds, prompt_embeds_o, prompt_embeds_b], dim=0
|
198 |
+
)
|
199 |
add_text_embeds = torch.cat(
|
200 |
+
[
|
201 |
+
negative_pooled_prompt_embeds,
|
202 |
+
pooled_prompt_embeds_o,
|
203 |
+
pooled_prompt_embeds_b,
|
204 |
+
],
|
205 |
+
dim=0,
|
206 |
+
)
|
207 |
add_time_ids = torch.cat(
|
208 |
+
[negative_add_time_ids, add_time_ids_o, add_time_ids_b], dim=0
|
209 |
+
)
|
210 |
|
211 |
prompt_embeds = prompt_embeds.to(self.device)
|
212 |
add_text_embeds = add_text_embeds.to(self.device)
|
|
|
280 |
embeddings : Callable
|
281 |
embeddings
|
282 |
"""
|
|
|
|
|
283 |
|
284 |
+
def v(_x, _e):
|
|
|
|
|
|
|
|
|
|
|
|
|
285 |
"""v.
|
286 |
|
287 |
Parameters
|
|
|
291 |
_e :
|
292 |
_e
|
293 |
"""
|
294 |
+
return self.model(
|
295 |
+
_x / ((sigma**2 + 1) ** 0.5), t, encoder_hidden_states=_e
|
296 |
+
).sample
|
297 |
+
|
298 |
embeds = torch.cat(embeddings)
|
299 |
latent_input = latents
|
300 |
vel = v(latent_input, embeds)
|
|
|
349 |
self.seed
|
350 |
) # Seed generator to create the initial latent noise
|
351 |
|
352 |
+
latents = torch.randn(
|
353 |
+
(batch_size, self.unet.in_channels, height // 8, width // 8),
|
354 |
+
generator=self.generator,
|
355 |
+
dtype=self.dtype,
|
356 |
+
device=self.device,
|
357 |
+
)
|
358 |
prompt_embeds, added_cond_kwargs = self.prepare_prompt_input(
|
359 |
+
prompt_1, prompt_2, batch_size, height, width
|
360 |
+
)
|
361 |
|
362 |
return {
|
363 |
"latents": latents,
|
|
|
383 |
added_cond_kwargs = model_inputs["added_cond_kwargs"]
|
384 |
|
385 |
t = torch.tensor(1.0)
|
386 |
+
dt = 1.0 / self.num_inference_steps
|
387 |
train_number_steps = 1000
|
388 |
+
latents = latents * (sigma(t) ** 2 + 1) ** 0.5
|
389 |
with torch.no_grad():
|
390 |
for i in tqdm(range(self.num_inference_steps)):
|
391 |
latent_model_input = torch.cat([latents] * 3)
|
392 |
sigma_t = sigma(t)
|
393 |
+
dsigma = sigma(t - dt) - sigma_t
|
394 |
+
latent_model_input /= (sigma_t**2 + 1) ** 0.5
|
395 |
with torch.no_grad():
|
396 |
+
noise_pred = self.unet(
|
397 |
+
latent_model_input,
|
398 |
+
t * train_number_steps,
|
399 |
+
encoder_hidden_states=prompt_embeds,
|
400 |
+
added_cond_kwargs=added_cond_kwargs,
|
401 |
+
return_dict=False,
|
402 |
+
)[0]
|
403 |
+
|
404 |
+
(
|
405 |
+
noise_pred_uncond,
|
406 |
+
noise_pred_text_o,
|
407 |
+
noise_pred_text_b,
|
408 |
+
) = noise_pred.chunk(3)
|
409 |
|
410 |
# noise = torch.sqrt(2*torch.abs(dsigma)*sigma_t)*torch.randn_like(latents)
|
411 |
+
noise = torch.sqrt(2 * torch.abs(dsigma) * sigma_t) * torch.empty_like(
|
412 |
+
latents, device=self.device
|
413 |
+
).normal_(generator=self.generator)
|
414 |
+
|
415 |
+
dx_ind = (
|
416 |
+
2
|
417 |
+
* dsigma
|
418 |
+
* (
|
419 |
+
noise_pred_uncond
|
420 |
+
+ self.guidance_scale *
|
421 |
+
(noise_pred_text_b - noise_pred_uncond)
|
422 |
+
)
|
423 |
+
+ noise
|
424 |
+
)
|
425 |
+
kappa = (
|
426 |
+
torch.abs(dsigma)
|
427 |
+
* (noise_pred_text_b - noise_pred_text_o)
|
428 |
+
* (noise_pred_text_b + noise_pred_text_o)
|
429 |
+
).sum((1, 2, 3)) - (
|
430 |
+
dx_ind * ((noise_pred_text_o - noise_pred_text_b))
|
431 |
+
).sum(
|
432 |
+
(1, 2, 3)
|
433 |
+
)
|
434 |
+
kappa /= (
|
435 |
+
2
|
436 |
+
* dsigma
|
437 |
+
* self.guidance_scale
|
438 |
+
* ((noise_pred_text_o - noise_pred_text_b) ** 2).sum((1, 2, 3))
|
439 |
+
)
|
440 |
+
noise_pred = noise_pred_uncond + self.guidance_scale * (
|
441 |
+
(noise_pred_text_b - noise_pred_uncond)
|
442 |
+
+ kappa[:, None, None, None]
|
443 |
+
* (noise_pred_text_o - noise_pred_text_b)
|
444 |
+
)
|
445 |
|
446 |
if i < self.num_inference_steps - 1:
|
447 |
+
latents += 2 * dsigma * noise_pred + noise
|
448 |
else:
|
449 |
latents += dsigma * noise_pred
|
450 |
|
|
|
464 |
Callable
|
465 |
|
466 |
"""
|
467 |
+
latents = latents / self.vae.config.scaling_factor
|
468 |
latents = latents.to(torch.float32)
|
469 |
with torch.no_grad():
|
470 |
image = self.vae.decode(latents, return_dict=False)[0]
|