Delete pipeline.py
Browse files- pipeline.py +0 -376
pipeline.py
DELETED
@@ -1,376 +0,0 @@
|
|
1 |
-
import random
|
2 |
-
from typing import Callable, Dict, List, Optional
|
3 |
-
|
4 |
-
import torch
|
5 |
-
from tqdm import tqdm
|
6 |
-
from diffusers import DiffusionPipeline
|
7 |
-
from diffusers.configuration_utils import ConfigMixin
|
8 |
-
|
9 |
-
def get_scaled_coeffs():
|
10 |
-
beta_min = 0.85
|
11 |
-
beta_max = 12.0
|
12 |
-
return beta_min**0.5, beta_max**0.5-beta_min**0.5
|
13 |
-
def beta(t):
|
14 |
-
a, b = get_scaled_coeffs()
|
15 |
-
return (a+t*b)**2
|
16 |
-
def int_beta(t):
|
17 |
-
a, b = get_scaled_coeffs()
|
18 |
-
return ((a+b*t)**3-a**3)/(3*b)
|
19 |
-
def sigma(t):
|
20 |
-
return torch.expm1(int_beta(t))**0.5
|
21 |
-
def sigma_orig(t):
|
22 |
-
return (-torch.expm1(-int_beta(t)))**0.5
|
23 |
-
|
24 |
-
class SuperDiffSDXLPipeline(DiffusionPipeline, ConfigMixin):
|
25 |
-
"""SuperDiffSDXLPipeline."""
|
26 |
-
|
27 |
-
def __init__(self, unet: Callable, vae: Callable, text_encoder: Callable, text_encoder_2: Callable, tokenizer: Callable, tokenizer_2: Callable) -> None:
|
28 |
-
|
29 |
-
"""__init__.
|
30 |
-
|
31 |
-
Parameters
|
32 |
-
----------
|
33 |
-
model : Callable
|
34 |
-
model
|
35 |
-
vae : Callable
|
36 |
-
vae
|
37 |
-
text_encoder : Callable
|
38 |
-
text_encoder
|
39 |
-
scheduler : Callable
|
40 |
-
scheduler
|
41 |
-
tokenizer : Callable
|
42 |
-
tokenizer
|
43 |
-
kwargs :
|
44 |
-
kwargs
|
45 |
-
|
46 |
-
Returns
|
47 |
-
-------
|
48 |
-
None
|
49 |
-
|
50 |
-
"""
|
51 |
-
super().__init__()
|
52 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
53 |
-
|
54 |
-
vae.to(device)
|
55 |
-
unet.to(device)
|
56 |
-
text_encoder.to(device)
|
57 |
-
text_encoder_2.to(device)
|
58 |
-
|
59 |
-
|
60 |
-
self.register_modules(unet=unet,
|
61 |
-
vae=vae,
|
62 |
-
text_encoder=text_encoder,
|
63 |
-
text_encoder_2=text_encoder_2,
|
64 |
-
tokenizer=tokenizer,
|
65 |
-
tokenizer_2=tokenizer_2,
|
66 |
-
)
|
67 |
-
|
68 |
-
def prepare_prompt_input(self, prompt_o, prompt_b, batch_size, height, width):
|
69 |
-
text_input = self.tokenizer(prompt_o* batch_size, padding="max_length", max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt")
|
70 |
-
text_input_2 = self.tokenizer_2(prompt_o* batch_size, padding="max_length", max_length=self.tokenizer_2.model_max_length, truncation=True, return_tensors="pt")
|
71 |
-
with torch.no_grad():
|
72 |
-
text_embeddings = self.text_encoder(text_input.input_ids.to(self.device), output_hidden_states=True)
|
73 |
-
text_embeddings_2 = self.text_encoder_2(text_input_2.input_ids.to(self.device), output_hidden_states=True)
|
74 |
-
prompt_embeds_o = torch.concat((text_embeddings.hidden_states[-2], text_embeddings_2.hidden_states[-2]), dim=-1)
|
75 |
-
pooled_prompt_embeds_o = text_embeddings_2[0]
|
76 |
-
negative_prompt_embeds = torch.zeros_like(prompt_embeds_o)
|
77 |
-
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds_o)
|
78 |
-
|
79 |
-
text_input = self.tokenizer(prompt_b* batch_size, padding="max_length", max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt")
|
80 |
-
text_input_2 = self.tokenizer_2(prompt_b* batch_size, padding="max_length", max_length=self.tokenizer_2.model_max_length, truncation=True, return_tensors="pt")
|
81 |
-
with torch.no_grad():
|
82 |
-
text_embeddings = self.text_encoder(text_input.input_ids.to(self.device), output_hidden_states=True)
|
83 |
-
text_embeddings_2 = self.text_encoder_2(text_input_2.input_ids.to(self.device), output_hidden_states=True)
|
84 |
-
prompt_embeds_b = torch.concat((text_embeddings.hidden_states[-2], text_embeddings_2.hidden_states[-2]), dim=-1)
|
85 |
-
pooled_prompt_embeds_b = text_embeddings_2[0]
|
86 |
-
add_time_ids_o = torch.tensor([(height,width,0,0,height,width)])
|
87 |
-
add_time_ids_b = torch.tensor([(height,width,0,0,height,width)])
|
88 |
-
negative_add_time_ids = torch.tensor([(height,width,0,0,height,width)])
|
89 |
-
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds_o, prompt_embeds_b], dim=0)
|
90 |
-
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds_o, pooled_prompt_embeds_b], dim=0)
|
91 |
-
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids_o, add_time_ids_b], dim=0)
|
92 |
-
|
93 |
-
prompt_embeds = prompt_embeds.to(self.device)
|
94 |
-
add_text_embeds = add_text_embeds.to(self.device)
|
95 |
-
add_time_ids = add_time_ids.to(self.device).repeat(batch_size, 1)
|
96 |
-
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
97 |
-
return prompt_embeds, added_cond_kwargs
|
98 |
-
@torch.no_grad
|
99 |
-
def get_batch(self, latents: Callable, nrow: int, ncol: int) -> Callable:
|
100 |
-
"""get_batch.
|
101 |
-
|
102 |
-
Parameters
|
103 |
-
----------
|
104 |
-
latents : Callable
|
105 |
-
latents
|
106 |
-
nrow : int
|
107 |
-
nrow
|
108 |
-
ncol : int
|
109 |
-
ncol
|
110 |
-
|
111 |
-
Returns
|
112 |
-
-------
|
113 |
-
Callable
|
114 |
-
|
115 |
-
"""
|
116 |
-
image = self.vae.decode(
|
117 |
-
latents / self.vae.config.scaling_factor, return_dict=False
|
118 |
-
)[0]
|
119 |
-
image = (image / 2 + 0.5).clamp(0, 1).squeeze()
|
120 |
-
if len(image.shape) < 4:
|
121 |
-
image = image.unsqueeze(0)
|
122 |
-
image = (image.permute(0, 2, 3, 1) * 255).to(torch.uint8)
|
123 |
-
return image
|
124 |
-
|
125 |
-
@torch.no_grad
|
126 |
-
def get_text_embedding(self, prompt: str) -> Callable:
|
127 |
-
"""get_text_embedding.
|
128 |
-
|
129 |
-
Parameters
|
130 |
-
----------
|
131 |
-
prompt : str
|
132 |
-
prompt
|
133 |
-
|
134 |
-
Returns
|
135 |
-
-------
|
136 |
-
Callable
|
137 |
-
|
138 |
-
"""
|
139 |
-
text_input = self.tokenizer(
|
140 |
-
prompt,
|
141 |
-
padding="max_length",
|
142 |
-
max_length=self.tokenizer.model_max_length,
|
143 |
-
truncation=True,
|
144 |
-
return_tensors="pt",
|
145 |
-
)
|
146 |
-
return self.text_encoder(text_input.input_ids.to(self.device))[0]
|
147 |
-
|
148 |
-
@torch.no_grad
|
149 |
-
def get_vel(self, t: float, sigma: float, latents: Callable, embeddings: Callable):
|
150 |
-
"""get_vel.
|
151 |
-
|
152 |
-
Parameters
|
153 |
-
----------
|
154 |
-
t : float
|
155 |
-
t
|
156 |
-
sigma : float
|
157 |
-
sigma
|
158 |
-
latents : Callable
|
159 |
-
latents
|
160 |
-
embeddings : Callable
|
161 |
-
embeddings
|
162 |
-
"""
|
163 |
-
def v(_x, _e): return self.model(
|
164 |
-
_x / ((sigma**2 + 1) ** 0.5), t, encoder_hidden_states=_e
|
165 |
-
).sample
|
166 |
-
embeds = torch.cat(embeddings)
|
167 |
-
latent_input = latents
|
168 |
-
vel = v(latent_input, embeds)
|
169 |
-
return vel
|
170 |
-
|
171 |
-
def preprocess(
|
172 |
-
self,
|
173 |
-
prompt_1: str,
|
174 |
-
prompt_2: str,
|
175 |
-
seed: int = None,
|
176 |
-
num_inference_steps: int = 1000,
|
177 |
-
batch_size: int = 1,
|
178 |
-
lift: int = 0.0,
|
179 |
-
height: int = 512,
|
180 |
-
width: int = 512,
|
181 |
-
guidance_scale: int = 7.5,
|
182 |
-
) -> Callable:
|
183 |
-
"""preprocess.
|
184 |
-
|
185 |
-
Parameters
|
186 |
-
----------
|
187 |
-
prompt_1 : str
|
188 |
-
prompt_1
|
189 |
-
prompt_2 : str
|
190 |
-
prompt_2
|
191 |
-
seed : int
|
192 |
-
seed
|
193 |
-
num_inference_steps : int
|
194 |
-
num_inference_steps
|
195 |
-
batch_size : int
|
196 |
-
batch_size
|
197 |
-
lift : int
|
198 |
-
lift
|
199 |
-
height : int
|
200 |
-
height
|
201 |
-
width : int
|
202 |
-
width
|
203 |
-
guidance_scale : int
|
204 |
-
guidance_scale
|
205 |
-
|
206 |
-
Returns
|
207 |
-
-------
|
208 |
-
Callable
|
209 |
-
|
210 |
-
"""
|
211 |
-
# Tokenize the input
|
212 |
-
self.batch_size = batch_size
|
213 |
-
self.num_inference_steps = num_inference_steps
|
214 |
-
self.guidance_scale = guidance_scale
|
215 |
-
self.lift = lift
|
216 |
-
self.seed = seed
|
217 |
-
if self.seed is None:
|
218 |
-
self.seed = random.randint(0, 2**32 - 1)
|
219 |
-
|
220 |
-
#obj_prompt = [prompt_1]
|
221 |
-
#bg_prompt = [prompt_2]
|
222 |
-
#obj_embeddings = self.get_text_embedding(obj_prompt * batch_size)
|
223 |
-
#bg_embeddings = self.get_text_embedding(bg_prompt * batch_size)
|
224 |
-
|
225 |
-
#uncond_embeddings = self.get_text_embedding([""] * batch_size)
|
226 |
-
|
227 |
-
generator = torch.cuda.manual_seed(
|
228 |
-
self.seed
|
229 |
-
) # Seed generator to create the initial latent noise
|
230 |
-
latents = torch.randn((batch_size, self.unet.in_channels, height // 8, width // 8), generator=generator, dtype=self.dtype, device=self.device,)
|
231 |
-
prompt_embeds, added_cond_kwargs = self.prepare_prompt_input(prompt_1, prompt_2, batch_size, height, width)
|
232 |
-
#latents = torch.randn(
|
233 |
-
# (batch_size, self.model.config.in_channels, height // 8, width // 8),
|
234 |
-
# generator=generator,
|
235 |
-
# device=self.device,
|
236 |
-
#)
|
237 |
-
|
238 |
-
#latents_og = latents.clone().detach()
|
239 |
-
#latents_uncond_og = latents.clone().detach()
|
240 |
-
|
241 |
-
#self.scheduler.set_timesteps(num_inference_steps)
|
242 |
-
#latents = latents * self.scheduler.init_noise_sigma
|
243 |
-
|
244 |
-
#latents_uncond = latents.clone().detach()
|
245 |
-
return {
|
246 |
-
"latents": latents,
|
247 |
-
"prompt_embeds": prompt_embeds,
|
248 |
-
"added_cond_kwargs": added_cond_kwargs,
|
249 |
-
}
|
250 |
-
|
251 |
-
def _forward(self, model_inputs: Dict) -> Callable:
|
252 |
-
"""_forward.
|
253 |
-
|
254 |
-
Parameters
|
255 |
-
----------
|
256 |
-
model_inputs : Dict
|
257 |
-
model_inputs
|
258 |
-
|
259 |
-
Returns
|
260 |
-
-------
|
261 |
-
Callable
|
262 |
-
|
263 |
-
"""
|
264 |
-
latents = model_inputs["latents"]
|
265 |
-
prompt_embeds = model_inputs["prompt_embeds"]
|
266 |
-
added_cond_kwargs = model_inputs["added_cond_kwargs"]
|
267 |
-
|
268 |
-
t = torch.tensor(1.0)
|
269 |
-
dt = 1.0/self.num_inference_steps
|
270 |
-
train_number_steps = 1000
|
271 |
-
latents = latents * (sigma(t)**2+1)**0.5
|
272 |
-
with torch.no_grad():
|
273 |
-
for i in tqdm(range(self.num_inference_steps)):
|
274 |
-
latent_model_input = torch.cat([latents] * 3)
|
275 |
-
sigma_t = sigma(t)
|
276 |
-
dsigma = sigma(t-dt) - sigma_t
|
277 |
-
latent_model_input /= (sigma_t**2+1)**0.5
|
278 |
-
with torch.no_grad():
|
279 |
-
noise_pred = self.unet(latent_model_input, t*train_number_steps, encoder_hidden_states=prompt_embeds, added_cond_kwargs=added_cond_kwargs, return_dict=False)[0]
|
280 |
-
|
281 |
-
noise_pred_uncond, noise_pred_text_o, noise_pred_text_b = noise_pred.chunk(3)
|
282 |
-
|
283 |
-
noise = torch.sqrt(2*torch.abs(dsigma)*sigma_t)*torch.randn_like(latents)
|
284 |
-
|
285 |
-
|
286 |
-
dx_ind = 2*dsigma*(noise_pred_uncond + self.guidance_scale*(noise_pred_text_b - noise_pred_uncond)) + noise
|
287 |
-
kappa = (torch.abs(dsigma)*(noise_pred_text_b-noise_pred_text_o)*(noise_pred_text_b+noise_pred_text_o)).sum((1,2,3))-(dx_ind*((noise_pred_text_o-noise_pred_text_b))).sum((1,2,3))
|
288 |
-
kappa /= 2*dsigma*self.guidance_scale*((noise_pred_text_o-noise_pred_text_b)**2).sum((1,2,3))
|
289 |
-
noise_pred = noise_pred_uncond + self.guidance_scale*((noise_pred_text_b - noise_pred_uncond) + kappa[:,None,None,None]*(noise_pred_text_o-noise_pred_text_b))
|
290 |
-
|
291 |
-
latents += 2*dsigma * noise_pred + noise
|
292 |
-
t -= dt
|
293 |
-
return latents
|
294 |
-
|
295 |
-
def postprocess(self, latents: Callable) -> Callable:
|
296 |
-
"""postprocess.
|
297 |
-
|
298 |
-
Parameters
|
299 |
-
----------
|
300 |
-
latents : Callable
|
301 |
-
latents
|
302 |
-
|
303 |
-
Returns
|
304 |
-
-------
|
305 |
-
Callable
|
306 |
-
|
307 |
-
"""
|
308 |
-
latents = latents/self.vae.config.scaling_factor
|
309 |
-
latents = latents.to(torch.float32)
|
310 |
-
with torch.no_grad():
|
311 |
-
image = self.vae.decode(latents, return_dict=False)[0]
|
312 |
-
|
313 |
-
image = (image / 2 + 0.5).clamp(0, 1)
|
314 |
-
image = image.detach().cpu().permute(0, 2, 3, 1).numpy()
|
315 |
-
images = (image * 255).round().astype("uint8")
|
316 |
-
return images
|
317 |
-
|
318 |
-
def __call__(
|
319 |
-
self,
|
320 |
-
prompt_1: str,
|
321 |
-
prompt_2: str,
|
322 |
-
seed: int = None,
|
323 |
-
num_inference_steps: int = 1000,
|
324 |
-
batch_size: int = 1,
|
325 |
-
lift: int = 0.0,
|
326 |
-
height: int = 1024,
|
327 |
-
width: int = 1024,
|
328 |
-
guidance_scale: int = 7.5,
|
329 |
-
) -> Callable:
|
330 |
-
"""__call__.
|
331 |
-
|
332 |
-
Parameters
|
333 |
-
----------
|
334 |
-
prompt_1 : str
|
335 |
-
prompt_1
|
336 |
-
prompt_2 : str
|
337 |
-
prompt_2
|
338 |
-
seed : int
|
339 |
-
seed
|
340 |
-
num_inference_steps : int
|
341 |
-
num_inference_steps
|
342 |
-
batch_size : int
|
343 |
-
batch_size
|
344 |
-
lift : int
|
345 |
-
lift
|
346 |
-
height : int
|
347 |
-
height
|
348 |
-
width : int
|
349 |
-
width
|
350 |
-
guidance_scale : int
|
351 |
-
guidance_scale
|
352 |
-
|
353 |
-
Returns
|
354 |
-
-------
|
355 |
-
Callable
|
356 |
-
|
357 |
-
"""
|
358 |
-
# Preprocess inputs
|
359 |
-
model_inputs = self.preprocess(
|
360 |
-
prompt_1,
|
361 |
-
prompt_2,
|
362 |
-
seed,
|
363 |
-
num_inference_steps,
|
364 |
-
batch_size,
|
365 |
-
lift,
|
366 |
-
height,
|
367 |
-
width,
|
368 |
-
guidance_scale,
|
369 |
-
)
|
370 |
-
|
371 |
-
# Forward pass through the pipeline
|
372 |
-
latents = self._forward(model_inputs)
|
373 |
-
|
374 |
-
# Postprocess to generate the final output
|
375 |
-
images = self.postprocess(latents)
|
376 |
-
return images
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|