File size: 10,883 Bytes
e80c177 e4e13db e80c177 b4f7900 e80c177 a8cf307 e80c177 0b0d11b e80c177 e4e13db 1732305 6a798c0 e4e13db a8cf307 6a798c0 e4e13db e80c177 0b0d11b e80c177 0b0d11b e80c177 e4e13db e80c177 89b9678 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 |
import random
from tqdm import tqdm
from typing import Callable, Dict, List, Optional
import torch
from diffusers import DiffusionPipeline
from diffusers.configuration_utils import ConfigMixin
class SuperDiffPipeline(DiffusionPipeline, ConfigMixin):
"""SuperDiffPipeline."""
def __init__(self, unet: Callable, vae: Callable, text_encoder: Callable, scheduler: Callable, tokenizer: Callable) -> None:
"""__init__.
Parameters
----------
unet : Callable
unet
vae : Callable
vae
text_encoder : Callable
text_encoder
scheduler : Callable
scheduler
tokenizer : Callable
tokenizer
kwargs :
kwargs
Returns
-------
None
"""
super().__init__()
# Register additional parameters for flexibility
# Explicitly assign required components
#self.unet = unet
#self.vae = vae
#self.text_encoder = text_encoder
#self.tokenizer = tokenizer
#self.scheduler = scheduler
device = "cuda" if torch.cuda.is_available() else "cpu"
vae.to(device)
unet.to(device)
text_encoder.to(device)
self.register_modules(unet=unet,
scheduler=scheduler,
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,)
#self.register_to_config(
# vae=vae.__class__.__name__,
# scheduler=scheduler.__class__.__name__,
# tokenizer=tokenizer.__class__.__name__,
# unet=unet.__class__.__name__,
# text_encoder=text_encoder.__class__.__name__,
# device=device,
# batch_size=None,
# num_inference_steps=None,
# guidance_scale=None,
# lift=None,
# seed=None,
#)
@torch.no_grad
def get_batch(self, latents: Callable, nrow: int, ncol: int) -> Callable:
"""get_batch.
Parameters
----------
latents : Callable
latents
nrow : int
nrow
ncol : int
ncol
Returns
-------
Callable
"""
image = self.vae.decode(
latents / self.vae.config.scaling_factor, return_dict=False
)[0]
image = (image / 2 + 0.5).clamp(0, 1).squeeze()
if len(image.shape) < 4:
image = image.unsqueeze(0)
image = (image.permute(0, 2, 3, 1) * 255).to(torch.uint8)
return image
@torch.no_grad
def get_text_embedding(self, prompt: str) -> Callable:
"""get_text_embedding.
Parameters
----------
prompt : str
prompt
Returns
-------
Callable
"""
text_input = self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
return self.text_encoder(text_input.input_ids.to(self.device))[0]
@torch.no_grad
def get_vel(self, t: float, sigma: float, latents: Callable, embeddings: Callable):
"""get_vel.
Parameters
----------
t : float
t
sigma : float
sigma
latents : Callable
latents
embeddings : Callable
embeddings
"""
def v(_x, _e): return self.unet(
_x / ((sigma**2 + 1) ** 0.5), t, encoder_hidden_states=_e
).sample
embeds = torch.cat(embeddings)
latent_input = latents
vel = v(latent_input, embeds)
return vel
def preprocess(
self,
prompt_1: str,
prompt_2: str,
seed: int = None,
num_inference_steps: int = 1000,
batch_size: int = 1,
lift: int = 0.0,
height: int = 512,
width: int = 512,
guidance_scale: int = 7.5,
) -> Callable:
"""preprocess.
Parameters
----------
prompt_1 : str
prompt_1
prompt_2 : str
prompt_2
seed : int
seed
num_inference_steps : int
num_inference_steps
batch_size : int
batch_size
lift : int
lift
height : int
height
width : int
width
guidance_scale : int
guidance_scale
Returns
-------
Callable
"""
# Tokenize the input
self.batch_size = batch_size
self.num_inference_steps = num_inference_steps
self.guidance_scale = guidance_scale
self.lift = lift
self.seed = seed
if self.seed is None:
self.seed = random.randint(0, 2**32 - 1)
obj_prompt = [prompt_1]
bg_prompt = [prompt_2]
obj_embeddings = self.get_text_embedding(obj_prompt * batch_size)
bg_embeddings = self.get_text_embedding(bg_prompt * batch_size)
uncond_embeddings = self.get_text_embedding([""] * batch_size)
generator = torch.cuda.manual_seed(
self.seed
) # Seed generator to create the initial latent noise
latents = torch.randn(
(batch_size, self.unet.config.in_channels, height // 8, width // 8),
generator=generator,
device=self.device,
)
latents_og = latents.clone().detach()
latents_uncond_og = latents.clone().detach()
self.scheduler.set_timesteps(num_inference_steps)
latents = latents * self.scheduler.init_noise_sigma
latents_uncond = latents.clone().detach()
return {
"latents": latents,
"obj_embeddings": obj_embeddings,
"uncond_embeddings": uncond_embeddings,
"bg_embeddings": bg_embeddings,
}
def _forward(self, model_inputs: Dict) -> Callable:
"""_forward.
Parameters
----------
model_inputs : Dict
model_inputs
Returns
-------
Callable
"""
latents = model_inputs["latents"]
obj_embeddings = model_inputs["obj_embeddings"]
uncond_embeddings = model_inputs["uncond_embeddings"]
bg_embeddings = model_inputs["bg_embeddings"]
kappa = 0.5 * torch.ones(
(self.num_inference_steps + 1, self.batch_size), device=self.device
)
ll_obj = torch.ones(
(self.num_inference_steps + 1, self.batch_size), device=self.device
)
ll_bg = torch.ones(
(self.num_inference_steps + 1, self.batch_size), device=self.device
)
ll_uncond = torch.ones(
(self.num_inference_steps + 1, self.batch_size), device=self.device
)
with torch.no_grad():
for i, t in tqdm(enumerate(self.scheduler.timesteps)):
dsigma = self.scheduler.sigmas[i +
1] - self.scheduler.sigmas[i]
sigma = self.scheduler.sigmas[i]
vel_obj = self.get_vel(t, sigma, latents, [obj_embeddings])
vel_uncond = self.get_vel(
t, sigma, latents, [uncond_embeddings])
vel_bg = self.get_vel(t, sigma, latents, [bg_embeddings])
noise = torch.sqrt(2 * torch.abs(dsigma) * sigma) * torch.randn_like(
latents
)
dx_ind = (
2
* dsigma
* (vel_uncond + self.guidance_scale * (vel_bg - vel_uncond))
+ noise
)
kappa[i + 1] = (
(torch.abs(dsigma) * (vel_bg - vel_obj) * (vel_bg + vel_obj)).sum(
(1, 2, 3)
)
- (dx_ind * ((vel_obj - vel_bg))).sum((1, 2, 3))
+ sigma * self.lift / self.num_inference_steps
)
kappa[i + 1] /= (
2
* dsigma
* self.guidance_scale
* ((vel_obj - vel_bg) ** 2).sum((1, 2, 3))
)
vf = vel_uncond + self.guidance_scale * (
(vel_bg - vel_uncond)
+ kappa[i + 1][:, None, None, None] * (vel_obj - vel_bg)
)
dx = 2 * dsigma * vf + noise
latents += dx
ll_obj[i + 1] = ll_obj[i] + (
-torch.abs(dsigma) / sigma * (vel_obj) ** 2
- (dx * (vel_obj / sigma))
).sum((1, 2, 3))
ll_bg[i + 1] = ll_bg[i] + (
-torch.abs(dsigma) / sigma * (vel_bg) ** 2 -
(dx * (vel_bg / sigma))
).sum((1, 2, 3))
return latents
def postprocess(self, latents: Callable) -> Callable:
"""postprocess.
Parameters
----------
latents : Callable
latents
Returns
-------
Callable
"""
image = self.get_batch(latents, 1, self.batch_size)
# Ensure the shape is (height, width, 3)
assert image.shape[-1] == 3 # Handle grayscale or invalid shapes
# Convert to uint8 if not already
image = image.to(torch.uint8) # Ensure it's uint8 for PIL
return image
def __call__(
self,
prompt_1: str,
prompt_2: str,
seed: int = None,
num_inference_steps: int = 1000,
batch_size: int = 1,
lift: int = 0.0,
height: int = 512,
width: int = 512,
guidance_scale: int = 7.5,
) -> Callable:
"""__call__.
Parameters
----------
prompt_1 : str
prompt_1
prompt_2 : str
prompt_2
seed : int
seed
num_inference_steps : int
num_inference_steps
batch_size : int
batch_size
lift : int
lift
height : int
height
width : int
width
guidance_scale : int
guidance_scale
Returns
-------
Callable
"""
# Preprocess inputs
model_inputs = self.preprocess(
prompt_1,
prompt_2,
seed,
num_inference_steps,
batch_size,
lift,
height,
width,
guidance_scale,
)
# Forward pass through the pipeline
latents = self._forward(model_inputs)
# Postprocess to generate the final output
images = self.postprocess(latents)
return images
|