File size: 25,285 Bytes
6fc683c |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 |
import argparse
import gc
import os
from torchvision.transforms import InterpolationMode
BICUBIC = InterpolationMode.BICUBIC
from tiktoken.core import Encoding
from torchvision.transforms import CenterCrop, Compose, Resize
from diffusers import ControlNetModel
from diffusers.schedulers import DPMSolverMultistepScheduler, PNDMScheduler, DDIMScheduler
from app_utils import *
from controlnet.preprocessor import ControlNet_Preprocessor
from fairseq import checkpoint_utils, tasks, utils
from fairseq.dataclass.utils import convert_namespace_to_omegaconf
from unilm.data.vl.openimage_loader import NumpyNormalize
class AppModel:
def __init__(self, cfg):
if isinstance(cfg, argparse.Namespace):
cfg = convert_namespace_to_omegaconf(cfg)
cfg.model.align = False
cfg.model.checkpoint_activations = False
utils.import_user_module(cfg.common)
task = tasks.setup_task(cfg.task)
model = task.build_model(cfg.model)
model.freeze_params(model.parameters())
model.half()
model.cuda()
model.eval()
self.cfg = cfg
self.model = model
self.model_cache = {}
self.bos_id = task.dictionary.bos()
self.eos_id = task.dictionary.eos()
self.boi_id = task.dictionary.index(BOI_SYMBOL)
self.eoi_id = task.dictionary.index(EOI_SYMBOL)
self.dictionary = task.dictionary
self.tokenizer = task.tokenizer
self.text_transform = self._build_text_transform()
self.image_transform = self._build_image_transform()
self.task_name = ""
self.controlnet = None
self.controlnet_preprocessor = ControlNet_Preprocessor()
def _build_image_transform(self):
preprocess_image = {
'gpt': Compose([
Resize(224, interpolation=BICUBIC),
CenterCrop(224),
NumpyNormalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
]),
'diff': Compose([
Resize(512),
CenterCrop(512),
NumpyNormalize([0.5], [0.5])
])
}
return preprocess_image
def _build_text_transform(self):
def text_transform(text):
append_eos = False
fs_dict = self.dictionary
if isinstance(self.tokenizer, Encoding):
words = list(map(str, self.tokenizer.encode(text, allowed_special="all")))
else:
words = self.tokenizer.encode(text, out_type=str)
# ids = [fs_dict.bos_index]
ids = []
for i, word in enumerate(words):
idx = fs_dict.index(word)
ids.append(idx)
if append_eos:
ids.append(fs_dict.eos_index)
return ids
return text_transform
def load_controlnet_weight(self, task_name):
if task_name == self.task_name:
return
torch.cuda.empty_cache()
gc.collect()
self.controlnet = ControlNetModel.from_pretrained(CONTROLNET_MODEL_IDS[task_name], torch_dtype=torch.float16)
self.model.freeze_params(self.controlnet.parameters())
self.controlnet.cuda()
self.controlnet.eval()
torch.cuda.empty_cache()
gc.collect()
self.task_name = task_name
def get_available_lora(self):
# traverse all the available lora in cfg.model.lora_dir in created time order
if self.cfg.model.lora_dir == '':
return []
files = [f for f in os.listdir(self.cfg.model.lora_dir) if
os.path.isfile(os.path.join(self.cfg.model.lora_dir, f))]
files_sorted = sorted(files, key=lambda x: os.path.getctime(os.path.join(self.cfg.model.lora_dir, x)))
return files_sorted
def load_lora(self, lora_name):
if lora_name == 'None':
if self.cfg.model.lora_name != 'None':
for _, module in self.model.diffusion_unet.unet.named_modules():
if hasattr(module, "set_lora_layer"):
module.set_lora_layer(None)
self.model.diffusion_unet.lora = False
self.cfg.model.lora_name = 'None'
return 'None'
try:
state_dict, network_alphas = self.model.diffusion_unet.lora_state_dict(
os.path.join(self.cfg.model.lora_dir, lora_name)
)
self.model.diffusion_unet.load_lora_into_unet(
state_dict, network_alphas=network_alphas, unet=self.model.diffusion_unet.unet, low_cpu_mem_usage=True
)
self.model.diffusion_unet.lora = True
self.cfg.model.lora_name = lora_name
return lora_name
except:
return 'None'
def set_ckpt_scheduler_fn(self, ckpt, scheduler):
# reset scheduler if the class is changed
if scheduler == 'dpms' and not isinstance(self.model.diffusion_unet.scheduler, DPMSolverMultistepScheduler):
self.model.diffusion_unet.scheduler = DPMSolverMultistepScheduler.from_pretrained(
self.cfg.model.pretrained_model_name_or_path, subfolder="scheduler", torch_dtype=torch.float16,
revision="fp16"
)
elif scheduler == 'pndm' and not isinstance(self.model.diffusion_unet.scheduler, PNDMScheduler):
self.model.diffusion_unet.scheduler = PNDMScheduler.from_pretrained(
self.cfg.model.pretrained_model_name_or_path, subfolder="scheduler", torch_dtype=torch.float16,
revision="fp16"
)
elif scheduler == 'ddim' and not isinstance(self.model.diffusion_unet.scheduler, DDIMScheduler):
self.model.diffusion_unet.scheduler = DDIMScheduler.from_pretrained(
self.cfg.model.pretrained_model_name_or_path, subfolder="scheduler", torch_dtype=torch.float16,
revision="fp16"
)
if ckpt != self.cfg.model.pretrained_ckpt_path:
try:
if ckpt not in self.model_cache:
state = checkpoint_utils.load_checkpoint_to_cpu(ckpt)
self.model_cache[ckpt] = state["model"]
msg = self.model.load_state_dict(self.model_cache[ckpt], strict=False, args=self.cfg.model)
self.cfg.model.pretrained_ckpt_path = ckpt
except:
if ckpt in self.model_cache:
del self.model_cache[ckpt]
exp = self.cfg.model.pretrained_ckpt_path.split('/')[-2]
pt = self.cfg.model.pretrained_ckpt_path.split('/')[-1].split('.')[0].split('_')[-1]
return f'exp: {exp} pt: {pt}'
def kosmosg_preprocess(self, prompt, negative_prompt, *args, single_batch=True):
img_src_tokens = [im for im in args if im is not None]
assert len(img_src_tokens) == prompt.count('<i>'), \
"Number of images in prompt does not match the number of images uploaded"
gpt_img_src_tokens = [torch.tensor(self.image_transform['gpt'](im)) for im in img_src_tokens]
src_tokens = [self.bos_id]
img_gpt_input_mask = [0]
for i in range(len(img_src_tokens)):
text_snippet = prompt.split('<i>', 1)[0]
prompt = prompt.split('<i>', 1)[1]
text_token = self.text_transform(text_snippet)
src_tokens.extend(text_token + [self.boi_id] * (self.cfg.task.image_token_length + 1) + [self.eoi_id])
img_gpt_input_mask.extend([0] * len(text_token) + [0] + [1] * self.cfg.task.image_token_length + [0])
text_token = self.text_transform(prompt)
src_tokens.extend(text_token)
img_gpt_input_mask.extend([0] * len(text_token))
src_tokens = torch.LongTensor(src_tokens)
gpt_img_src_tokens = torch.stack(gpt_img_src_tokens).to(torch.float16) \
if len(gpt_img_src_tokens) > 0 else None
img_gpt_input_mask = torch.tensor(img_gpt_input_mask, dtype=torch.bool)
negative_tokens = torch.LongTensor([self.bos_id] + self.text_transform(negative_prompt))
if single_batch:
return src_tokens.unsqueeze(0), gpt_img_src_tokens, img_gpt_input_mask.unsqueeze(0), \
negative_tokens.unsqueeze(0)
else:
return src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens
@torch.inference_mode()
def kosmosg_generation(self, prompt, lora_scale, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, *args):
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': lora_scale,
}
image = self.model.sample(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens, **kwargs)
return image
@torch.inference_mode()
def controlnet_generation_canny(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, low_threshold,
high_threshold, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_canny(control_image, image_resolution, low_threshold,
high_threshold)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('Canny')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_mlsd(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, preprocess_resolution,
value_threshold, distance_threshold, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_mlsd(control_image, image_resolution,
preprocess_resolution, value_threshold,
distance_threshold)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('MLSD')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_scribble(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, preprocess_resolution,
preprocessor_name, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_scribble(control_image, image_resolution,
preprocess_resolution, preprocessor_name)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('scribble')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_scribble_interactive(self, prompt, num_inference_steps, text_guidance_scale,
negative_prompt, num_images_per_prompt, control_image_and_mask,
image_resolution, *args):
assert control_image_and_mask is not None, 'Image is required'
control_image = self.controlnet_preprocessor.preprocess_scribble_interactive(control_image_and_mask,
image_resolution)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('scribble')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_softedge(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, preprocess_resolution,
preprocessor_name, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_softedge(control_image, image_resolution,
preprocess_resolution, preprocessor_name)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('softedge')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_openpose(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, preprocess_resolution,
preprocessor_name, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_openpose(control_image, image_resolution,
preprocess_resolution, preprocessor_name)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('Openpose')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_segmentation(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution,
preprocess_resolution, preprocessor_name, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_segmentation(control_image, image_resolution,
preprocess_resolution, preprocessor_name)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('segmentation')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_depth(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, preprocess_resolution,
preprocessor_name, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_depth(control_image, image_resolution,
preprocess_resolution, preprocessor_name)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('depth')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_normal(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, preprocess_resolution,
preprocessor_name, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_normal(control_image, image_resolution,
preprocess_resolution, preprocessor_name)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('NormalBae')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_lineart(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, preprocess_resolution,
preprocessor_name, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_lineart(control_image, image_resolution,
preprocess_resolution, preprocessor_name)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
if 'anime' in preprocessor_name:
self.load_controlnet_weight('lineart_anime')
else:
self.load_controlnet_weight('lineart')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, **kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_shuffle(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, preprocessor_name, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_shuffle(control_image, image_resolution,
preprocessor_name)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('shuffle')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, kwargs)
return [control_image] + image
@torch.inference_mode()
def controlnet_generation_ip2p(self, prompt, num_inference_steps, text_guidance_scale, negative_prompt,
num_images_per_prompt, control_image, image_resolution, *args):
assert control_image is not None, 'Image is required'
assert image_resolution <= MAX_IMAGE_RESOLUTION, 'Image resolution is too high'
control_image = self.controlnet_preprocessor.preprocess_ip2p(control_image, image_resolution)
src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens = \
self.kosmosg_preprocess(prompt, negative_prompt, *args)
self.load_controlnet_weight('ip2p')
kwargs = {
'num_inference_steps': num_inference_steps,
'text_guidance_scale': text_guidance_scale,
'num_images_per_prompt': num_images_per_prompt,
'lora_scale': 0.0,
}
image = self.model.sample_controlnet(src_tokens, gpt_img_src_tokens, img_gpt_input_mask, negative_tokens,
control_image, self.controlnet, kwargs)
return [control_image] + image
|