File size: 23,407 Bytes
8741abe |
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 |
import os
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "False"
os.environ["TOKENIZERS_PARALLELISM"] = "true"
import numpy as np
import gradio as gr
import torch
from PIL import Image
from omegaconf import OmegaConf
from transformers import AutoTokenizer
import torch.nn.functional as F
from transformers import CLIPImageProcessor
import sys
sys.path.insert(0, ".")
from training import conversation as conversation_lib
from prompting_utils import UniversalPrompting, create_attention_mask_predict_next, create_attention_mask_for_mmu
from training_utils import image_transform
from models import Showo, MAGVITv2, get_mask_chedule, CLIPVisionTower
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
conversation_lib.default_conversation = conversation_lib.conv_templates["phi1.5"]
SYSTEM_PROMPT = "A chat between a curious user and an artificial intelligence assistant. " \
"The assistant gives helpful, detailed, and polite answers to the user's questions."
SYSTEM_PROMPT_LEN = 28
config = OmegaConf.load("configs/showo_demo.yaml")
tokenizer = AutoTokenizer.from_pretrained(config.model.showo.llm_model_path, padding_side="left")
uni_prompting = UniversalPrompting(tokenizer, max_text_len=config.dataset.preprocessing.max_seq_length,
special_tokens=("<|soi|>", "<|eoi|>", "<|sov|>", "<|eov|>", "<|t2i|>", "<|mmu|>",
"<|t2v|>", "<|v2v|>", "<|lvg|>"),
ignore_id=-100, cond_dropout_prob=config.training.cond_dropout_prob)
vq_model = MAGVITv2()
vq_model = vq_model.from_pretrained(config.model.vq_model.vq_model_name).to(device)
vq_model.requires_grad_(False)
vq_model.eval()
model = Showo.from_pretrained(config.model.showo.pretrained_model_path).to(device)
model.eval()
mask_token_id = model.config.mask_token_id
def text_to_image_generation(input_text, guidance_scale, generation_timesteps):
prompts = [input_text]
config.training.batch_size = config.batch_size = 1
config.training.guidance_scale = config.guidance_scale = guidance_scale
config.training.generation_timesteps = config.generation_timesteps = generation_timesteps
image_tokens = torch.ones((len(prompts), config.model.showo.num_vq_tokens),
dtype=torch.long, device=device) * mask_token_id
input_ids, _ = uni_prompting((prompts, image_tokens), 't2i_gen')
if config.training.guidance_scale > 0:
uncond_input_ids, _ = uni_prompting(([''] * len(prompts), image_tokens), 't2i_gen')
attention_mask = create_attention_mask_predict_next(torch.cat([input_ids, uncond_input_ids], dim=0),
pad_id=int(uni_prompting.sptids_dict['<|pad|>']),
soi_id=int(uni_prompting.sptids_dict['<|soi|>']),
eoi_id=int(uni_prompting.sptids_dict['<|eoi|>']),
rm_pad_in_image=True)
else:
attention_mask = create_attention_mask_predict_next(input_ids,
pad_id=int(uni_prompting.sptids_dict['<|pad|>']),
soi_id=int(uni_prompting.sptids_dict['<|soi|>']),
eoi_id=int(uni_prompting.sptids_dict['<|eoi|>']),
rm_pad_in_image=True)
uncond_input_ids = None
if config.get("mask_schedule", None) is not None:
schedule = config.mask_schedule.schedule
args = config.mask_schedule.get("params", {})
mask_schedule = get_mask_chedule(schedule, **args)
else:
mask_schedule = get_mask_chedule(config.training.get("mask_schedule", "cosine"))
with torch.no_grad():
gen_token_ids = model.t2i_generate(
input_ids=input_ids,
uncond_input_ids=uncond_input_ids,
attention_mask=attention_mask,
guidance_scale=config.training.guidance_scale,
temperature=config.training.get("generation_temperature", 1.0),
timesteps=config.training.generation_timesteps,
noise_schedule=mask_schedule,
noise_type=config.training.get("noise_type", "mask"),
seq_len=config.model.showo.num_vq_tokens,
uni_prompting=uni_prompting,
config=config,
)
gen_token_ids = torch.clamp(gen_token_ids, max=config.model.showo.codebook_size - 1, min=0)
images = vq_model.decode_code(gen_token_ids)
images = torch.clamp((images + 1.0) / 2.0, min=0.0, max=1.0)
images *= 255.0
images = images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
return images[0]
def text_guided_inpainting(input_text, inpainting_image, inpainting_mask, guidance_scale, generation_timesteps):
prompt = [input_text]
config.training.batch_size = config.batch_size = 1
config.training.guidance_scale = config.guidance_scale = guidance_scale
config.training.generation_timesteps = config.generation_timesteps = generation_timesteps
inpainting_image = image_transform(inpainting_image, resolution=config.dataset.params.resolution).to(device)
inpainting_mask = image_transform(inpainting_mask, resolution=config.dataset.params.resolution, normalize=False)
inpainting_image = inpainting_image.unsqueeze(0).repeat(config.training.batch_size, 1, 1, 1)
inpainting_mask = inpainting_mask.unsqueeze(0).to(device)
inpainting_mask = F.interpolate(inpainting_mask, size=config.dataset.params.resolution // 16, mode='bicubic')
inpainting_mask = inpainting_mask.repeat(config.training.batch_size, 1, 1, 1)
inpainting_mask[inpainting_mask < 0.5] = 0
inpainting_mask[inpainting_mask >= 0.5] = 1
inpainting_mask = inpainting_mask.reshape(config.training.batch_size, -1)
inpainting_mask = inpainting_mask.to(torch.bool)
inpainting_image_tokens = vq_model.get_code(inpainting_image) + len(uni_prompting.text_tokenizer)
inpainting_image_tokens[inpainting_mask] = mask_token_id
input_ids, _ = uni_prompting((prompt, inpainting_image_tokens), 't2i_gen')
if config.training.guidance_scale > 0:
uncond_input_ids, _ = uni_prompting(([''] * len(prompt), inpainting_image_tokens), 't2i_gen')
attention_mask = create_attention_mask_predict_next(torch.cat([input_ids, uncond_input_ids], dim=0),
pad_id=int(uni_prompting.sptids_dict['<|pad|>']),
soi_id=int(uni_prompting.sptids_dict['<|soi|>']),
eoi_id=int(uni_prompting.sptids_dict['<|eoi|>']),
rm_pad_in_image=True)
else:
attention_mask = create_attention_mask_predict_next(input_ids,
pad_id=int(uni_prompting.sptids_dict['<|pad|>']),
soi_id=int(uni_prompting.sptids_dict['<|soi|>']),
eoi_id=int(uni_prompting.sptids_dict['<|eoi|>']),
rm_pad_in_image=True)
uncond_input_ids = None
if config.get("mask_schedule", None) is not None:
schedule = config.mask_schedule.schedule
args = config.mask_schedule.get("params", {})
mask_schedule = get_mask_chedule(schedule, **args)
else:
mask_schedule = get_mask_chedule(config.training.get("mask_schedule", "cosine"))
with torch.no_grad():
gen_token_ids = model.t2i_generate(
input_ids=input_ids,
uncond_input_ids=uncond_input_ids,
attention_mask=attention_mask,
guidance_scale=config.training.guidance_scale,
temperature=config.training.get("generation_temperature", 1.0),
timesteps=config.training.generation_timesteps,
noise_schedule=mask_schedule,
noise_type=config.training.get("noise_type", "mask"),
seq_len=config.model.showo.num_vq_tokens,
uni_prompting=uni_prompting,
config=config,
)
gen_token_ids = torch.clamp(gen_token_ids, max=config.model.showo.codebook_size - 1, min=0)
images = vq_model.decode_code(gen_token_ids)
images = torch.clamp((images + 1.0) / 2.0, min=0.0, max=1.0)
images *= 255.0
images = images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
return images[0]
def text_guided_extrapolation(input_img, input_text, left_ext, right_ext, guidance_scale, generation_timesteps):
config.offset = 0
config.training.batch_size = config.batch_size = 1
config.training.guidance_scale = config.guidance_scale = guidance_scale
config.training.generation_timesteps = config.generation_timesteps = generation_timesteps
extra_direction = ['right'] * int(right_ext) + ['left'] * int(left_ext)
prompt = [input_text] * len(extra_direction)
W = config.dataset.params.resolution // 16
for id, (prt, direction) in enumerate(zip(prompt, extra_direction)):
prt = [prt] * config.training.batch_size
if id == 0:
# extrapolation_image = Image.open(config.image_path).convert("RGB")
extrapolation_image = input_img
extrapolation_image = image_transform(extrapolation_image,
resolution=config.dataset.params.resolution).to(device)
B, _, _ = extrapolation_image.shape
extrapolation_image = extrapolation_image.unsqueeze(0)
extrapolation_image_tokens = vq_model.get_code(extrapolation_image) + len(uni_prompting.text_tokenizer)
extrapolation_image_tokens = extrapolation_image_tokens.reshape(1,
config.dataset.params.resolution // 16,
config.dataset.params.resolution // 16)
extrapolation_image_tokens = extrapolation_image_tokens.repeat(config.training.batch_size, 1, 1)
else:
extrapolation_image_tokens = gen_token_ids + len(uni_prompting.text_tokenizer)
image_left_part = extrapolation_image_tokens[:, :, :-(W // 2 - config.offset)] - len(
uni_prompting.text_tokenizer)
image_right_part = extrapolation_image_tokens[:, :, W // 2 - config.offset:] - len(uni_prompting.text_tokenizer)
image_up_part = extrapolation_image_tokens[:, :-(W // 2 - config.offset), :] - len(uni_prompting.text_tokenizer)
image_down_part = extrapolation_image_tokens[:, W // 2 - config.offset:, :] - len(uni_prompting.text_tokenizer)
if direction in ['left', 'right']:
extrapolation_mask = torch.zeros((config.training.batch_size,
config.dataset.params.resolution // 16,
config.dataset.params.resolution // 16 // 2 + config.offset),
dtype=torch.int64, device=device) + mask_token_id
else:
extrapolation_mask = torch.zeros((config.training.batch_size,
config.dataset.params.resolution // 16 // 2 + config.offset,
config.dataset.params.resolution // 16),
dtype=torch.int64, device=device) + mask_token_id
if direction == 'left':
extrapolation_image_tokens = torch.cat(
[extrapolation_mask, extrapolation_image_tokens[:, :, :W // 2 - config.offset]], dim=-1)
elif direction == 'right':
extrapolation_image_tokens = torch.cat(
[extrapolation_image_tokens[:, :, -(W // 2 - config.offset):], extrapolation_mask], dim=-1)
elif direction == 'up':
extrapolation_image_tokens = torch.cat(
[extrapolation_mask, extrapolation_image_tokens[:, :W // 2 - config.offset, :]], dim=-2)
else:
extrapolation_image_tokens = torch.cat(
[extrapolation_image_tokens[:, -(W // 2 - config.offset):, :], extrapolation_mask], dim=-2)
extrapolation_image_tokens = extrapolation_image_tokens.reshape(config.training.batch_size, -1)
input_ids, _ = uni_prompting((prt, extrapolation_image_tokens), 't2i_gen')
if config.training.guidance_scale > 0:
uncond_input_ids, _ = uni_prompting(([''] * len(prt), extrapolation_image_tokens), 't2i_gen')
attention_mask = create_attention_mask_predict_next(torch.cat([input_ids, uncond_input_ids], dim=0),
pad_id=int(uni_prompting.sptids_dict['<|pad|>']),
soi_id=int(uni_prompting.sptids_dict['<|soi|>']),
eoi_id=int(uni_prompting.sptids_dict['<|eoi|>']),
rm_pad_in_image=True)
else:
attention_mask = create_attention_mask_predict_next(input_ids,
pad_id=int(uni_prompting.sptids_dict['<|pad|>']),
soi_id=int(uni_prompting.sptids_dict['<|soi|>']),
eoi_id=int(uni_prompting.sptids_dict['<|eoi|>']),
rm_pad_in_image=True)
uncond_input_ids = None
if config.get("mask_schedule", None) is not None:
schedule = config.mask_schedule.schedule
args = config.mask_schedule.get("params", {})
mask_schedule = get_mask_chedule(schedule, **args)
else:
mask_schedule = get_mask_chedule(config.training.get("mask_schedule", "cosine"))
with torch.no_grad():
gen_token_ids = model.t2i_generate(
input_ids=input_ids,
uncond_input_ids=uncond_input_ids,
attention_mask=attention_mask,
guidance_scale=config.training.guidance_scale,
temperature=config.training.get("generation_temperature", 1.0),
timesteps=config.training.generation_timesteps,
noise_schedule=mask_schedule,
noise_type=config.training.get("noise_type", "mask"),
seq_len=config.model.showo.num_vq_tokens,
uni_prompting=uni_prompting,
config=config,
)
gen_token_ids = torch.clamp(gen_token_ids, max=config.model.showo.codebook_size - 1, min=0)
gen_token_ids = gen_token_ids.reshape(config.training.batch_size,
config.dataset.params.resolution // 16,
config.dataset.params.resolution // 16)
if direction == 'left':
gen_token_ids = torch.cat([gen_token_ids, image_right_part], dim=-1)
elif direction == 'right':
gen_token_ids = torch.cat([image_left_part, gen_token_ids], dim=-1)
elif direction == 'up':
gen_token_ids = torch.cat([gen_token_ids, image_down_part], dim=-2)
else:
gen_token_ids = torch.cat([image_left_part, gen_token_ids], dim=-2)
_, h, w = gen_token_ids.shape
gen_token_ids = gen_token_ids.reshape(config.training.batch_size, -1)
images = vq_model.decode_code(gen_token_ids, shape=(h, w))
images = torch.clamp((images + 1.0) / 2.0, min=0.0, max=1.0)
images *= 255.0
images = images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
return images[0]
def multimodal_understanding(input_img, input_text, chat_history):
top_k = 1 # retain only the top_k most likely tokens, clamp others to have 0 probability
image_ori = input_img
image = image_transform(image_ori, resolution=config.dataset.params.resolution).to(device)
image = image.unsqueeze(0)
image_tokens = vq_model.get_code(image) + len(uni_prompting.text_tokenizer)
question = input_text
input_ids = uni_prompting.text_tokenizer(['USER: \n' + question + ' ASSISTANT:'])[
'input_ids']
input_ids = torch.tensor(input_ids).to(device)
input_ids = torch.cat([
(torch.ones(input_ids.shape[0], 1) * uni_prompting.sptids_dict['<|mmu|>']).to(device),
(torch.ones(input_ids.shape[0], 1) * uni_prompting.sptids_dict['<|soi|>']).to(device),
image_tokens,
(torch.ones(input_ids.shape[0], 1) * uni_prompting.sptids_dict['<|eoi|>']).to(device),
(torch.ones(input_ids.shape[0], 1) * uni_prompting.sptids_dict['<|sot|>']).to(device),
input_ids
], dim=1).long()
attention_mask = create_attention_mask_for_mmu(input_ids.to(device),
eoi_id=int(uni_prompting.sptids_dict['<|eoi|>']))
cont_toks_list = model.mmu_generate(input_ids, attention_mask=attention_mask,
max_new_tokens=100, top_k=top_k,
eot_token=uni_prompting.sptids_dict['<|eot|>'])
cont_toks_list = torch.stack(cont_toks_list).squeeze()[None]
output_text = uni_prompting.text_tokenizer.batch_decode(cont_toks_list, skip_special_tokens=True)
output_text = output_text[0].strip()
chat_history.append((input_text, output_text))
return "", chat_history
with gr.Blocks() as demo:
gr.HTML("""
<h1 class="display-2 fw-bold title">
<a style="color: #70a8dc;">S</a><a style="color: #6fb051;">h</a><a style="color: #e06766;">o</a><a style="color: #f7b26b;">w</a>-o
</h1>
<p>This is the official Gradio demo for the Show-o model, a unified model that can do multimodal understanding and generation.</p>
<strong>Paper:</strong> <a href="https://arxiv.org/abs/2408.12528" target="_blank">Show-o: One Single Transformer To Unify Multimodal Understanding and Generation </a>
<br/>
<strong>Project Website:</strong> <a href="https://showlab.github.io/Show-o/" target="_blank">Show-o Website</a>
<br/>
<strong>Code and Models:</strong> <a href="https://github.com/showlab/Show-o" target="_blank">GitHub</a>
<br/>
<br/>
""")
with gr.Row():
with gr.Column():
text_prompt_t2i = gr.Textbox(
label="Text prompt",
lines=2,
placeholder="Input the text prompt here for image generation."
)
guidance_scale_t2i = gr.Slider(
label="guidance scale",
minimum=0,
maximum=5,
step=0.05,
value=1.75
)
generation_timesteps_t2i = gr.Slider(
label="timesteps",
minimum=1,
maximum=30,
step=1,
value=18
)
generated_img_t2i = gr.Image(
label="Output image"
)
submit_btn_t2i = gr.Button("Generate: Text-to-image")
submit_btn_t2i.click(text_to_image_generation,
[text_prompt_t2i, guidance_scale_t2i, generation_timesteps_t2i],
[generated_img_t2i])
with gr.Row():
inpainting_input_img = gr.Image(
label="Input image",
type="pil",
)
inpainting_input_mask = gr.Image(
label="Inpainting mask",
image_mode="L",
type="pil",
)
with gr.Column():
text_prompt_inpainting = gr.Textbox(
label="Text prompt",
lines=2,
placeholder="Input the text prompt here for image inpainting."
)
guidance_scale_inpainting = gr.Slider(
label="guidance scale",
minimum=0,
maximum=5,
step=0.05,
value=1.75
)
generation_timesteps_inpainting = gr.Slider(
label="timesteps",
minimum=1,
maximum=30,
step=1,
value=16
)
generated_img_inpainting = gr.Image(
label="Output image"
)
submit_btn_inpainting = gr.Button("Generate: Text-guided Inpainting")
submit_btn_inpainting.click(text_guided_inpainting,
[text_prompt_inpainting, inpainting_input_img, inpainting_input_mask,
guidance_scale_inpainting, generation_timesteps_inpainting],
[generated_img_inpainting])
with gr.Row():
extra_input_img = gr.Image(
label="Input image",
type="pil",
image_mode="RGB",
)
with gr.Column():
text_prompt_extrapolation = gr.Textbox(
label="Text prompt",
lines=1,
placeholder="Input the text prompt here for image extrapolation."
)
guidance_scale_extrapolation = gr.Slider(
label="guidance scale",
minimum=0,
maximum=5,
step=0.05,
value=1.75
)
generation_timesteps_extrapolation = gr.Slider(
label="timesteps",
minimum=1,
maximum=30,
step=1,
value=16
)
left_extrapolation = gr.Slider(
label="left extrapolation",
minimum=0,
maximum=5,
step=1,
value=1
)
right_extrapolation = gr.Slider(
label="right extrapolation",
minimum=0,
maximum=5,
step=1,
value=1
)
generated_img_extrapolation = gr.Image(
label="Output image"
)
submit_btn_inpainting = gr.Button("Generate: Text-guided Extrapolation")
submit_btn_inpainting.click(text_guided_extrapolation,
[extra_input_img, text_prompt_extrapolation, left_extrapolation, right_extrapolation,
guidance_scale_extrapolation, generation_timesteps_extrapolation],
[generated_img_extrapolation])
with gr.Row():
with gr.Row():
chat_input_img = gr.Image(
label="Input image",
type="pil",
image_mode="RGB",
)
with gr.Column():
chatbot = gr.Chatbot()
msg = gr.Textbox(label="Press Enter to send a message for chat")
clear = gr.ClearButton([msg, chatbot])
msg.submit(multimodal_understanding, [chat_input_img, msg, chatbot], [msg, chatbot])
demo.launch()
|