File size: 25,649 Bytes
a31b327 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 |
from transformers import Qwen2Config, Qwen2Model, Qwen2ForCausalLM, StoppingCriteria, TextStreamer
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from typing import List, Optional, Tuple, Union
from transformers.cache_utils import Cache
import requests
from PIL import Image
from io import BytesIO
import torch
import torch.nn as nn
from torch.nn import CrossEntropyLoss
from .got_vision_b import build_GOT_vit_b
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode
import dataclasses
DEFAULT_IMAGE_TOKEN = "<image>"
DEFAULT_IMAGE_PATCH_TOKEN = '<imgpad>'
DEFAULT_IM_START_TOKEN = '<img>'
DEFAULT_IM_END_TOKEN = '</img>'
from enum import auto, Enum
class SeparatorStyle(Enum):
"""Different separator style."""
SINGLE = auto()
TWO = auto()
MPT = auto()
@dataclasses.dataclass
class Conversation:
"""A class that keeps all conversation history."""
system: str
roles: List[str]
messages: List[List[str]]
offset: int
sep_style: SeparatorStyle = SeparatorStyle.SINGLE
sep: str = "<|im_end|>"
sep2: str = None
version: str = "Unknown"
skip_next: bool = False
def get_prompt(self):
if self.sep_style == SeparatorStyle.SINGLE:
ret = self.system + self.sep + '\n'
for role, message in self.messages:
if message:
if type(message) is tuple:
message, _, _ = message
ret += role + ": " + message + self.sep
else:
ret += role + ":"
return ret
elif self.sep_style == SeparatorStyle.TWO:
seps = [self.sep, self.sep2]
ret = self.system + seps[0]
for i, (role, message) in enumerate(self.messages):
if message:
if type(message) is tuple:
message, _, _ = message
ret += role + ": " + message + seps[i % 2]
else:
ret += role + ":"
return ret
if self.sep_style == SeparatorStyle.MPT:
if self.system:
ret = self.system + self.sep
else:
ret = ''
for role, message in self.messages:
if message:
if type(message) is tuple:
message, _, _ = message
ret += role + message + self.sep
else:
ret += role
return ret
else:
raise ValueError(f"Invalid style: {self.sep_style}")
def append_message(self, role, message):
self.messages.append([role, message])
def copy(self):
return Conversation(
system=self.system,
roles=self.roles,
messages=[[x, y] for x, y in self.messages],
offset=self.offset,
sep_style=self.sep_style,
sep=self.sep,
sep2=self.sep2)
class KeywordsStoppingCriteria(StoppingCriteria):
def __init__(self, keywords, tokenizer, input_ids):
self.keywords = keywords
self.keyword_ids = [tokenizer(keyword).input_ids for keyword in keywords]
self.keyword_ids = [keyword_id[0] for keyword_id in self.keyword_ids if type(keyword_id) is list and len(keyword_id) == 1]
self.tokenizer = tokenizer
self.start_len = None
self.input_ids = input_ids
def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
if self.start_len is None:
self.start_len = self.input_ids.shape[1]
else:
for keyword_id in self.keyword_ids:
if output_ids[0, -1] == keyword_id:
return True
outputs = self.tokenizer.batch_decode(output_ids[:, self.start_len:], skip_special_tokens=True)[0]
for keyword in self.keywords:
if keyword in outputs:
return True
return False
class GOTImageEvalProcessor:
def __init__(self, image_size=384, mean=None, std=None):
if mean is None:
mean = (0.48145466, 0.4578275, 0.40821073)
if std is None:
std = (0.26862954, 0.26130258, 0.27577711)
self.normalize = transforms.Normalize(mean, std)
self.transform = transforms.Compose(
[
transforms.Resize(
(image_size, image_size), interpolation=InterpolationMode.BICUBIC
),
transforms.ToTensor(),
self.normalize,
]
)
def __call__(self, item):
return self.transform(item)
class GOTConfig(Qwen2Config):
model_type = "GOT"
class GOTQwenModel(Qwen2Model):
config_class = GOTConfig
def __init__(self, config: Qwen2Config):
super(GOTQwenModel, self).__init__(config)
self.vision_tower_high = build_GOT_vit_b()
self.mm_projector_vary = nn.Linear(1024, 1024)
def initialize_vision_modules(
self,
vision_tower,
pretrained_stage1_model=None,
freeze_vision_tower=False,
use_im_start_end=False,
vision_select_layer=-1,
dtype=torch.float16,
device="cuda"
):
image_processor_high = GOTImageEvalProcessor(image_size=1024)
self.vision_tower_high = self.vision_tower_high.to(dtype=dtype, device=device)
self.mm_projector_vary = self.mm_projector_vary.to(dtype=dtype, device=device)
image_token_len = 256
self.config.vision_tower = vision_tower
self.config.image_token_len = image_token_len
self.config.use_im_start_end = True
self.config.vision_select_layer = vision_select_layer
self.config.freeze_vision_tower = freeze_vision_tower
return dict(
image_processor_high=image_processor_high,
image_token_len=image_token_len,
)
def forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
images: Optional[torch.FloatTensor] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPast]:
# HACK: replace back original embeddings for LLaVA pretraining
orig_embeds_params = getattr(self, 'orig_embeds_params', None)
if orig_embeds_params is not None:
with torch.no_grad():
self.get_input_embeddings().weight[:-self.num_new_tokens] = orig_embeds_params[:-self.num_new_tokens].data
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
vision_tower_high = getattr(self, 'vision_tower_high', None)
if vision_tower_high is not None and (input_ids.shape[1] != 1 or self.training) and images is not None:
use_im_start_end = getattr(self.config, "use_im_start_end", -1)
vision_select_layer = getattr(self.config, "vision_select_layer", -1)
im_patch_token = getattr(self.config, "im_patch_token", -1)
im_start_token = getattr(self.config, "im_start_token", -1)
im_end_token = getattr(self.config, "im_end_token", -1)
freeze_vision_tower = getattr(self.config, "freeze_vision_tower", False)
im_patch_token = 151859
im_start_token = 151857
im_end_token = 151858
image_features = []
for image in images:
P, C, H, W = image.shape
if P == 1:
with torch.set_grad_enabled(False):
cnn_feature = vision_tower_high(image)
cnn_feature = cnn_feature.flatten(2).permute(0, 2, 1) # 256*1024
image_feature = self.mm_projector_vary(cnn_feature)
image_features.append(image_feature)
else:
image_patches = torch.unbind(image)
image_patches_features = []
for image_patch in image_patches:
image_p = torch.stack([image_patch])
with torch.set_grad_enabled(False):
cnn_feature_p = vision_tower_high(image_p)
cnn_feature_p = cnn_feature_p.flatten(2).permute(0, 2, 1)
image_feature_p = self.mm_projector_vary(cnn_feature_p)
image_patches_features.append(image_feature_p)
image_feature = torch.cat(image_patches_features, dim=1)
image_features.append(image_feature)
dummy_image_features_2 = torch.zeros(256, 1024, device=inputs_embeds.device, dtype=inputs_embeds.dtype)
dummy_image_features = dummy_image_features_2
use_im_start_end = True
new_input_embeds = []
for cur_input_ids, cur_input_embeds, cur_image_features in zip(input_ids, inputs_embeds, image_features):
if (cur_input_ids == im_patch_token).sum() == 0:
cur_input_embeds = cur_input_embeds + (0. * dummy_image_features).sum()
new_input_embeds.append(cur_input_embeds)
continue
if use_im_start_end:
if (cur_input_ids == im_start_token).sum() != (cur_input_ids == im_end_token).sum():
raise ValueError("The number of image start tokens and image end tokens should be the same.")
image_start_tokens = torch.where(cur_input_ids == im_start_token)[0]
for image_start_token_pos, per_cur_image_features in zip(image_start_tokens, cur_image_features):
per_cur_image_features = per_cur_image_features.to(device=cur_input_embeds.device)
num_patches = per_cur_image_features.shape[0]
if cur_input_ids[image_start_token_pos + num_patches + 1] != im_end_token:
raise ValueError("The image end token should follow the image start token.")
cur_input_embeds = torch.cat(
(
cur_input_embeds[:image_start_token_pos+1],
per_cur_image_features,
cur_input_embeds[image_start_token_pos + num_patches + 1:]
),
dim=0
)
new_input_embeds.append(cur_input_embeds)
else:
raise NotImplementedError
inputs_embeds = torch.stack(new_input_embeds, dim=0)
return super(GOTQwenModel, self).forward(
input_ids=None, attention_mask=attention_mask, past_key_values=past_key_values,
inputs_embeds=inputs_embeds, use_cache=use_cache, position_ids = position_ids,
output_attentions=output_attentions, output_hidden_states=output_hidden_states,
return_dict=return_dict
)
class GOTQwenForCausalLM(Qwen2ForCausalLM):
config_class = GOTConfig
# supports_gradient_checkpointing = True
def __init__(self, config):
super(Qwen2ForCausalLM, self).__init__(config)
self.model = GOTQwenModel(config)
self.vocab_size = config.vocab_size
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
# Initialize weights and apply final processing
self.post_init()
def get_model(self):
return self.model
def forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
images: Optional[torch.FloatTensor] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, CausalLMOutputWithPast]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.model(
input_ids=input_ids,
past_key_values=past_key_values,
attention_mask=attention_mask,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
images=images,
return_dict=return_dict
)
hidden_states = outputs[0]
logits = self.lm_head(hidden_states)
logits = logits.float()
# logits
loss = None
if labels is not None:
# Shift so that tokens < n predict n
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
shift_logits = shift_logits.view(-1, self.config.vocab_size)
shift_labels = shift_labels.view(-1)
# Enable model parallelism
shift_labels = shift_labels.to(shift_logits.device)
loss = loss_fct(shift_logits, shift_labels)
if not return_dict:
output = (logits,) + outputs[1:]
return (loss,) + output if loss is not None else output
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
def prepare_inputs_for_generation(
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
):
# Omit tokens covered by past_key_values
if past_key_values is not None:
if isinstance(past_key_values, Cache):
cache_length = past_key_values.get_seq_length()
past_length = past_key_values.seen_tokens
max_cache_length = past_key_values.get_max_length()
else:
cache_length = past_length = past_key_values[0][0].shape[2]
max_cache_length = None
# Keep only the unprocessed tokens:
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
# input)
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
# input_ids based on the past_length.
elif past_length < input_ids.shape[1]:
input_ids = input_ids[:, past_length:]
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
if (
max_cache_length is not None
and attention_mask is not None
and cache_length + input_ids.shape[1] > max_cache_length
):
attention_mask = attention_mask[:, -max_cache_length:]
position_ids = kwargs.get("position_ids", None)
if attention_mask is not None and position_ids is None:
# create position_ids on the fly for batch generation
position_ids = attention_mask.long().cumsum(-1) - 1
position_ids.masked_fill_(attention_mask == 0, 1)
if past_key_values:
position_ids = position_ids[:, -input_ids.shape[1] :]
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
if inputs_embeds is not None and past_key_values is None:
model_inputs = {"inputs_embeds": inputs_embeds}
else:
model_inputs = {"input_ids": input_ids}
model_inputs.update(
{
"position_ids": position_ids,
"past_key_values": past_key_values,
"use_cache": kwargs.get("use_cache"),
"attention_mask": attention_mask,
"images": kwargs.get("images", None),
}
)
return model_inputs
def initialize_vision_tokenizer(
self,
tokenizer,
freeze_lm_model=False,
pretrained_stage1_model=None,
device="cuda"
):
config = self.get_model().config
self.resize_token_embeddings(len(tokenizer))
config.im_patch_token = 151859
config.use_im_start_end = True
if config.use_im_start_end:
self.resize_token_embeddings(len(tokenizer))
config.im_start_token, config.im_end_token = 151857, 151858
def load_image(self, image_file):
if image_file.startswith('http') or image_file.startswith('https'):
response = requests.get(image_file)
image = Image.open(BytesIO(response.content)).convert('RGB')
else:
image = Image.open(image_file).convert('RGB')
return image
def disable_torch_init(self):
"""
Disable the redundant torch default initialization to accelerate model creation.
"""
import torch
setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)
def chat(self, tokenizer, image_file, ocr_type, ocr_box='', ocr_color='', render=False):
self.disable_torch_init()
image_processor_high = GOTImageEvalProcessor(image_size=1024)
use_im_start_end = True
image_token_len = 256
image = self.load_image(image_file)
w, h = image.size
if ocr_type == 'format':
qs = 'OCR with format: '
else:
qs = 'OCR: '
if ocr_box:
bbox = eval(ocr_box)
if len(bbox) == 2:
bbox[0] = int(bbox[0]/w*1000)
bbox[1] = int(bbox[1]/h*1000)
if len(bbox) == 4:
bbox[0] = int(bbox[0]/w*1000)
bbox[1] = int(bbox[1]/h*1000)
bbox[2] = int(bbox[2]/w*1000)
bbox[3] = int(bbox[3]/h*1000)
if ocr_type == 'format':
qs = str(bbox) + ' ' + 'OCR with format: '
else:
qs = str(bbox) + ' ' + 'OCR: '
if ocr_color:
if ocr_type == 'format':
qs = '[' + ocr_color + ']' + ' ' + 'OCR with format: '
else:
qs = '[' + ocr_color + ']' + ' ' + 'OCR: '
if use_im_start_end:
qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_PATCH_TOKEN*image_token_len + DEFAULT_IM_END_TOKEN + '\n' + qs
else:
qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
conv_mpt = Conversation(
system="""<|im_start|>system
You should follow the instructions carefully and explain your answers in detail.""",
# system = None,
roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
version="mpt",
messages=(),
offset=0,
sep_style=SeparatorStyle.MPT,
sep="<|im_end|>",
)
conv = conv_mpt.copy()
conv.append_message(conv.roles[0], qs)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
print(prompt)
inputs = tokenizer([prompt])
image_tensor_1 = image_processor_high(image)
input_ids = torch.as_tensor(inputs.input_ids).cuda()
stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
keywords = [stop_str]
stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
with torch.autocast("cuda", dtype=torch.bfloat16):
output_ids = self.generate(
input_ids,
images=[image_tensor_1.unsqueeze(0).half().cuda()],
do_sample=False,
num_beams = 1,
no_repeat_ngram_size = 20,
streamer=streamer,
max_new_tokens=4096,
stopping_criteria=[stopping_criteria]
)
# if render:
# print('==============rendering===============')
# outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
# if outputs.endswith(stop_str):
# outputs = outputs[:-len(stop_str)]
# outputs = outputs.strip()
# if '**kern' in outputs:
# import verovio
# from cairosvg import svg2png
# import cv2
# import numpy as np
# tk = verovio.toolkit()
# tk.loadData(outputs)
# tk.setOptions({"pageWidth": 2100, "footer": 'none',
# 'barLineWidth': 0.5, 'beamMaxSlope': 15,
# 'staffLineWidth': 0.2, 'spacingStaff': 6})
# tk.getPageCount()
# svg = tk.renderToSVG()
# svg = svg.replace("overflow=\"inherit\"", "overflow=\"visible\"")
# svg_to_html(svg, "./results/demo.html")
# if ocr_type == 'format' and '**kern' not in outputs:
# if '\\begin{tikzpicture}' not in outputs:
# html_path = "./render_tools/" + "/content-mmd-to-html.html"
# html_path_2 = "./results/demo.html"
# right_num = outputs.count('\\right')
# left_num = outputs.count('\left')
# if right_num != left_num:
# outputs = outputs.replace('\left(', '(').replace('\\right)', ')').replace('\left[', '[').replace('\\right]', ']').replace('\left{', '{').replace('\\right}', '}').replace('\left|', '|').replace('\\right|', '|').replace('\left.', '.').replace('\\right.', '.')
# outputs = outputs.replace('"', '``').replace('$', '')
# outputs_list = outputs.split('\n')
# gt= ''
# for out in outputs_list:
# gt += '"' + out.replace('\\', '\\\\') + r'\n' + '"' + '+' + '\n'
# gt = gt[:-2]
# with open(html_path, 'r') as web_f:
# lines = web_f.read()
# lines = lines.split("const text =")
# new_web = lines[0] + 'const text =' + gt + lines[1]
# else:
# html_path = "./render_tools/" + "/tikz.html"
# html_path_2 = "./results/demo.html"
# outputs = outputs.translate(translation_table)
# outputs_list = outputs.split('\n')
# gt= ''
# for out in outputs_list:
# if out:
# if '\\begin{tikzpicture}' not in out and '\\end{tikzpicture}' not in out:
# while out[-1] == ' ':
# out = out[:-1]
# if out is None:
# break
# if out:
# if out[-1] != ';':
# gt += out[:-1] + ';\n'
# else:
# gt += out + '\n'
# else:
# gt += out + '\n'
# with open(html_path, 'r') as web_f:
# lines = web_f.read()
# lines = lines.split("const text =")
# new_web = lines[0] + gt + lines[1]
# with open(html_path_2, 'w') as web_f_new:
# web_f_new.write(new_web)
|