| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from itertools import product |
| 4 | from math import ceil |
| 5 | |
| 6 | import numpy as np |
| 7 | import torch |
| 8 | from PIL import Image |
| 9 | from torchvision import transforms |
| 10 | from torchvision.transforms.functional import InterpolationMode |
| 11 | from transformers import ( |
| 12 | BatchFeature, |
| 13 | ImageProcessingMixin, |
| 14 | ProcessorMixin, |
| 15 | TensorType, |
| 16 | ) |
| 17 | |
| 18 | from vllm.tokenizers import TokenizerLike |
| 19 | |
| 20 | MAX_IMAGE_SIZE: int = 3024 |
| 21 | |
| 22 | ImageWithPatches = tuple[Image.Image, list[Image.Image], list[bool]] |
| 23 | |
| 24 | |
| 25 | class Step3VisionProcessor: |
| 26 | def __init__(self, size, interpolation_mode="bicubic", patch_size=None): |
| 27 | mean = [0.48145466, 0.4578275, 0.40821073] |
| 28 | std = [0.26862954, 0.26130258, 0.27577711] |
| 29 | patch_size = patch_size if patch_size is not None else size |
| 30 | |
| 31 | self.transform = transforms.Compose( |
| 32 | [ |
| 33 | transforms.ToTensor(), |
| 34 | transforms.Normalize(mean, std), |
| 35 | transforms.Resize( |
| 36 | (size, size), |
| 37 | interpolation=InterpolationMode.BICUBIC |
| 38 | if interpolation_mode == "bicubic" |
| 39 | else InterpolationMode.BILINEAR, |
| 40 | antialias=True, |
| 41 | ), |
| 42 | ] |
| 43 | ) |
| 44 | |
| 45 | self.patch_transform = ( |
| 46 | transforms.Compose( |
| 47 | [ |
| 48 | transforms.ToTensor(), |
| 49 | transforms.Normalize(mean, std), |
| 50 | transforms.Resize( |
| 51 | (patch_size, patch_size), |
| 52 | interpolation=InterpolationMode.BICUBIC |
| 53 | if interpolation_mode == "bicubic" |
| 54 | else InterpolationMode.BILINEAR, |
| 55 | antialias=True, |
| 56 | ), |
| 57 | ] |
| 58 | ) |
| 59 | if patch_size is not None |
| 60 | else None |
| 61 | ) |
| 62 | |
| 63 | def __call__(self, image, is_patch=False): |
| 64 | if is_patch: |
| 65 | assert self.patch_transform is not None |
| 66 | return {"pixel_values": self.patch_transform(image).unsqueeze(0)} |
| 67 | |
| 68 | return {"pixel_values": self.transform(image).unsqueeze(0)} |
| 69 | |
| 70 | |
| 71 | class ImagePatcher: |
| 72 | def __init__(self, enable_patch: bool = True) -> None: |
| 73 | self.enable_patch = enable_patch |
| 74 | |
| 75 | def determine_window_size(self, long: int, short: int) -> int: |
| 76 | if long < 728: |
| 77 | return short if long / short > 1.5 else 0 |
| 78 | return min(short, 504) if long / short > 4 else 504 |
| 79 | |
| 80 | def slide_window( |
| 81 | self, |
| 82 | width: int, |
| 83 | height: int, |
| 84 | sizes: list[tuple[int, int]], |
| 85 | steps: list[tuple[int, int]], |
| 86 | img_rate_thr: float = 0.6, |
| 87 | ) -> tuple[list[tuple[int, int, int, int]], tuple[int, int]]: |
| 88 | assert 1 >= img_rate_thr >= 0, "The `in_rate_thr` should lie in 0~1" |
| 89 | windows = [] |
| 90 | # Sliding windows. |
| 91 | for size, step in zip(sizes, steps): |
| 92 | size_w, size_h = size |
| 93 | step_w, step_h = step |
| 94 | |
| 95 | x_num = 1 if width <= size_w else ceil((width - size_w) / step_w + 1) |
| 96 | x_start = [step_w * i for i in range(x_num)] |
| 97 | if len(x_start) > 1 and x_start[-1] + size_w > width: |
| 98 | x_start[-1] = width - size_w |
| 99 | |
| 100 | y_num = 1 if height <= size_h else ceil((height - size_h) / step_h + 1) |
| 101 | y_start = [step_h * i for i in range(y_num)] |
| 102 | if len(y_start) > 1 and y_start[-1] + size_h > height: |
| 103 | y_start[-1] = height - size_h |
| 104 | |
| 105 | start = np.array(list(product(y_start, x_start)), dtype=int) |
| 106 | start[:, [0, 1]] = start[:, [1, 0]] |
| 107 | windows.append(np.concatenate([start, start + size], axis=1)) |
| 108 | windows = np.concatenate(windows, axis=0) |
| 109 | |
| 110 | return [ |
| 111 | (int(box[0]), int(box[1]), int(box[2] - box[0]), int(box[3] - box[1])) |
| 112 | for box in windows |
| 113 | ], (x_num, y_num) |
| 114 | |
| 115 | def square_pad(self, img: Image.Image) -> Image.Image: |
| 116 | w, h = img.size |
| 117 | if w == h: |
| 118 | return img |
| 119 | size = max(w, h) |
| 120 | padded = Image.new(img.mode, (size, size), 0) |
| 121 | padded.paste(img, (0, 0)) |
| 122 | return padded |
| 123 | |
| 124 | def get_image_size_for_padding( |
| 125 | self, img_width: int, img_height: int |
| 126 | ) -> tuple[int, int]: |
| 127 | ratio = img_width / img_height |
| 128 | if min(img_height, img_width) < 32 and (ratio > 4 or ratio < 1 / 4): |
| 129 | new_size = max(img_height, img_width) |
| 130 | return new_size, new_size |
| 131 | return img_width, img_height |
| 132 | |
| 133 | def get_image_size_for_preprocess( |
| 134 | self, img_width: int, img_height: int |
| 135 | ) -> tuple[int, int]: |
| 136 | if max(img_height, img_width) > MAX_IMAGE_SIZE: |
| 137 | scale_factor = MAX_IMAGE_SIZE / max(img_height, img_width) |
| 138 | img_width = int(img_width * scale_factor) |
| 139 | img_height = int(img_height * scale_factor) |
| 140 | return img_width, img_height |
| 141 | |
| 142 | def get_image_size_for_crop( |
| 143 | self, img_width: int, img_height: int, window_size: int |
| 144 | ): |
| 145 | w_ratio = img_width / window_size |
| 146 | h_ratio = img_height / window_size |
| 147 | |
| 148 | if w_ratio < 1: |
| 149 | width_new = img_width |
| 150 | else: |
| 151 | decimal_w = w_ratio - img_width // window_size |
| 152 | w_ratio = int(w_ratio) + 1 if decimal_w > 0.2 else int(w_ratio) |
| 153 | width_new = window_size * w_ratio |
| 154 | if h_ratio < 1: |
| 155 | height_new = img_height |
| 156 | else: |
| 157 | decimal_h = h_ratio - img_height // window_size |
| 158 | h_ratio = int(h_ratio) + 1 if decimal_h > 0.2 else int(h_ratio) |
| 159 | height_new = window_size * h_ratio |
| 160 | return int(width_new), int(height_new) |
| 161 | |
| 162 | def patch_crop(self, img: Image.Image, i: int, j: int, th: int, tw: int): |
| 163 | target = img.crop((j, i, j + tw, i + th)) |
| 164 | return target |
| 165 | |
| 166 | def get_num_patches(self, img_width: int, img_height: int) -> tuple[int, int]: |
| 167 | img_width, img_height = self.get_image_size_for_padding(img_width, img_height) |
| 168 | img_width, img_height = self.get_image_size_for_preprocess( |
| 169 | img_width, img_height |
| 170 | ) |
| 171 | window_size = self.determine_window_size( |
| 172 | max(img_height, img_width), min(img_height, img_width) |
| 173 | ) |
| 174 | if window_size == 0 or not self.enable_patch: |
| 175 | return 0, 0 |
| 176 | else: |
| 177 | img_width, img_height = self.get_image_size_for_crop( |
| 178 | img_width, img_height, window_size |
| 179 | ) |
| 180 | center_list, (x_num, y_num) = self.slide_window( |
| 181 | img_width, |
| 182 | img_height, |
| 183 | [(window_size, window_size)], |
| 184 | [(window_size, window_size)], |
| 185 | ) |
| 186 | full_rows = (len(center_list) - 1) // x_num + 1 |
| 187 | if len(center_list) > 0 and len(center_list) % x_num == 0: |
| 188 | full_rows -= 1 |
| 189 | return len(center_list), full_rows |
| 190 | |
| 191 | def __call__( |
| 192 | self, img: Image.Image |
| 193 | ) -> tuple[Image.Image, list[Image.Image], list[bool]]: |
| 194 | img_width, img_height = img.size |
| 195 | new_img_width, new_img_height = self.get_image_size_for_padding( |
| 196 | img_width, img_height |
| 197 | ) |
| 198 | if new_img_width != img_width or new_img_height != img_height: |
| 199 | img = self.square_pad(img) |
| 200 | img_width, img_height = img.size |
| 201 | |
| 202 | new_img_width, new_img_height = self.get_image_size_for_preprocess( |
| 203 | img_width, img_height |
| 204 | ) |
| 205 | img = img.resize((new_img_width, new_img_height), Image.Resampling.BILINEAR) |
| 206 | window_size = self.determine_window_size( |
| 207 | max(new_img_height, new_img_width), min(new_img_height, new_img_width) |
| 208 | ) |
| 209 | |
| 210 | if window_size == 0 or not self.enable_patch: |
| 211 | return img, [], [] |
| 212 | else: |
| 213 | new_img_width, new_img_height = self.get_image_size_for_crop( |
| 214 | new_img_width, new_img_height, window_size |
| 215 | ) |
| 216 | if (new_img_width, new_img_height) != (img_width, img_height): |
| 217 | img_for_crop = img.resize( |
| 218 | (new_img_width, new_img_height), Image.Resampling.BILINEAR |
| 219 | ) |
| 220 | else: |
| 221 | img_for_crop = img |
| 222 | |
| 223 | patches = [] |
| 224 | newlines = [] |
| 225 | center_list, (x_num, y_num) = self.slide_window( |
| 226 | new_img_width, |
| 227 | new_img_height, |
| 228 | [(window_size, window_size)], |
| 229 | [(window_size, window_size)], |
| 230 | ) |
| 231 | for patch_id, center_lf_point in enumerate(center_list): |
| 232 | x, y, patch_w, patch_h = center_lf_point |
| 233 | big_patch = self.patch_crop(img_for_crop, y, x, patch_h, patch_w) |
| 234 | patches.append(big_patch) |
| 235 | if (patch_id + 1) % x_num == 0: |
| 236 | newlines.append(patch_id) |
| 237 | |
| 238 | if newlines and newlines[-1] == len(patches) - 1: |
| 239 | newlines.pop() |
| 240 | |
| 241 | return ( |
| 242 | img, |
| 243 | patches, |
| 244 | [i in newlines for i in range(len(patches))], |
| 245 | ) |
| 246 | |
| 247 | |
| 248 | class Step3VLImageProcessor(ImageProcessingMixin): |
| 249 | def __init__( |
| 250 | self, |
| 251 | image_size: int = 728, |
| 252 | patch_size: int = 504, |
| 253 | num_image_feature_size: int = 169, |
| 254 | num_patch_feature_size: int = 81, |
| 255 | enable_patch: bool = True, |
| 256 | ) -> None: |
| 257 | self.image_size = image_size |
| 258 | self.patch_size = patch_size |
| 259 | self.num_image_feature_size = num_image_feature_size |
| 260 | self.num_patch_feature_size = num_patch_feature_size |
| 261 | self.image_preprocessor = Step3VisionProcessor( |
| 262 | image_size, "bilinear", patch_size |
| 263 | ) |
| 264 | self.patcher = ImagePatcher(enable_patch=enable_patch) |
| 265 | |
| 266 | def get_num_image_tokens(self, img_width: int, img_height: int) -> int: |
| 267 | num_patches, num_newlines = self.patcher.get_num_patches(img_width, img_height) |
| 268 | |
| 269 | return ( |
| 270 | num_patches * (self.num_patch_feature_size + 2) |
| 271 | + self.num_image_feature_size |
| 272 | + 2 |
| 273 | + num_newlines |
| 274 | ) |
| 275 | |
| 276 | def _split_images(self, images: list[Image.Image]) -> list[ImageWithPatches]: |
| 277 | result = [] |
| 278 | for img in images: |
| 279 | result.append(self.patcher(img)) |
| 280 | return result |
| 281 | |
| 282 | def _convert_images_to_pixel_values( |
| 283 | self, |
| 284 | images: list[Image.Image], |
| 285 | is_patch: bool = False, |
| 286 | ) -> list[torch.Tensor]: |
| 287 | return [ |
| 288 | self.image_preprocessor(img, is_patch=is_patch)["pixel_values"] |
| 289 | for img in images |
| 290 | ] |
| 291 | |
| 292 | def __call__( |
| 293 | self, |
| 294 | images: Image.Image | list[Image.Image], |
| 295 | return_tensors: str | TensorType | None = None, |
| 296 | ) -> BatchFeature: |
| 297 | if not isinstance(images, list): |
| 298 | images = [images] |
| 299 | |
| 300 | split_images_data = self._split_images(images) |
| 301 | pixel_values_lst = [] |
| 302 | patch_pixel_values_lst = [] |
| 303 | patch_newline_mask_lst = [] |
| 304 | num_patches = [] |
| 305 | for raw_img, img_patches, patch_newline_mask in split_images_data: |
| 306 | pixel_values_lst.extend(self._convert_images_to_pixel_values([raw_img])) |
| 307 | num_patches.append(len(img_patches)) |
| 308 | patch_pixel_values_lst.extend( |
| 309 | self._convert_images_to_pixel_values(img_patches, is_patch=True) |
| 310 | ) |
| 311 | patch_newline_mask_lst.extend(patch_newline_mask) |
| 312 | |
| 313 | pixel_values = torch.cat(pixel_values_lst) |
| 314 | patch_size = self.patch_size |
| 315 | image_inputs = { |
| 316 | "pixel_values": pixel_values, |
| 317 | "num_patches": num_patches, |
| 318 | "patch_pixel_values": ( |
| 319 | torch.cat(patch_pixel_values_lst) |
| 320 | if patch_pixel_values_lst |
| 321 | else pixel_values.new_empty((0, 3, patch_size, patch_size)) |
| 322 | ), |
| 323 | "patch_newline_mask": torch.tensor( |
| 324 | patch_newline_mask_lst, dtype=torch.bool |
| 325 | ), |
| 326 | } |
| 327 | return BatchFeature(image_inputs, tensor_type=return_tensors) |
| 328 | |
| 329 | |
| 330 | class Step3VLProcessor(ProcessorMixin): |
| 331 | attributes = ["image_processor", "tokenizer"] |
| 332 | |
| 333 | def __init__( |
| 334 | self, |
| 335 | image_processor: Step3VLImageProcessor, |
| 336 | tokenizer: TokenizerLike, |
| 337 | ) -> None: |
| 338 | self.image_processor = image_processor |
| 339 | self.tokenizer = tokenizer |
| 340 | |
| 341 | self.image_start_token = image_start_token = "<im_start>" |
| 342 | self.image_end_token = image_end_token = "<im_end>" |
| 343 | self.patch_start_token = patch_start_token = "<patch_start>" |
| 344 | self.patch_end_token = patch_end_token = "<patch_end>" |
| 345 | self.patch_newline_token = patch_newline_token = "<patch_newline>" |
| 346 | self.image_start_token_id = tokenizer.convert_tokens_to_ids(image_start_token) |
| 347 | self.image_end_token_id = tokenizer.convert_tokens_to_ids(image_end_token) |
| 348 | self.patch_start_token_id = tokenizer.convert_tokens_to_ids(patch_start_token) |
| 349 | self.patch_end_token_id = tokenizer.convert_tokens_to_ids(patch_end_token) |
| 350 | self.patch_newline_token_id = tokenizer.convert_tokens_to_ids( |
| 351 | patch_newline_token |
| 352 | ) |
| 353 | |
| 354 | self.image_token = image_token = "<im_patch>" |
| 355 | self.image_feature_tokens = image_token * image_processor.num_image_feature_size |
| 356 | self.patch_feature_tokens = image_token * image_processor.num_patch_feature_size |
| 357 | |
| 358 | self.image_token_id = image_token_id = tokenizer.convert_tokens_to_ids( |
| 359 | image_token |
| 360 | ) |
| 361 | self.image_feature_token_ids = [ |
| 362 | image_token_id |
| 363 | ] * image_processor.num_image_feature_size |
| 364 | self.patch_feature_token_ids = [ |
| 365 | image_token_id |
| 366 | ] * image_processor.num_patch_feature_size |
| 367 | |
| 368 | def _get_patch_repl_text( |
| 369 | self, |
| 370 | num_patches: int, |
| 371 | patch_newline_mask: list[bool], |
| 372 | ) -> str: |
| 373 | assert len(patch_newline_mask) == num_patches |
| 374 | |
| 375 | parts = [] |
| 376 | for i in range(num_patches): |
| 377 | parts.extend( |
| 378 | [ |
| 379 | self.patch_start_token, |
| 380 | self.patch_feature_tokens, |
| 381 | self.patch_end_token, |
| 382 | ] |
| 383 | ) |
| 384 | if patch_newline_mask[i]: |
| 385 | parts.append(self.patch_newline_token) |
| 386 | |
| 387 | return "".join(parts) |
| 388 | |
| 389 | def _get_patch_repl_ids( |
| 390 | self, |
| 391 | num_patches: int, |
| 392 | patch_newline_mask: list[bool], |
| 393 | ) -> list[int]: |
| 394 | assert len(patch_newline_mask) == num_patches |
| 395 | |
| 396 | parts = [] |
| 397 | for i in range(num_patches): |
| 398 | parts.extend( |
| 399 | [ |
| 400 | self.patch_start_token_id, |
| 401 | *self.patch_feature_token_ids, |
| 402 | self.patch_end_token_id, |
| 403 | ] |
| 404 | ) |
| 405 | if patch_newline_mask[i]: |
| 406 | parts.append(self.patch_newline_token_id) |
| 407 | |
| 408 | return parts |
| 409 | |
| 410 | def _get_image_repl_text( |
| 411 | self, |
| 412 | num_images: int, |
| 413 | ) -> str: |
| 414 | parts = [ |
| 415 | self.image_start_token, |
| 416 | self.image_feature_tokens, |
| 417 | self.image_end_token, |
| 418 | ] * num_images |
| 419 | |
| 420 | return "".join(parts) |
| 421 | |
| 422 | def _get_image_repl_ids( |
| 423 | self, |
| 424 | num_images: int, |
| 425 | ) -> list[int]: |
| 426 | part = [ |
| 427 | self.image_start_token_id, |
| 428 | *self.image_feature_token_ids, |
| 429 | self.image_end_token_id, |
| 430 | ] |
| 431 | return part * num_images |
| 432 | |
| 433 | def get_image_repl_feature_text( |
| 434 | self, |
| 435 | num_images: int, |
| 436 | num_patches: int, |
| 437 | patch_new_line_idx: list[bool], |
| 438 | ) -> str: |
| 439 | patch_repl = self._get_patch_repl_text(num_patches, patch_new_line_idx) |
| 440 | image_repl = self._get_image_repl_text(num_images) |
| 441 | return patch_repl + image_repl |
| 442 | |
| 443 | def get_image_repl_feature_ids( |
| 444 | self, |
| 445 | num_images: int, |
| 446 | num_patches: int, |
| 447 | patch_new_line_idx: list[bool], |
| 448 | ) -> list[int]: |
| 449 | patch_repl = self._get_patch_repl_ids(num_patches, patch_new_line_idx) |
| 450 | image_repl = self._get_image_repl_ids(num_images) |
| 451 | return patch_repl + image_repl |
| 452 | |
| 453 | def replace_placeholder(self, text: str, placeholder: str, repls: list[str]) -> str: |
| 454 | parts = text.split(placeholder) |
| 455 | |
| 456 | if len(parts) - 1 != len(repls): |
| 457 | raise ValueError( |
| 458 | "The number of placeholders does not match the number of replacements." |
| 459 | ) |
| 460 | |
| 461 | result = [parts[0]] |
| 462 | for i, repl in enumerate(repls): |
| 463 | result.append(repl) |
| 464 | result.append(parts[i + 1]) |
| 465 | |
| 466 | return "".join(result) |
| 467 | |
| 468 | def __call__( |
| 469 | self, |
| 470 | text: str | list[str] | None = None, |
| 471 | images: Image.Image | list[Image.Image] | None = None, |
| 472 | return_tensors: str | TensorType | None = None, |
| 473 | ) -> BatchFeature: |
| 474 | if images is not None: |
| 475 | image_inputs = self.image_processor( |
| 476 | images=images, |
| 477 | return_tensors=return_tensors, |
| 478 | ) |
| 479 | num_patches = image_inputs["num_patches"] |
| 480 | patch_newline_mask = image_inputs["patch_newline_mask"] |
| 481 | else: |
| 482 | image_inputs = {} |
| 483 | num_patches = [] |
| 484 | patch_newline_mask = [] |
| 485 | |
| 486 | if text is not None: |
| 487 | if not isinstance(text, list): |
| 488 | text = [text] |
| 489 | |
| 490 | if image_inputs: |
| 491 | image_token = self.image_token |
| 492 | image_repl_str_lst = [] |
| 493 | start = 0 |
| 494 | for n_patches in num_patches: |
| 495 | image_repl_str = self.get_image_repl_feature_text( |
| 496 | 1, n_patches, patch_newline_mask[start : start + n_patches] |
| 497 | ) |
| 498 | image_repl_str_lst.append(image_repl_str) |
| 499 | |
| 500 | start += n_patches |
| 501 | |
| 502 | text = [ |
| 503 | self.replace_placeholder(t, image_token, image_repl_str_lst) |
| 504 | for t in text |
| 505 | ] |
| 506 | |
| 507 | text_inputs = self.tokenizer(text) |
| 508 | else: |
| 509 | text_inputs = {} |
| 510 | |
| 511 | return BatchFeature( |
| 512 | data={**text_inputs, **image_inputs}, |
| 513 | tensor_type=return_tensors, |
| 514 | ) |