| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # adapted from https://huggingface.co/h2oai/h2ovl-mississippi-2b/blob/main/modeling_h2ovl_chat.py |
| 5 | # https://huggingface.co/h2oai/h2ovl-mississippi-2b/blob/main/image_process.py |
| 6 | # -------------------------------------------------------- |
| 7 | # H2OVL-Mississippi |
| 8 | # Copyright (c) 2024 H2O.AI |
| 9 | # Licensed under Apache 2.0 License [see LICENSE for details] |
| 10 | # -------------------------------------------------------- |
| 11 | import torch |
| 12 | from PIL import Image |
| 13 | |
| 14 | from vllm.tokenizers.hf import HfTokenizer |
| 15 | |
| 16 | from .internvl import ( |
| 17 | InternVLImageProcessor, |
| 18 | InternVLProcessor, |
| 19 | build_transform, |
| 20 | find_closest_aspect_ratio, |
| 21 | get_internvl_target_ratios, |
| 22 | ) |
| 23 | |
| 24 | |
| 25 | def resolve_h2ovl_min_max_num( |
| 26 | *, |
| 27 | min_dynamic_patch: int, |
| 28 | max_dynamic_patch: int, |
| 29 | dynamic_image_size: bool, |
| 30 | use_thumbnail: bool, |
| 31 | ) -> tuple[int, int]: |
| 32 | min_dynamic_patch = min_dynamic_patch if dynamic_image_size else 1 |
| 33 | max_dynamic_patch = max_dynamic_patch if dynamic_image_size else 1 |
| 34 | |
| 35 | if use_thumbnail and max_dynamic_patch != 1: |
| 36 | max_dynamic_patch += 1 |
| 37 | |
| 38 | return min_dynamic_patch, max_dynamic_patch |
| 39 | |
| 40 | |
| 41 | def get_h2ovl_target_ratios( |
| 42 | min_num: int, |
| 43 | max_num: int, |
| 44 | *, |
| 45 | prior_aspect_ratio: tuple[int, int] | None, |
| 46 | ) -> list[tuple[int, int]]: |
| 47 | target_ratios = get_internvl_target_ratios(min_num, max_num) |
| 48 | |
| 49 | # if prior_aspect_ratio is provided, filter the target ratios |
| 50 | if prior_aspect_ratio is not None: |
| 51 | target_ratios = [ |
| 52 | ratio |
| 53 | for ratio in target_ratios |
| 54 | if prior_aspect_ratio[0] % ratio[0] != 0 |
| 55 | and prior_aspect_ratio[1] % ratio[1] != 0 |
| 56 | ] |
| 57 | |
| 58 | return target_ratios |
| 59 | |
| 60 | |
| 61 | # modified to include blocks generated in second pass |
| 62 | def calculate_h2ovl_targets( |
| 63 | *, |
| 64 | orig_width: int, |
| 65 | orig_height: int, |
| 66 | target_ratios: list[tuple[int, int]], |
| 67 | image_size: int, |
| 68 | use_thumbnail: bool, |
| 69 | ) -> tuple[int, int, int, tuple[int, int]]: |
| 70 | aspect_ratio = orig_width / orig_height |
| 71 | |
| 72 | # find the closest aspect ratio to the target |
| 73 | target_aspect_ratio = find_closest_aspect_ratio( |
| 74 | aspect_ratio, |
| 75 | target_ratios, |
| 76 | width=orig_width, |
| 77 | height=orig_height, |
| 78 | image_size=image_size, |
| 79 | ) |
| 80 | |
| 81 | # calculate the target width and height |
| 82 | target_width = image_size * target_aspect_ratio[0] |
| 83 | target_height = image_size * target_aspect_ratio[1] |
| 84 | blocks = target_aspect_ratio[0] * target_aspect_ratio[1] |
| 85 | |
| 86 | # add thumbnail image if num_blocks != 1 |
| 87 | if use_thumbnail and blocks != 1: |
| 88 | blocks += 1 |
| 89 | |
| 90 | return blocks, target_width, target_height, target_aspect_ratio |
| 91 | |
| 92 | |
| 93 | # adapted from https://huggingface.co/OpenGVLab/InternVL2-1B |
| 94 | # refactored to handle prior_aspect_ratio |
| 95 | def dynamic_preprocess_h2ovl( |
| 96 | image: Image.Image, |
| 97 | *, |
| 98 | target_ratios: list[tuple[int, int]], |
| 99 | image_size: int, |
| 100 | use_thumbnail: bool, |
| 101 | ) -> tuple[list[Image.Image], tuple[int, int]]: |
| 102 | orig_width, orig_height = image.size |
| 103 | |
| 104 | # calculate the number of blocks without thumbnail |
| 105 | ( |
| 106 | blocks, |
| 107 | target_width, |
| 108 | target_height, |
| 109 | target_aspect_ratio, |
| 110 | ) = calculate_h2ovl_targets( |
| 111 | orig_width=orig_width, |
| 112 | orig_height=orig_height, |
| 113 | target_ratios=target_ratios, |
| 114 | image_size=image_size, |
| 115 | use_thumbnail=False, |
| 116 | ) |
| 117 | |
| 118 | # resize the image |
| 119 | resized_img = image.resize((target_width, target_height)) |
| 120 | processed_images = [] |
| 121 | for i in range(blocks): |
| 122 | box = ( |
| 123 | (i % (target_width // image_size)) * image_size, |
| 124 | (i // (target_width // image_size)) * image_size, |
| 125 | ((i % (target_width // image_size)) + 1) * image_size, |
| 126 | ((i // (target_width // image_size)) + 1) * image_size, |
| 127 | ) |
| 128 | # split the image |
| 129 | split_img = resized_img.crop(box) |
| 130 | processed_images.append(split_img) |
| 131 | |
| 132 | assert len(processed_images) == blocks |
| 133 | |
| 134 | if use_thumbnail and len(processed_images) != 1: |
| 135 | thumbnail_img = image.resize((image_size, image_size)) |
| 136 | processed_images.append(thumbnail_img) |
| 137 | |
| 138 | return processed_images, target_aspect_ratio |
| 139 | |
| 140 | |
| 141 | def _preprocess_image( |
| 142 | image: Image.Image, |
| 143 | *, |
| 144 | input_size: int, |
| 145 | min_num: int, |
| 146 | max_num: int, |
| 147 | use_thumbnail: bool, |
| 148 | prior_aspect_ratio: tuple[int, int] | None, |
| 149 | ) -> tuple[torch.Tensor, tuple[int, int]]: |
| 150 | target_ratios = get_h2ovl_target_ratios( |
| 151 | min_num, |
| 152 | max_num, |
| 153 | prior_aspect_ratio=prior_aspect_ratio, |
| 154 | ) |
| 155 | |
| 156 | transform = build_transform(input_size=input_size) |
| 157 | images, target_aspect_ratio = dynamic_preprocess_h2ovl( |
| 158 | image, |
| 159 | image_size=input_size, |
| 160 | use_thumbnail=use_thumbnail, |
| 161 | target_ratios=target_ratios, |
| 162 | ) |
| 163 | |
| 164 | pixel_values = torch.stack([transform(image) for image in images]) |
| 165 | return pixel_values, target_aspect_ratio |
| 166 | |
| 167 | |
| 168 | # refactored to use the _preprocess_image function |
| 169 | def image_to_pixel_values_h2ovl( |
| 170 | image: Image.Image, |
| 171 | *, |
| 172 | input_size: int, |
| 173 | min_num: int, |
| 174 | max_num: int, |
| 175 | use_thumbnail: bool, |
| 176 | use_msac: bool, |
| 177 | ) -> torch.Tensor: |
| 178 | # when MSAC is turned on, we need to process the image twice |
| 179 | if use_msac: |
| 180 | # first pass |
| 181 | pixel_values1, aspect_ratio1 = _preprocess_image( |
| 182 | image, |
| 183 | input_size=input_size, |
| 184 | min_num=1, |
| 185 | max_num=max_num, |
| 186 | use_thumbnail=True, |
| 187 | prior_aspect_ratio=None, |
| 188 | ) |
| 189 | # second pass |
| 190 | pixel_values2, _ = _preprocess_image( |
| 191 | image, |
| 192 | input_size=input_size, |
| 193 | min_num=3, |
| 194 | max_num=max_num, |
| 195 | use_thumbnail=True, |
| 196 | prior_aspect_ratio=aspect_ratio1, |
| 197 | ) |
| 198 | # combine pixel values |
| 199 | pixel_values = torch.cat( |
| 200 | [pixel_values2[:-1], pixel_values1[:-1], pixel_values2[-1:]], 0 |
| 201 | ) |
| 202 | |
| 203 | else: |
| 204 | pixel_values, _ = _preprocess_image( |
| 205 | image, |
| 206 | input_size=input_size, |
| 207 | min_num=min_num, |
| 208 | max_num=max_num, |
| 209 | use_thumbnail=use_thumbnail, |
| 210 | prior_aspect_ratio=None, |
| 211 | ) |
| 212 | |
| 213 | return pixel_values |
| 214 | |
| 215 | |
| 216 | class H2OVLImageProcessor(InternVLImageProcessor): |
| 217 | def __init__( |
| 218 | self, |
| 219 | image_size: int, |
| 220 | min_dynamic_patch: int, |
| 221 | max_dynamic_patch: int, |
| 222 | dynamic_image_size: bool, |
| 223 | use_thumbnail: bool, |
| 224 | use_msac: bool, |
| 225 | ) -> None: |
| 226 | super().__init__( |
| 227 | image_size=image_size, |
| 228 | min_dynamic_patch=min_dynamic_patch, |
| 229 | max_dynamic_patch=max_dynamic_patch, |
| 230 | dynamic_image_size=dynamic_image_size, |
| 231 | use_thumbnail=use_thumbnail, |
| 232 | ) |
| 233 | |
| 234 | self.use_msac = use_msac |
| 235 | |
| 236 | def resolve_min_max_num( |
| 237 | self, |
| 238 | *, |
| 239 | min_dynamic_patch: int | None = None, |
| 240 | max_dynamic_patch: int | None = None, |
| 241 | dynamic_image_size: bool | None = None, |
| 242 | use_thumbnail: bool | None = None, |
| 243 | ) -> tuple[int, int]: |
| 244 | if min_dynamic_patch is None: |
| 245 | min_dynamic_patch = self.min_dynamic_patch |
| 246 | if max_dynamic_patch is None: |
| 247 | max_dynamic_patch = self.max_dynamic_patch |
| 248 | if dynamic_image_size is None: |
| 249 | dynamic_image_size = self.dynamic_image_size |
| 250 | if use_thumbnail is None: |
| 251 | use_thumbnail = self.use_thumbnail |
| 252 | |
| 253 | return resolve_h2ovl_min_max_num( |
| 254 | min_dynamic_patch=min_dynamic_patch, |
| 255 | max_dynamic_patch=max_dynamic_patch, |
| 256 | dynamic_image_size=dynamic_image_size, |
| 257 | use_thumbnail=use_thumbnail, |
| 258 | ) |
| 259 | |
| 260 | def _images_to_pixel_values_lst( |
| 261 | self, |
| 262 | images: list[Image.Image], |
| 263 | min_dynamic_patch: int | None = None, |
| 264 | max_dynamic_patch: int | None = None, |
| 265 | dynamic_image_size: bool | None = None, |
| 266 | ) -> list[torch.Tensor]: |
| 267 | use_msac = self.use_msac if len(images) == 1 else False |
| 268 | |
| 269 | min_num, max_num = self.resolve_min_max_num( |
| 270 | min_dynamic_patch=min_dynamic_patch, |
| 271 | max_dynamic_patch=max_dynamic_patch, |
| 272 | dynamic_image_size=dynamic_image_size, |
| 273 | use_thumbnail=False, # Applied in image_to_pixel_values |
| 274 | ) |
| 275 | |
| 276 | return [ |
| 277 | image_to_pixel_values_h2ovl( |
| 278 | image, |
| 279 | input_size=self.image_size, |
| 280 | min_num=min_num, |
| 281 | max_num=max_num, |
| 282 | use_thumbnail=self.use_thumbnail, |
| 283 | use_msac=use_msac, |
| 284 | ) |
| 285 | for image in images |
| 286 | ] |
| 287 | |
| 288 | |
| 289 | class H2OVLProcessor(InternVLProcessor): |
| 290 | def __init__( |
| 291 | self, |
| 292 | image_processor: H2OVLImageProcessor, |
| 293 | tokenizer: HfTokenizer, |
| 294 | *, |
| 295 | image_seq_length: int, |
| 296 | start_image_token: str = "<img>", |
| 297 | end_image_token: str = "</img>", |
| 298 | ctx_image_token: str = "<IMG_CONTEXT>", |
| 299 | ) -> None: |
| 300 | super().__init__( |
| 301 | image_processor=image_processor, |
| 302 | tokenizer=tokenizer, |
| 303 | image_seq_length=image_seq_length, |
| 304 | start_image_token=start_image_token, |
| 305 | end_image_token=end_image_token, |
| 306 | ctx_image_token=ctx_image_token, |
| 307 | ) |
| 308 | |
| 309 | self.image_processor: H2OVLImageProcessor |
| 310 | |
| 311 | def resolve_target_ratios( |
| 312 | self, |
| 313 | *, |
| 314 | min_dynamic_patch: int | None = None, |
| 315 | max_dynamic_patch: int | None = None, |
| 316 | dynamic_image_size: bool | None = None, |
| 317 | use_thumbnail: bool | None = None, |
| 318 | prior_aspect_ratio: tuple[int, int] | None = None, |
| 319 | override_min_num: int | None = None, |
| 320 | ) -> list[tuple[int, int]]: |
| 321 | min_num, max_num = self.image_processor.resolve_min_max_num( |
| 322 | min_dynamic_patch=min_dynamic_patch, |
| 323 | max_dynamic_patch=max_dynamic_patch, |
| 324 | dynamic_image_size=dynamic_image_size, |
| 325 | use_thumbnail=use_thumbnail, |
| 326 | ) |
| 327 | if override_min_num is not None: |
| 328 | min_num = override_min_num |
| 329 | |
| 330 | return get_h2ovl_target_ratios( |
| 331 | min_num, |
| 332 | max_num, |
| 333 | prior_aspect_ratio=prior_aspect_ratio, |
| 334 | ) |
| 335 | |
| 336 | def get_num_image_tokens( |
| 337 | self, |
| 338 | *, |
| 339 | image_width: int, |
| 340 | image_height: int, |
| 341 | use_msac: bool | None = None, |
| 342 | ) -> int: |
| 343 | image_processor = self.image_processor |
| 344 | use_msac = image_processor.use_msac if use_msac is None else use_msac |
| 345 | |
| 346 | use_thumbnail = image_processor.use_thumbnail |
| 347 | |
| 348 | if use_msac: |
| 349 | target_ratios_1 = self.resolve_target_ratios( |
| 350 | use_thumbnail=False, # Applied in calculate_targets |
| 351 | override_min_num=1, |
| 352 | ) |
| 353 | num_patches_1, _, _, aspect_ratio_1 = calculate_h2ovl_targets( |
| 354 | orig_width=image_width, |
| 355 | orig_height=image_height, |
| 356 | image_size=image_processor.image_size, |
| 357 | target_ratios=target_ratios_1, |
| 358 | use_thumbnail=True, |
| 359 | ) |
| 360 | |
| 361 | target_ratios_2 = self.resolve_target_ratios( |
| 362 | use_thumbnail=False, # Applied in calculate_targets |
| 363 | prior_aspect_ratio=aspect_ratio_1, |
| 364 | override_min_num=3, |
| 365 | ) |
| 366 | num_patches_2, _, _, _ = calculate_h2ovl_targets( |
| 367 | orig_width=image_width, |
| 368 | orig_height=image_height, |
| 369 | image_size=image_processor.image_size, |
| 370 | target_ratios=target_ratios_2, |
| 371 | use_thumbnail=True, |
| 372 | ) |
| 373 | |
| 374 | num_patches = num_patches_1 + num_patches_2 - 1 |
| 375 | else: |
| 376 | target_ratios = self.resolve_target_ratios( |
| 377 | use_thumbnail=False, # Applied in calculate_targets |
| 378 | ) |
| 379 | num_patches, _, _, _ = calculate_h2ovl_targets( |
| 380 | orig_width=image_width, |
| 381 | orig_height=image_height, |
| 382 | image_size=image_processor.image_size, |
| 383 | target_ratios=target_ratios, |
| 384 | use_thumbnail=use_thumbnail, |
| 385 | ) |
| 386 | |
| 387 | return num_patches * self.image_seq_length |