| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| import math |
| import warnings |
| from collections.abc import Iterable |
|
|
| import torch |
| from torchvision.transforms.v2 import functional as tvF |
|
|
| from transformers.image_processing_backends import TorchvisionBackend |
| from transformers.image_processing_utils import BatchFeature |
| from transformers.image_transforms import group_images_by_shape, reorder_images |
| from transformers.image_utils import ImageInput, PILImageResampling, SizeDict |
| from transformers.processing_utils import ImagesKwargs, Unpack |
| from transformers.utils import TensorType, auto_docstring |
|
|
|
|
| class Lfm3VlImageProcessorKwargs(ImagesKwargs, total=False): |
| r""" |
| min_pixels (`int`, *optional*, defaults to `56 * 56`): |
| The min pixels of the image to resize the image. |
| max_pixels (`int`, *optional*, defaults to `28 * 28 * 1280`): |
| The max pixels of the image to resize the image. |
| patch_size (`int`, *optional*, defaults to 14): |
| The spatial patch size of the vision encoder. |
| temporal_patch_size (`int`, *optional*, defaults to 2): |
| The temporal patch size of the vision encoder. |
| merge_size (`int`, *optional*, defaults to 2): |
| The merge size of the vision encoder to llm encoder. |
| """ |
|
|
| min_pixels: int |
| max_pixels: int |
| patch_size: int |
| temporal_patch_size: int |
| merge_size: int |
|
|
|
|
| def smart_resize( |
| height: int, width: int, factor: int = 28, min_pixels: int = 56 * 56, max_pixels: int = 14 * 14 * 4 * 1280 |
| ): |
| """Rescales the image so that the following conditions are met: |
| |
| 1. Both dimensions (height and width) are divisible by 'factor'. |
| |
| 2. The total number of pixels is within the range ['min_pixels', 'max_pixels']. |
| |
| 3. The aspect ratio of the image is maintained as closely as possible. |
| |
| """ |
| if max(height, width) / min(height, width) > 200: |
| raise ValueError( |
| f"absolute aspect ratio must be smaller than 200, got {max(height, width) / min(height, width)}" |
| ) |
| h_bar = round(height / factor) * factor |
| w_bar = round(width / factor) * factor |
| if h_bar * w_bar > max_pixels: |
| beta = math.sqrt((height * width) / max_pixels) |
| h_bar = max(factor, math.floor(height / beta / factor) * factor) |
| w_bar = max(factor, math.floor(width / beta / factor) * factor) |
| elif h_bar * w_bar < min_pixels: |
| beta = math.sqrt(min_pixels / (height * width)) |
| h_bar = math.ceil(height * beta / factor) * factor |
| w_bar = math.ceil(width * beta / factor) * factor |
| return h_bar, w_bar |
|
|
|
|
| @auto_docstring |
| class Lfm3VlImageProcessor(TorchvisionBackend): |
| do_resize = True |
| resample = PILImageResampling.BICUBIC |
| size = {"shortest_edge": 65_536, "longest_edge": 16_777_216} |
| default_to_square = False |
| do_rescale = True |
| do_normalize = True |
| image_mean = [0.5, 0.5, 0.5] |
| image_std = [0.5, 0.5, 0.5] |
| do_convert_rgb = True |
| patch_size = 16 |
| temporal_patch_size = 2 |
| merge_size = 2 |
| valid_kwargs = Lfm3VlImageProcessorKwargs |
| model_input_names = ["pixel_values", "image_grid_thw"] |
|
|
| def __init__(self, **kwargs: Unpack[Lfm3VlImageProcessorKwargs]): |
| |
| size = kwargs.pop("size", None) |
| size = dict(self.size) if size is None else size |
| if (min_pixels := kwargs.pop("min_pixels", None)) is not None: |
| size["shortest_edge"] = min_pixels |
| size.pop("min_pixels", None) |
| if (max_pixels := kwargs.pop("max_pixels", None)) is not None: |
| size["longest_edge"] = max_pixels |
| size.pop("max_pixels", None) |
| super().__init__(size=size, **kwargs) |
|
|
| def _standardize_kwargs( |
| self, |
| size: int | Iterable[int] | dict[str, int] | SizeDict | None = None, |
| min_pixels: int | None = None, |
| max_pixels: int | None = None, |
| **kwargs, |
| ) -> dict: |
| if min_pixels is not None or max_pixels is not None: |
| warnings.warn( |
| "Passing `min_pixels` and `max_pixels` to a processor call is deprecated and will be removed in v5.23. " |
| "Pass in `size={'longest_edge': xxx, 'shortest_edge': xxx} to override the target size.`", |
| FutureWarning, |
| ) |
|
|
| size_dict = dict(size) if isinstance(size, (dict, SizeDict)) else {} |
| if min_pixels is not None: |
| size_dict["shortest_edge"] = min_pixels |
| if max_pixels is not None: |
| size_dict["longest_edge"] = max_pixels |
| size = SizeDict(**size_dict) |
| return super()._standardize_kwargs(size=size, **kwargs) |
|
|
| @auto_docstring |
| def preprocess( |
| self, |
| images: ImageInput, |
| **kwargs: Unpack[Lfm3VlImageProcessorKwargs], |
| ) -> BatchFeature: |
| return super().preprocess(images, **kwargs) |
|
|
| def resize( |
| self, |
| images: "torch.Tensor", |
| size: SizeDict, |
| resample: "PILImageResampling | tvF.InterpolationMode | int | None", |
| factor: int, |
| **kwargs, |
| ) -> "torch.Tensor": |
| """Resize dynamically based on input image aspect ratio.""" |
| if not size.shortest_edge or not size.longest_edge: |
| raise ValueError(f"`size` dict must contain 'shortest_edge' and 'longest_edge' keys but got {size}.") |
|
|
| height, width = images.shape[-2:] |
| resized_height, resized_width = smart_resize( |
| height, |
| width, |
| factor=factor, |
| min_pixels=size.shortest_edge, |
| max_pixels=size.longest_edge, |
| ) |
| return super().resize( |
| image=images, |
| size=SizeDict(height=resized_height, width=resized_width), |
| resample=resample, |
| ) |
|
|
| def patchify( |
| self, |
| images: "torch.Tensor", |
| patch_size: int, |
| merge_size: int, |
| temporal_patch_size: int, |
| ) -> tuple["torch.Tensor", int, int]: |
| """Patchifies each image into flat layout of shape (`seq_len`, `patch_dim`) so we can concat dynamically shaped pixels.""" |
| batch_size, channel, resized_height, resized_width = images.shape |
| grid_h, grid_w = resized_height // patch_size, resized_width // patch_size |
| patches = images.reshape( |
| batch_size, |
| channel, |
| grid_h // merge_size, |
| merge_size, |
| patch_size, |
| grid_w // merge_size, |
| merge_size, |
| patch_size, |
| ) |
| |
| |
| patches = patches.permute(0, 2, 5, 3, 6, 1, 4, 7) |
|
|
| flatten_patches = ( |
| patches.unsqueeze(6) |
| .expand(-1, -1, -1, -1, -1, -1, temporal_patch_size, -1, -1) |
| .reshape( |
| batch_size, |
| grid_h * grid_w, |
| channel * temporal_patch_size * patch_size * patch_size, |
| ) |
| ) |
| return flatten_patches, grid_h, grid_w |
|
|
| def _preprocess( |
| self, |
| images: list["torch.Tensor"], |
| do_resize: bool, |
| size: SizeDict, |
| resample: "PILImageResampling | tvF.InterpolationMode | int | None", |
| do_rescale: bool, |
| rescale_factor: float, |
| do_normalize: bool, |
| image_mean: float | list[float] | None, |
| image_std: float | list[float] | None, |
| patch_size: int, |
| temporal_patch_size: int, |
| merge_size: int, |
| disable_grouping: bool | None, |
| return_tensors: str | TensorType | None, |
| **kwargs, |
| ) -> BatchFeature: |
| grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping) |
| resized_images_grouped = {} |
| for shape, stacked_images in grouped_images.items(): |
| if do_resize: |
| stacked_images = self.resize( |
| images=stacked_images, |
| size=size, |
| resample=resample, |
| factor=patch_size * merge_size, |
| ) |
| resized_images_grouped[shape] = stacked_images |
| resized_images = reorder_images(resized_images_grouped, grouped_images_index) |
|
|
| grouped_images, grouped_images_index = group_images_by_shape(resized_images, disable_grouping=disable_grouping) |
| processed_images_grouped = {} |
| processed_grids = {} |
| for shape, stacked_images in grouped_images.items(): |
| stacked_images = self.rescale_and_normalize( |
| stacked_images, do_rescale, rescale_factor, do_normalize, image_mean, image_std |
| ) |
| patches, grid_h, grid_w = self.patchify( |
| stacked_images, |
| patch_size=patch_size, |
| merge_size=merge_size, |
| temporal_patch_size=temporal_patch_size, |
| ) |
|
|
| processed_images_grouped[shape] = patches |
| processed_grids[shape] = [[1, grid_h, grid_w]] * len(stacked_images) |
|
|
| processed_images = reorder_images(processed_images_grouped, grouped_images_index) |
| processed_grids_ordered = reorder_images(processed_grids, grouped_images_index) |
| pixel_values = processed_images[0] if len(processed_images) == 1 else torch.cat(processed_images, dim=0) |
| image_grid_thw = torch.tensor(processed_grids_ordered, dtype=torch.long) |
|
|
| return BatchFeature( |
| data={"pixel_values": pixel_values, "image_grid_thw": image_grid_thw}, tensor_type=return_tensors |
| ) |
|
|
| def get_number_of_image_patches(self, height: int, width: int, images_kwargs: dict | None = None) -> int: |
| """ |
| A utility that returns number of image patches for a given image size. |
| |
| Note: Do not remove this method! It is used by vLLM to infer the number of patches and placeholders |
| without an image input. |
| |
| Args: |
| height (`int`): |
| Height of the input image. |
| width (`int`): |
| Width of the input image. |
| images_kwargs (`dict`, *optional*) |
| Any kwargs to override defaults of the image processor. |
| Returns: |
| `int`: Number of image patches per image. |
| """ |
| |
| |
| images_kwargs = dict(images_kwargs or {}) |
| size = images_kwargs.pop("size", None) or self.size |
| images_kwargs.setdefault("min_pixels", size["shortest_edge"]) |
| images_kwargs.setdefault("max_pixels", size["longest_edge"]) |
| images_kwargs = images_kwargs or {} |
| min_pixels = images_kwargs["min_pixels"] if "min_pixels" in images_kwargs else self.size["shortest_edge"] |
| max_pixels = images_kwargs["max_pixels"] if "max_pixels" in images_kwargs else self.size["longest_edge"] |
| patch_size = images_kwargs.get("patch_size", self.patch_size) |
| merge_size = images_kwargs.get("merge_size", self.merge_size) |
|
|
| factor = patch_size * merge_size |
| resized_height, resized_width = smart_resize( |
| height, width, factor, min_pixels=min_pixels, max_pixels=max_pixels |
| ) |
| grid_h, grid_w = resized_height // patch_size, resized_width // patch_size |
| return grid_h * grid_w |
|
|
|
|
| __all__ = ["Lfm3VlImageProcessor"] |
|
|
