# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # This file was automatically generated from src/transformers/models/llava_onevision/modular_llava_onevision.py. # Do NOT edit this file manually as any edits will be overwritten by the generation of # the file from the modular. If any change should be done, please apply the change to the # modular_llava_onevision.py file directly. One of our CI enforces this. # 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 # coding=utf-8 # Copyright 2024 the HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from typing import List, Optional, Union import torch from ...image_processing_utils import BatchFeature, get_patch_output_size, select_best_resolution from ...image_processing_utils_fast import ( BaseImageProcessorFast, DefaultFastImageProcessorKwargs, divide_to_patches, group_images_by_shape, reorder_images, ) from ...image_utils import ( OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, ChannelDimension, ImageInput, PILImageResampling, SizeDict, get_image_size, make_flat_list_of_images, ) from ...processing_utils import Unpack from ...utils import TensorType, auto_docstring, is_torchvision_v2_available if is_torchvision_v2_available(): from torchvision.transforms.v2 import functional as F else: from torchvision.transforms import functional as F class LlavaOnevisionFastImageProcessorKwargs(DefaultFastImageProcessorKwargs): """ image_grid_pinpoints (`List[List[int]]`, *optional*): A list of possible resolutions to use for processing high resolution images. The best resolution is selected based on the original size of the image. Can be overridden by `image_grid_pinpoints` in the `preprocess` method. do_pad (`bool`, *optional*): Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest number of patches in the batch. Padding will be applied to the bottom and right with zeros. """ image_grid_pinpoints: Optional[List[List[int]]] do_pad: Optional[bool] @auto_docstring class LlavaOnevisionImageProcessorFast(BaseImageProcessorFast): resample = PILImageResampling.BICUBIC image_mean = OPENAI_CLIP_MEAN image_std = OPENAI_CLIP_STD size = {"height": 384, "width": 384} default_to_square = False crop_size = None do_resize = True do_center_crop = None do_rescale = True do_normalize = True do_convert_rgb = True do_pad = True image_grid_pinpoints = [[384, 384], [384, 768], [384, 1152], [384, 1536], [384, 1920], [384, 2304], [768, 384], [768, 768], [768, 1152], [768, 1536], [768, 1920], [768, 2304], [1152, 384], [1152, 768], [1152, 1152], [1152, 1536], [1152, 1920], [1152, 2304], [1536, 384], [1536, 768], [1536, 1152], [1536, 1536], [1536, 1920], [1536, 2304], [1920, 384], [1920, 768], [1920, 1152], [1920, 1536], [1920, 1920], [1920, 2304], [2304, 384], [2304, 768], [2304, 1152], [2304, 1536], [2304, 1920], [2304, 2304]] # fmt: skip valid_kwargs = LlavaOnevisionFastImageProcessorKwargs model_input_names = ["pixel_values_videos"] def __init__(self, **kwargs: Unpack[LlavaOnevisionFastImageProcessorKwargs]): super().__init__(**kwargs) @auto_docstring def preprocess(self, images: ImageInput, **kwargs: Unpack[LlavaOnevisionFastImageProcessorKwargs]) -> BatchFeature: return super().preprocess(images, **kwargs) def _prepare_images_structure( self, images: ImageInput, ) -> ImageInput: """ Prepare the images structure for processing. Args: images (`ImageInput`): The input images to process. Returns: `ImageInput`: The images with a valid nesting. """ return make_flat_list_of_images(images) def _resize_for_patching( self, image: "torch.Tensor", target_resolution: tuple, interpolation: "F.InterpolationMode", input_data_format: ChannelDimension, ) -> "torch.Tensor": """ Resizes an image to a target resolution while maintaining aspect ratio. Args: image ("torch.Tensor"): The input image. target_resolution (tuple): The target resolution (height, width) of the image. interpolation (`InterpolationMode`): Resampling filter to use if resizing the image. input_data_format (`ChannelDimension` or `str`): The channel dimension format of the input image. Returns: "torch.Tensor": The resized and padded image. """ new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format) # Resize the image resized_image = F.resize(image, (new_height, new_width), interpolation=interpolation) return resized_image def _pad_for_patching( self, image: "torch.Tensor", target_resolution: tuple, input_data_format: ChannelDimension ) -> "torch.Tensor": """ Pad an image to a target resolution while maintaining aspect ratio. """ target_height, target_width = target_resolution new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format) paste_x, r_x = divmod(target_width - new_width, 2) paste_y, r_y = divmod(target_height - new_height, 2) padded_image = F.pad(image, padding=[paste_x, paste_y, paste_x + r_x, paste_y + r_y]) return padded_image def _get_image_patches( self, image: "torch.Tensor", grid_pinpoints, size: tuple, patch_size: int, interpolation: "F.InterpolationMode", ) -> List["torch.Tensor"]: """ Process an image with variable resolutions by dividing it into patches. Args: image ("torch.Tensor"): The input image to be processed. grid_pinpoints (List): A string representation of a list of possible resolutions. size (`tuple`): Size to resize the original image to. patch_size (`int`): Size of the patches to divide the image into. interpolation (`"InterpolationMode"`): Resampling filter to use if resizing the image. Returns: List["torch.Tensor"]: A list of NumPy arrays containing the processed image patches. """ if not isinstance(grid_pinpoints, list): raise TypeError("grid_pinpoints must be a list of possible resolutions.") possible_resolutions = grid_pinpoints image_size = get_image_size(image, channel_dim=ChannelDimension.FIRST) best_resolution = select_best_resolution(image_size, possible_resolutions) resized_image = self._resize_for_patching( image, best_resolution, interpolation=interpolation, input_data_format=ChannelDimension.FIRST ) padded_image = self._pad_for_patching(resized_image, best_resolution, input_data_format=ChannelDimension.FIRST) patches = divide_to_patches(padded_image, patch_size=patch_size) resized_original_image = F.resize(image, size=size, interpolation=interpolation) image_patches = [resized_original_image] + patches return image_patches def _pad_for_batching( self, pixel_values: List["torch.Tensor"], ) -> List["torch.Tensor"]: """ Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches. Args: pixel_values (`List[torch.Tensor]`): An array of pixel values of each images of shape (`batch_size`, `num_patches`, `image_in_3D`) Returns: List[`torch.Tensor`]: The padded images. """ max_patch = max(len(x) for x in pixel_values) pixel_values = [ torch.nn.functional.pad(image, pad=[0, 0, 0, 0, 0, 0, 0, max_patch - image.shape[0]]) for image in pixel_values ] return pixel_values def _preprocess( self, images: List["torch.Tensor"], do_resize: bool, size: SizeDict, image_grid_pinpoints: List[List[int]], interpolation: Optional["F.InterpolationMode"], do_center_crop: bool, crop_size: SizeDict, do_rescale: bool, rescale_factor: float, do_normalize: bool, image_mean: Optional[Union[float, List[float]]], image_std: Optional[Union[float, List[float]]], do_pad: bool, return_tensors: Optional[Union[str, TensorType]], ) -> BatchFeature: processed_images = [] image_sizes = [] # Determine the size tuple if size and size.height and size.width: size_tuple = (size.height, size.width) else: size_tuple = (size.shortest_edge, size.shortest_edge) # Determine the patch size if crop_size and crop_size.height: patch_size = crop_size.height elif size and size.height: patch_size = size.height else: patch_size = size.shortest_edge for image in images: image_patches = self._get_image_patches( image, image_grid_pinpoints, size=size_tuple, patch_size=patch_size, interpolation=interpolation, ) # Group images by size for batched processing processed_image_patches_grouped = {} grouped_image_patches, grouped_image_patches_index = group_images_by_shape(image_patches) for shape, stacked_image_patches in grouped_image_patches.items(): if do_resize: stacked_image_patches = self.resize( image=stacked_image_patches, size=size, interpolation=interpolation, ) if do_center_crop: stacked_image_patches = self.center_crop(stacked_image_patches, crop_size) # Fused rescale and normalize stacked_image_patches = self.rescale_and_normalize( stacked_image_patches, do_rescale, rescale_factor, do_normalize, image_mean, image_std ) processed_image_patches_grouped[shape] = stacked_image_patches processed_image_patches = reorder_images(processed_image_patches_grouped, grouped_image_patches_index) processed_image_patches = ( torch.stack(processed_image_patches, dim=0) if return_tensors else processed_image_patches ) processed_images.append(processed_image_patches) image_sizes.append(get_image_size(image, ChannelDimension.FIRST)) if do_pad: processed_images = self._pad_for_batching(processed_images) processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images return BatchFeature( data={"pixel_values": processed_images, "image_sizes": image_sizes}, tensor_type=return_tensors ) __all__ = ["LlavaOnevisionImageProcessorFast"]