# coding=utf-8 # Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved. # # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX # and OPT implementations in this library. It has been modified from its # original forms to accommodate minor architectural differences compared # to GPT-NeoX and OPT used by the Meta AI team that trained the model. # # 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. """Fast Image processor class for Qwen2-VL.""" from typing import Dict, List, Optional, Union from ...image_processing_utils import BatchFeature from ...image_processing_utils_fast import ( BaseImageProcessorFast, DefaultFastImageProcessorKwargs, 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, valid_images, ) from ...processing_utils import Unpack from ...utils import ( TensorType, auto_docstring, is_torch_available, is_torchvision_available, is_torchvision_v2_available, logging, ) from ...video_utils import VideoInput, make_batched_videos from .image_processing_qwen2_vl import smart_resize if is_torch_available(): import torch if is_torchvision_available(): from ...image_utils import pil_torch_interpolation_mapping if is_torchvision_v2_available(): from torchvision.transforms.v2 import functional as F else: from torchvision.transforms import functional as F logger = logging.get_logger(__name__) class Qwen2VLFastImageProcessorKwargs(DefaultFastImageProcessorKwargs): """ 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: Optional[int] max_pixels: Optional[int] patch_size: Optional[int] temporal_patch_size: Optional[int] merge_size: Optional[int] @auto_docstring class Qwen2VLImageProcessorFast(BaseImageProcessorFast): do_resize = True resample = PILImageResampling.BICUBIC size = {"shortest_edge": 56 * 56, "longest_edge": 28 * 28 * 1280} do_rescale = True do_normalize = True image_mean = OPENAI_CLIP_MEAN image_std = OPENAI_CLIP_STD do_convert_rgb = True patch_size = 14 temporal_patch_size = 2 merge_size = 2 min_pixels = None max_pixels = None valid_kwargs = Qwen2VLFastImageProcessorKwargs model_input_names = ["pixel_values", "image_grid_thw", "pixel_values_videos", "video_grid_thw"] def __init__(self, **kwargs: Unpack[Qwen2VLFastImageProcessorKwargs]): size = kwargs.pop("size", None) min_pixels = kwargs.pop("min_pixels", None) max_pixels = kwargs.pop("max_pixels", None) if size is not None and ("shortest_edge" not in size or "longest_edge" not in size): raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.") else: size = self.size # backward compatibility: override size with min_pixels and max_pixels if they are provided if min_pixels is not None: size["shortest_edge"] = min_pixels if max_pixels is not None: size["longest_edge"] = max_pixels super().__init__(size=size, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs) def _preprocess( self, images: List["torch.Tensor"], do_resize: bool, size: SizeDict, interpolation: Optional["F.InterpolationMode"], do_rescale: bool, rescale_factor: float, do_normalize: bool, image_mean: Optional[Union[float, List[float]]], image_std: Optional[Union[float, List[float]]], patch_size: int, temporal_patch_size: int, merge_size: int, do_convert_rgb: bool, input_data_format: Optional[Union[str, ChannelDimension]], device: Optional[Union[str, torch.device]], ): """ Preprocess an image or batch of images. Copy of the `preprocess` method from `CLIPImageProcessor`. Args: images (`ImageInput`): Image or batch of images to preprocess. Expects pixel values ranging from 0 to 255. If pixel values range from 0 to 1, set `do_rescale=False`. vision_info (`List[Dict]`, *optional*): Optional list of dictionaries containing additional information about vision inputs. do_resize (`bool`, *optional*, defaults to `self.do_resize`): Whether to resize the image. size (`Dict[str, int]`, *optional*, defaults to `self.size`): Size of the image after resizing. `shortest_edge` and `longest_edge` keys must be present. interpolation (`InterpolationMode`): Resampling filter to use if resizing the image. do_rescale (`bool`, *optional*, defaults to `self.do_rescale`): Whether to rescale the image. rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`): Scale factor to use if rescaling the image. do_normalize (`bool`, *optional*, defaults to `self.do_normalize`): Whether to normalize the image. image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`): Mean to use if normalizing the image. Can be a float or a list of floats corresponding to the number of channels in the image. image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`): Standard deviation to use if normalizing the image. Can be a float or a list of floats corresponding to the number of channels in the image. patch_size (`int`, *optional*, defaults to `self.patch_size`): The spatial patch size of the vision encoder. temporal_patch_size (`int`, *optional*, defaults to `self.temporal_patch_size`): The temporal patch size of the vision encoder. merge_size (`int`, *optional*, defaults to `self.merge_size`): The merge size of the vision encoder to llm encoder. do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`): Whether to convert the image to RGB. input_data_format (`ChannelDimension` or `str`, *optional*): The channel dimension format for the input image. Can be one of: - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format. - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format. - `"none"` or `ChannelDimension.NONE`: image in (height, width) format. - `"none"` or `ChannelDimension.NONE`: image in (height, width) format. device (`torch.device`, *optional*): The device to process the images on. If unset, the device is inferred from the input images. """ images = self._prepare_input_images( images=images, do_convert_rgb=do_convert_rgb, input_data_format=input_data_format, device=device, ) height, width = get_image_size(images[0], channel_dim=ChannelDimension.FIRST) resized_height, resized_width = height, width # Group images by size for batched resizing grouped_images, grouped_images_index = group_images_by_shape(images) resized_images_grouped = {} for shape, stacked_images in grouped_images.items(): if do_resize: resized_height, resized_width = smart_resize( height, width, factor=patch_size * merge_size, min_pixels=size["shortest_edge"], max_pixels=size["longest_edge"], ) stacked_images = F.resize( stacked_images, size=(resized_height, resized_width), interpolation=interpolation ) resized_images_grouped[shape] = stacked_images resized_images = reorder_images(resized_images_grouped, grouped_images_index) # Group images by size for further processing # Needed in case do_resize is False, or resize returns images with different sizes grouped_images, grouped_images_index = group_images_by_shape(resized_images) processed_images_grouped = {} for shape, stacked_images in grouped_images.items(): # Fused rescale and normalize stacked_images = self.rescale_and_normalize( stacked_images, do_rescale, rescale_factor, do_normalize, image_mean, image_std ) processed_images_grouped[shape] = stacked_images processed_images = reorder_images(processed_images_grouped, grouped_images_index) patches = torch.stack(processed_images, dim=0) if patches.shape[0] % temporal_patch_size != 0: repeats = patches[-1].unsqueeze(0).repeat(temporal_patch_size - 1, 1, 1, 1) patches = torch.cat([patches, repeats], dim=0) channel = patches.shape[1] grid_t = patches.shape[0] // temporal_patch_size grid_h, grid_w = resized_height // patch_size, resized_width // patch_size patches = patches.view( grid_t, temporal_patch_size, channel, grid_h // merge_size, merge_size, patch_size, grid_w // merge_size, merge_size, patch_size, ) patches = patches.permute(0, 3, 6, 4, 7, 2, 1, 5, 8) flatten_patches = patches.reshape( grid_t * grid_h * grid_w, channel * temporal_patch_size * patch_size * patch_size ) return flatten_patches, (grid_t, grid_h, grid_w) @auto_docstring def preprocess( self, images: ImageInput, videos: VideoInput = None, do_resize: Optional[bool] = None, size: Optional[Dict[str, int]] = None, resample: Optional[Union["PILImageResampling", "F.InterpolationMode"]] = None, do_rescale: Optional[bool] = None, rescale_factor: Optional[float] = None, do_normalize: Optional[bool] = None, image_mean: Optional[Union[float, List[float]]] = None, image_std: Optional[Union[float, List[float]]] = None, min_pixels: Optional[int] = None, max_pixels: Optional[int] = None, patch_size: Optional[int] = None, temporal_patch_size: Optional[int] = None, merge_size: Optional[int] = None, do_convert_rgb: Optional[bool] = None, return_tensors: Optional[Union[str, TensorType]] = None, data_format: Optional[ChannelDimension] = ChannelDimension.FIRST, input_data_format: Optional[Union[str, ChannelDimension]] = None, device: Optional["torch.device"] = None, **kwargs, ): 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 = min_pixels if min_pixels is not None else self.min_pixels max_pixels = max_pixels if max_pixels is not None else self.max_pixels if size is not None: if "shortest_edge" not in size or "longest_edge" not in size: raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.") min_pixels = size["shortest_edge"] elif min_pixels is not None and max_pixels is not None: # backward compatibility: override size with min_pixels and max_pixels if they are provided size = {"shortest_edge": min_pixels, "longest_edge": max_pixels} else: size = {**self.size} do_resize = do_resize if do_resize is not None else self.do_resize size = size if size is not None else self.size resample = resample if resample is not None else self.resample do_rescale = do_rescale if do_rescale is not None else self.do_rescale rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor do_normalize = do_normalize if do_normalize is not None else self.do_normalize image_mean = image_mean if image_mean is not None else self.image_mean image_std = image_std if image_std is not None else self.image_std patch_size = patch_size if patch_size is not None else self.patch_size temporal_patch_size = temporal_patch_size if temporal_patch_size is not None else self.temporal_patch_size merge_size = merge_size if merge_size is not None else self.merge_size do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb # Make hashable for cache size = SizeDict(**size) if size is not None else None image_mean = tuple(image_mean) if image_mean is not None else None image_std = tuple(image_std) if image_std is not None else None self._validate_preprocess_kwargs( do_rescale=do_rescale, rescale_factor=rescale_factor, do_normalize=do_normalize, image_mean=image_mean, image_std=image_std, do_resize=do_resize, size=size, resample=resample, return_tensors=return_tensors, data_format=data_format, ) interpolation = ( pil_torch_interpolation_mapping[resample] if isinstance(resample, (PILImageResampling, int)) else resample ) if images is not None: images = make_flat_list_of_images(images) if images is not None and not valid_images(images): raise ValueError( "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, " "torch.Tensor, tf.Tensor or jax.ndarray." ) data = {} if images is not None: pixel_values, vision_grid_thws = [], [] for image in images: patches, image_grid_thw = self._preprocess( image, do_resize=do_resize, size=size, interpolation=interpolation, do_rescale=do_rescale, rescale_factor=rescale_factor, do_normalize=do_normalize, image_mean=image_mean, image_std=image_std, patch_size=patch_size, temporal_patch_size=temporal_patch_size, merge_size=merge_size, do_convert_rgb=do_convert_rgb, input_data_format=input_data_format, device=device, ) pixel_values.extend(patches) vision_grid_thws.append(image_grid_thw) pixel_values = torch.stack(pixel_values) vision_grid_thws = torch.tensor(vision_grid_thws) data.update({"pixel_values": pixel_values, "image_grid_thw": vision_grid_thws}) # kept for BC only and should be removed after v5.0 if videos is not None: logger.warning( "`Qwen2VLImageProcessorFast` works only with image inputs and doesn't process videos anymore. " "This is a deprecated behavior and will be removed in v5.0. " "Your videos should be forwarded to `Qwen2VLVideoProcessor`. " ) videos = make_batched_videos(videos) pixel_values_videos, vision_grid_thws_videos = [], [] for images in videos: patches, video_grid_thw = self._preprocess( images, do_resize=do_resize, size=size, interpolation=interpolation, do_rescale=do_rescale, rescale_factor=rescale_factor, do_normalize=do_normalize, image_mean=image_mean, image_std=image_std, patch_size=patch_size, temporal_patch_size=temporal_patch_size, merge_size=merge_size, do_convert_rgb=do_convert_rgb, input_data_format=input_data_format, device=device, ) pixel_values_videos.extend(patches) vision_grid_thws_videos.append(video_grid_thw) pixel_values_videos = torch.stack(pixel_values_videos) vision_grid_thws_videos = torch.tensor(vision_grid_thws_videos) data.update({"pixel_values_videos": pixel_values_videos, "video_grid_thw": vision_grid_thws_videos}) return BatchFeature(data=data, tensor_type=return_tensors) __all__ = ["Qwen2VLImageProcessorFast"]