yanziang/InternVideo3-8B-Instruct

1# coding=utf-82# Copyright 2025 The InternVideo Team. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15"""Video processor class for InternVideo3."""16 17import math18from typing import Optional, Union19 20import numpy as np21import torch22 23from transformers.feature_extraction_utils import BatchFeature24from transformers.image_utils import ChannelDimension, PILImageResampling, SizeDict, get_image_size25from transformers.processing_utils import Unpack, VideosKwargs26from transformers.utils import TensorType, logging27from transformers.video_processing_utils import BaseVideoProcessor28from transformers.video_utils import VideoMetadata, group_videos_by_shape, reorder_videos29 30 31logger = logging.get_logger(__name__)32 33 34def smart_resize(35    num_frames: int,36    height: int,37    width: int,38    temporal_factor: int = 2,39    factor: int = 32,40    min_pixels: int = 128 * 128,41    max_pixels: int = 16 * 16 * 2 * 2 * 2 * 6144,42):43    if num_frames < temporal_factor:44        raise ValueError(f"t:{num_frames} must be larger than temporal_factor:{temporal_factor}")45    if height < factor or width < factor:46        raise ValueError(f"height:{height} or width:{width} must be larger than factor:{factor}")47    elif max(height, width) / min(height, width) > 200:48        raise ValueError(49            f"absolute aspect ratio must be smaller than 200, got {max(height, width) / min(height, width)}"50        )51    h_bar = round(height / factor) * factor52    w_bar = round(width / factor) * factor53    t_bar = round(num_frames / temporal_factor) * temporal_factor54 55    if t_bar * h_bar * w_bar > max_pixels:56        beta = math.sqrt((num_frames * height * width) / max_pixels)57        h_bar = max(factor, math.floor(height / beta / factor) * factor)58        w_bar = max(factor, math.floor(width / beta / factor) * factor)59    elif t_bar * h_bar * w_bar < min_pixels:60        beta = math.sqrt(min_pixels / (num_frames * height * width))61        h_bar = math.ceil(height * beta / factor) * factor62        w_bar = math.ceil(width * beta / factor) * factor63 64    return h_bar, w_bar65 66 67class InternVideo3VideoProcessorInitKwargs(VideosKwargs):68    patch_size: Optional[int]69    temporal_patch_size: Optional[int]70    merge_size: Optional[int]71    min_frames: Optional[int]72    max_frames: Optional[int]73 74 75class InternVideo3VideoProcessor(BaseVideoProcessor):76    resample = PILImageResampling.BICUBIC77    size = {"shortest_edge": 128 * 32 * 32, "longest_edge": 32 * 32 * 768}78    image_mean = [0.5, 0.5, 0.5]79    image_std = [0.5, 0.5, 0.5]80    do_resize = True81    do_rescale = True82    do_normalize = True83    do_convert_rgb = True84    patch_size = 1685    temporal_patch_size = 286    merge_size = 287    fps = 288    min_frames = 489    max_frames = 76890    do_sample_frames = True91    valid_kwargs = InternVideo3VideoProcessorInitKwargs92    model_input_names = ["pixel_values_videos", "video_grid_thw"]93 94    def __init__(self, **kwargs: Unpack[InternVideo3VideoProcessorInitKwargs]):95        super().__init__(**kwargs)96        if self.size is not None and (97            self.size.get("shortest_edge", None) is None or self.size.get("longest_edge", None) is None98        ):99            raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.")100 101    def _further_process_kwargs(102        self,103        size: Optional[SizeDict] = None,104        **kwargs,105    ) -> dict:106        if size is not None and ("shortest_edge" not in size or "longest_edge" not in size):107            raise ValueError("size must contain 'shortest_edge' and 'longest_edge' keys.")108        return super()._further_process_kwargs(size=size, **kwargs)109 110    def sample_frames(111        self,112        metadata: VideoMetadata,113        num_frames: Optional[int] = None,114        fps: Optional[Union[int, float]] = None,115        **kwargs,116    ):117        if fps is not None and num_frames is not None:118            raise ValueError("`num_frames` and `fps` are mutually exclusive arguments, please use only one!")119 120        total_num_frames = metadata.total_num_frames121        fps = fps if fps is not None else self.fps122 123        if num_frames is None and fps is not None:124            if metadata.fps is None:125                metadata.fps = 24126                logger.warning_once(127                    "Asked to sample `fps` frames per second but no video metadata was provided. "128                    "Defaulting to `fps=24`. Please provide `video_metadata` for more accurate results."129                )130            num_frames = int(total_num_frames / metadata.fps * fps)131            num_frames = min(min(max(num_frames, self.min_frames), self.max_frames), total_num_frames)132 133        if num_frames is None:134            num_frames = min(max(total_num_frames, self.min_frames), self.max_frames)135 136        indices = np.linspace(0, total_num_frames - 1, num_frames).round().astype(int)137        return indices138 139    def _preprocess(140        self,141        videos: list[torch.Tensor],142        do_convert_rgb: bool = True,143        do_resize: bool = True,144        size: Optional[SizeDict] = None,145        interpolation: PILImageResampling = PILImageResampling.BICUBIC,146        do_rescale: bool = True,147        rescale_factor: float = 1 / 255.0,148        do_normalize: bool = True,149        image_mean: Optional[Union[float, list[float]]] = None,150        image_std: Optional[Union[float, list[float]]] = None,151        patch_size: Optional[int] = None,152        temporal_patch_size: Optional[int] = None,153        merge_size: Optional[int] = None,154        return_tensors: Optional[Union[str, TensorType]] = None,155        **kwargs,156    ):157        grouped_videos, grouped_videos_index = group_videos_by_shape(videos)158        resized_videos_grouped = {}159 160        for shape, stacked_videos in grouped_videos.items():161            B, T, C, H, W = stacked_videos.shape162            num_frames, height, width = T, H, W163            if do_resize:164                resized_height, resized_width = smart_resize(165                    num_frames=num_frames,166                    height=height,167                    width=width,168                    temporal_factor=temporal_patch_size,169                    factor=patch_size * merge_size,170                    min_pixels=size.shortest_edge,171                    max_pixels=size.longest_edge,172                )173                stacked_videos = stacked_videos.view(B * T, C, H, W)174                stacked_videos = self.resize(175                    stacked_videos,176                    size=SizeDict(height=resized_height, width=resized_width),177                    interpolation=interpolation,178                )179                stacked_videos = stacked_videos.view(B, T, C, resized_height, resized_width)180            resized_videos_grouped[shape] = stacked_videos181        resized_videos = reorder_videos(resized_videos_grouped, grouped_videos_index)182 183        grouped_videos, grouped_videos_index = group_videos_by_shape(resized_videos)184        processed_videos_grouped = {}185        processed_grids = {}186        for shape, stacked_videos in grouped_videos.items():187            resized_height, resized_width = get_image_size(stacked_videos[0], channel_dim=ChannelDimension.FIRST)188 189            stacked_videos = self.rescale_and_normalize(190                stacked_videos, do_rescale, rescale_factor, do_normalize, image_mean, image_std191            )192            patches = stacked_videos193 194            if patches.shape[1] % temporal_patch_size != 0:195                repeats = patches[:, -1:].repeat(1, temporal_patch_size - 1, 1, 1, 1)196                patches = torch.cat([patches, repeats], dim=1)197            batch_size, grid_t, channel = patches.shape[:3]198            grid_t = grid_t // temporal_patch_size199            grid_h, grid_w = resized_height // patch_size, resized_width // patch_size200 201            patches = patches.view(202                batch_size,203                grid_t,204                temporal_patch_size,205                channel,206                grid_h // merge_size,207                merge_size,208                patch_size,209                grid_w // merge_size,210                merge_size,211                patch_size,212            )213            patches = patches.permute(0, 1, 4, 7, 5, 8, 3, 2, 6, 9)214            flatten_patches = patches.reshape(215                batch_size,216                grid_t * grid_h * grid_w,217                channel * temporal_patch_size * patch_size * patch_size,218            )219 220            processed_videos_grouped[shape] = flatten_patches221            processed_grids[shape] = [[grid_t, grid_h, grid_w]] * batch_size222 223        processed_videos = reorder_videos(processed_videos_grouped, grouped_videos_index)224        processed_grids = reorder_videos(processed_grids, grouped_videos_index)225        pixel_values_videos = torch.cat(processed_videos, dim=0)226        video_grid_thw = torch.tensor(processed_grids)227        data = {228            "pixel_values_videos": pixel_values_videos,229            "video_grid_thw": video_grid_thw,230        }231 232        return BatchFeature(data=data, tensor_type=return_tensors)233 234 235__all__ = ["InternVideo3VideoProcessor"]236