语言:

• 英文 库名称: transformers 许可证: apache-2.0 评估指标:
• 准确率 标签:
• 多模态 管道标签: 视频文本到文本 模型索引:
• 名称: InternVL2.5_HiCo_R16 结果:
  • 任务: 类型: 多模态 数据集: 名称: MLVU 类型: mlvu 指标:
    • 类型: 准确率 值: 71.5 名称: 准确率 已验证: 是
  • 任务: 类型: 多模态 数据集: 名称: MVBench 类型: mvbench 指标:
    • 类型: 准确率 值: 74.0 名称: 准确率 已验证: 是
  • 任务: 类型: 多模态 数据集: 名称: Perception Test 类型: percepTest 指标:
    • 类型: 准确率 值: 71.4 名称: 准确率 已验证: 是
  • 任务: 类型: 多模态 数据集: 名称: LongVideoBench 类型: longvideobench 指标:
    • 类型: 准确率 值: 59.6 名称: 准确率 已验证: 是
  • 任务: 类型: 多模态 数据集: 名称: VideoMME (无字幕) 类型: videomme 指标:
    • 类型: 准确率 值: 64.9 名称: 准确率 已验证: 是

📕InternVL2.5_HiCo_R16⚡

[📂 GitHub]
[📜 技术报告]

InternVideo2.5 是一个基于 InternVL2.5 构建的视频多模态大语言模型（MLLM），通过长且丰富的上下文（LRC）建模进行了增强。它显著提升了现有 MLLM 的能力，使其能够感知细粒度细节并捕捉长时态结构。我们通过使用直接偏好优化（TPO）进行密集视觉任务标注，以及通过自适应分层令牌压缩（HiCo）实现紧凑的时空表示来实现这一目标。该模型是 InternVideo2.5 消融实验的一个变体，仅基于 HiCo 技术构建（R16 表示每帧 16 个令牌）。

📈 性能

🚀 如何使用该模型

首先，您需要安装 flash attention2 和一些其他模块。我们提供了一个简单的安装示例：

pip install transformers==4.40.1
pip install av
pip install imageio
pip install decord
pip install opencv-python
pip install flash-attn --no-build-isolation

然后您可以使用我们的模型：

import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer


# 模型设置
model_path = 'OpenGVLab/InternVL_2_5_HiCo_R16'

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(model_path, trust_remote_code=True).half().cuda()

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([T.Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img), T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC), T.ToTensor(), T.Normalize(mean=MEAN, std=STD)])
    return transform


def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float("inf")
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # 计算现有图像的宽高比
    target_ratios = set((i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # 找到最接近的目标宽高比
    target_aspect_ratio = find_closest_aspect_ratio(aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # 计算目标宽度和高度
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # 调整图像大小
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = ((i % (target_width // image_size)) * image_size, (i // (target_width // image_size)) * image_size, ((i % (target_width // image_size)) + 1) * image_size, ((i // (target_width // image_size)) + 1) * image_size)
        # 分割图像
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images


def load_image(image, input_size=448, max_num=6):
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values


def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
    if bound:
        start, end = bound[0], bound[1]
    else:
        start, end = -100000, 100000
    start_idx = max(first_idx, round(start * fps))
    end_idx = min(round(end * fps), max_frame)
    seg_size = float(end_idx - start_idx) / num_segments
    frame_indices = np.array([int(start_idx + (seg_size / 2) + np.round(seg_size * idx)) for idx in range(num_segments)])
    return frame_indices

def get_num_frames_by_duration(duration):
        local_num_frames = 4        
        num_segments = int(duration // local_num_frames)
        if num_segments == 0:
            num_frames = local_num_frames
        else:
            num_frames = local_num_frames * num_segments
        
        num_frames = min(512, num_frames)
        num_frames = max(128, num_frames)

        return num_frames

def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32, get_frame_by_duration = False):
    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
    max_frame = len(vr) - 1
    fps = float(vr.get_avg_fps())

    pixel_values_list, num_patches_list = [], []
    transform = build_transform(input_size=input_size)
    if get_frame_by_duration:
        duration = max_frame / fps
        num_segments = get_num_frames_by_duration(duration)
    frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
    for frame_index in frame_indices:
        img = Image.fromarray(vr[frame_index].asnumpy()).convert("RGB")
        img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
        pixel_values = [transform(tile) for tile in img]
        pixel_values = torch.stack(pixel_values)
        num_patches_list.append(pixel_values.shape[0])
        pixel_values_list.append(pixel_values)
    pixel_values = torch.cat(pixel_values_list)
    return pixel_values, num_patches_list

# 评估设置
max_num_frames = 512
generation_config = dict(
    do_sample=False,
    temperature=0.0,
    max_new_tokens=1024,
    top_p=0.1,
    num_beams=1
)
video_path = "your_video.mp4"
num_segments=128


with torch.no_grad():
  
  pixel_values, num_patches_list = load_video(video_path, num_segments=num_segments, max_num=1, get_frame_by_duration=False)
  pixel_values = pixel_values.to(torch.bfloat16).to(model.device)
  video_prefix = "".join([f"Frame{i+1}: <image>\n" for i in range(len(num_patches_list))])
  # 单轮对话
  question1 = "详细描述这个视频。"
  question = video_prefix + question1
  output1, chat_history = model.chat(tokenizer, pixel_values, question, generation_config, num_patches_list=num_patches_list, history=None, return_history=True)
  print(output1)
  
  # 多轮对话
  question2 = "视频中出现了多少人？"
  output2, chat_history = model.chat(tokenizer, pixel_values, question, generation_config, num_patches_list=num_patches_list, history=chat_history, return_history=True)
  
  print(output2)

✏️ 引用

@article{wang2025internvideo,
  title={InternVideo2.5: 通过长且丰富的上下文建模赋能视频 MLLM},
  author={王毅 and 李鑫豪 and 严子昂 and 何一楠 and 余佳硕 and 曾翔宇 and 王晨婷 and 马常莲 and 黄海安 and 高剑飞 and 窦敏 and 陈凯 and 王温海 and 乔宇 and 王雅丽 and 王立民},
  journal={arXiv 预印本 arXiv:2501.12386},
  year={2025}
}


@article{li2024videochatflash,
  title={VideoChat-Flash: 长上下文视频建模的分层压缩},
  author={李鑫豪 and 王毅 and 余佳硕 and 曾翔宇 and 朱宇涵 and 黄海安 and 高剑飞 and 李坤昌 and 何一楠 and 王晨婷 and 其他},
  journal={arXiv 预印本 arXiv:2501.00574},
  year={2024}
}