license: mit pipeline_tag: video-text-to-text extra_gated_prompt: >- 您同意不使用该模型进行对人类受试者造成伤害的实验。 extra_gated_fields: 姓名: text 公司/组织: text 国家: text 电子邮箱: text

InternVideo2-Chat-8B-InternLM2.5

[📂 GitHub] [📜 技术报告]

为了进一步丰富InternVideo2中嵌入的语义并提升其在人机交互中的友好性，我们通过将InternVideo2整合到一个带有大型语言模型(LLM)和视频BLIP的VideoLLM中来进行调优。我们采用VideoChat中的渐进式学习方案，使用InternVideo2作为视频编码器，并训练一个视频BLIP来与开源LLM进行交互。在训练过程中，视频编码器会进行更新。详细的训练方法参见VideoChat。该模型经过高清训练。

该模型的基础LLM是InternLM2.5-7B，支持100万token的长上下文窗口。

📈 性能表现

🚀 如何使用该模型

1. 确保安装transformers >= 4.38.0, peft==0.5.0

从pip_requirements安装必要的Python包

2. 使用视频输入进行推理

import os
import torch

from transformers import AutoTokenizer, AutoModel

tokenizer =  AutoTokenizer.from_pretrained('OpenGVLab/InternVideo2_Chat_8B_InternLM2_5',
    trust_remote_code=True,
    use_fast=False,)
if torch.cuda.is_available():
  model = AutoModel.from_pretrained(
      'OpenGVLab/InternVideo2_Chat_8B_InternLM2_5',
      torch_dtype=torch.bfloat16,
      trust_remote_code=True).cuda()
else:
  model = AutoModel.from_pretrained(
      'OpenGVLab/InternVideo2_Chat_8B_InternLM2_5',
      torch_dtype=torch.bfloat16,
      trust_remote_code=True)


from decord import VideoReader, cpu
from PIL import Image
import numpy as np
import numpy as np
import decord
from decord import VideoReader, cpu
import torch.nn.functional as F
import torchvision.transforms as T
from torchvision.transforms import PILToTensor
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode
decord.bridge.set_bridge("torch")

def get_index(num_frames, num_segments):
    seg_size = float(num_frames - 1) / num_segments
    start = int(seg_size / 2)
    offsets = np.array([
        start + int(np.round(seg_size * idx)) for idx in range(num_segments)
    ])
    return offsets


def load_video(video_path, num_segments=8, return_msg=False, resolution=224, hd_num=4, padding=False):
    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
    num_frames = len(vr)
    frame_indices = get_index(num_frames, num_segments)

    mean = (0.485, 0.456, 0.406)
    std = (0.229, 0.224, 0.225)

    transform = transforms.Compose([
        transforms.Lambda(lambda x: x.float().div(255.0)),
        transforms.Normalize(mean, std)
    ])

    frames = vr.get_batch(frame_indices)
    frames = frames.permute(0, 3, 1, 2)

    if padding:
        frames = HD_transform_padding(frames.float(), image_size=resolution, hd_num=hd_num)
    else:
        frames = HD_transform_no_padding(frames.float(), image_size=resolution, hd_num=hd_num)

    frames = transform(frames)
    # print(frames.shape)
    T_, C, H, W = frames.shape

    sub_img = frames.reshape(
        1, T_, 3, H//resolution, resolution, W//resolution, resolution
    ).permute(0, 3, 5, 1, 2, 4, 6).reshape(-1, T_, 3, resolution, resolution).contiguous()

    glb_img = F.interpolate(
        frames.float(), size=(resolution, resolution), mode='bicubic', align_corners=False
    ).to(sub_img.dtype).unsqueeze(0)

    frames = torch.cat([sub_img, glb_img]).unsqueeze(0)

    if return_msg:
        fps = float(vr.get_avg_fps())
        sec = ", ".join([str(round(f / fps, 1)) for f in frame_indices])
        # 开头和结尾需要添加空格
        msg = f"视频包含{len(frame_indices)}帧，采样于{sec}秒。"
        return frames, msg
    else:
        return frames

def HD_transform_padding(frames, image_size=224, hd_num=6):
    def _padding_224(frames):
        _, _, H, W = frames.shape
        tar = int(np.ceil(H / 224) * 224)
        top_padding = (tar - H) // 2
        bottom_padding = tar - H - top_padding
        left_padding = 0
        right_padding = 0

        padded_frames = F.pad(
            frames,
            pad=[left_padding, right_padding, top_padding, bottom_padding],
            mode='constant', value=255
        )
        return padded_frames

    _, _, H, W = frames.shape
    trans = False
    if W < H:
        frames = frames.flip(-2, -1)
        trans = True
        width, height = H, W
    else:
        width, height = W, H

    ratio = width / height
    scale = 1
    while scale * np.ceil(scale / ratio) <= hd_num:
        scale += 1
    scale -= 1
    new_w = int(scale * image_size)
    new_h = int(new_w / ratio)

    resized_frames = F.interpolate(
        frames, size=(new_h, new_w),
        mode='bicubic',
        align_corners=False
    )
    padded_frames = _padding_224(resized_frames)

    if trans:
        padded_frames = padded_frames.flip(-2, -1)

    return padded_frames

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 HD_transform_no_padding(frames, image_size=224, hd_num=6, fix_ratio=(2,1)):
    min_num = 1
    max_num = hd_num
    _, _, orig_height, orig_width = frames.shape
    aspect_ratio = orig_width / orig_height

    # 计算现有视频的宽高比
    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])

    # 找到最接近目标宽高比的比例
    if fix_ratio:
        target_aspect_ratio = fix_ratio
    else:
        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]

    # 调整帧大小
    resized_frame = F.interpolate(
        frames, size=(target_height, target_width),
        mode='bicubic', align_corners=False
    )
    return resized_frame

video_path = "yoga.mp4"
# 从视频中均匀采样8帧
video_tensor = load_video(video_path, num_segments=8, return_msg=False, resolution=224, hd_num=6)
video_tensor = video_tensor.to(model.device)

chat_history = []
response, chat_history = model.chat(tokenizer, '', '逐步描述视频内容',instruction= "仔细观察视频，注意事件的因果关系和顺序、物体的细节和运动，以及人物的动作和姿势。根据你的观察，选择最准确回答问题的选项。\n", media_type='video', media_tensor=video_tensor, chat_history= chat_history, return_history=True,generation_config={'do_sample':False,'max_new_tokens':512,})
print(response) 

✏️ 引用

如果这项工作对您的研究有所帮助，请考虑引用InternVideo和VideoChat。

@article{wang2024internvideo2,
  title={Internvideo2: Scaling video foundation models for multimodal video understanding},
  author={Wang, Yi and Li, Kunchang and Li, Xinhao and Yu, Jiashuo and He, Yinan and Wang, Chenting and Chen, Guo and Pei, Baoqi and Zheng, Rongkun and Xu, Jilan and Wang, Zun and others},
  journal={arXiv preprint arXiv:2403.15377},
  year={2024}
}

@article{li2023videochat,
  title={Videochat: Chat-centric video understanding},
  author={Li, KunChang and He, Yinan and Wang, Yi and Li, Yizhuo and Wang, Wenhai and Luo, Ping and Wang, Yali and Wang, Limin and Qiao, Yu},
  journal={arXiv preprint arXiv:2305.06355},
  year={2023}
}