Internvideo2 Chat 8B HD

🚀 InternVideo2-Chat-8B-HD

InternVideo2-Chat-8B-HD是一個視頻文本處理模型，它將InternVideo2融入到VideoLLM中，結合大語言模型和視頻BLIP，進一步豐富語義並提升人機交互的友好性。該模型在視頻理解任務中表現出色，具有較高的性能指標。

🚀 快速開始

申請權限

申請本項目的使用權限以及基礎大語言模型的訪問權限。此模型的基礎大語言模型是Mistral - 7B，使用前請確保已獲得Mistral - 7B的訪問權限，若尚未獲得，請前往[Mistral - 7B](https://huggingface.co/mistralai/Mistral - 7B - Instruct - v0.3)獲取訪問權限，並將你的HF_token添加到環境變量中。

設置環境變量

將HF用戶訪問令牌填充到環境變量中：

export HF_TOKEN=hf_....

若不知道如何獲取以“hf_”開頭的令牌，請參考：[How to Get HF User access Token](https://huggingface.co/docs/hub/security - tokens#user - access - tokens)

安裝依賴

確保安裝了transformers >= 4.38.0，並從pip_requirements安裝所需的Python包。

視頻輸入推理

import os
token = os.environ['HF_TOKEN']
import torch

from transformers import AutoTokenizer, AutoModel

tokenizer =  AutoTokenizer.from_pretrained('OpenGVLab/InternVideo2_chat_8B_HD',
    trust_remote_code=True,
    use_fast=False,
    token=token)
if torch.cuda.is_available():
  model = AutoModel.from_pretrained(
      'OpenGVLab/InternVideo2_chat_8B_HD',
      torch_dtype=torch.bfloat16,
      trust_remote_code=True).cuda()
else:
  model = AutoModel.from_pretrained(
      'OpenGVLab/InternVideo2_chat_8B_HD',
      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])
        # " " should be added in the start and end
        msg = f"The video contains {len(frame_indices)} frames sampled at {sec} seconds."
        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

    # calculate the existing video aspect ratio
    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])

    # find the closest aspect ratio to the target
    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)

    # calculate the target width and height
    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]

    # resize the frames
    resized_frame = F.interpolate(
        frames, size=(target_height, target_width),
        mode='bicubic', align_corners=False
    )
    return resized_frame

video_path = "yoga.mp4"
# sample uniformly 8 frames from the video
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, '', 'Describe the action step by step.', media_type='video', media_tensor=video_tensor, chat_history= chat_history, return_history=True,generation_config={'do_sample':False})
print(response)

response, chat_history = model.chat(tokenizer, '', 'What is she wearing?', media_type='video', media_tensor=video_tensor, chat_history= chat_history, return_history=True,generation_config={'do_sample':False})

✨ 主要特性

為了進一步豐富InternVideo2中嵌入的語義，並提高其在人機通信中的友好性，我們將InternVideo2融入到一個結合了大語言模型和視頻BLIP的VideoLLM中進行微調。我們採用了VideoChat中的漸進式學習方案，使用InternVideo2作為視頻編碼器，並訓練了一個視頻blip用於與開源大語言模型進行通信。在訓練過程中，視頻編碼器會被更新。詳細的訓練方法請參考VideoChat。該模型經過了高清訓練。

📈 性能表現

✏️ 引用

如果本工作對你的研究有幫助，請考慮引用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}
}

📄 許可證

本項目採用MIT許可證。

⚠️ 重要提示

你同意不使用該模型進行對人類受試者造成傷害的實驗。

💡 使用建議

使用前請確保已獲得Mistral - 7B的訪問權限，並將HF_token添加到環境變量中。