license: apache-2.0

Video-LLaVA：通过投影前对齐学习统一视觉表征

若喜欢本项目，请点击GitHub星标⭐获取最新动态。

📰 动态

• [2024.01.27] 👀👀👀 我们的MoE-LLaVA已发布！3B参数的稀疏模型性能超越7B参数的稠密模型。
• [2024.01.17] 🔥🔥🔥 我们的LanguageBind被ICLR 2024接收！
• [2024.01.16] 🔥🔥🔥 代码重构并支持LoRA微调，参见finetune_lora.sh。
• [2023.11.30] 🤝 感谢社区贡献，OpenXLab在线演示已开放。
• [2023.11.23] 正在训练更强大的新模型。
• [2023.11.21] 🤝 查看由@nateraw创建的replicate演示，感谢其对研究的支持！
• [2023.11.20] 🤗 Hugging Face演示及全部代码与数据集已开放！欢迎watch👀本仓库获取更新。

😮 亮点

尽管数据集中不含图像-视频对，Video-LLaVA仍展现出卓越的跨模态交互能力。

💡 简洁基线：通过投影前对齐学习统一视觉表征

• 将统一视觉表征绑定至语言特征空间，使大语言模型能同时处理图像与视频的视觉推理任务。

🔥 高性能：视频与图像的互补学习

• 大量实验证明模态互补性，相比单一模态专用模型具有显著优势。

🤗 演示

Gradio网页界面

强烈建议运行以下命令体验完整功能，我们也提供了Hugging Face在线演示。

python -m videollava.serve.gradio_web_server

命令行推理

python -m videollava.serve.cli --model-path "LanguageBind/Video-LLaVA-7B" --file "视频路径.mp4" --load-4bit

python -m videollava.serve.cli --model-path "LanguageBind/Video-LLaVA-7B" --file "图像路径.jpg" --load-4bit

🛠️ 安装要求

• Python ≥ 3.10
• Pytorch == 2.0.1
• CUDA ≥ 11.7
• 安装依赖：

git clone https://github.com/PKU-YuanGroup/Video-LLaVA
cd Video-LLaVA
conda create -n videollava python=3.10 -y
conda activate videollava
pip install --upgrade pip  # 启用PEP 660支持
pip install -e .
pip install -e ".[train]"
pip install flash-attn --no-build-isolation
pip install decord opencv-python git+https://github.com/facebookresearch/pytorchvideo.git@28fe037d212663c6a24f373b94cc5d478c8c1a1d

🤖 API接口

我们开源全部代码。若需本地加载模型（如LanguageBind/Video-LLaVA-7B），可参考以下代码片段。

图像推理

import torch
from videollava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
from videollava.conversation import conv_templates, SeparatorStyle
from videollava.model.builder import load_pretrained_model
from videollava.utils import disable_torch_init
from videollava.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria

def main():
    disable_torch_init()
    image = 'videollava/serve/examples/extreme_ironing.jpg'
    inp = '这张图片有何特别之处？'
    model_path = 'LanguageBind/Video-LLaVA-7B'
    cache_dir = 'cache_dir'
    device = 'cuda'
    load_4bit, load_8bit = True, False
    model_name = get_model_name_from_path(model_path)
    tokenizer, model, processor, _ = load_pretrained_model(model_path, None, model_name, load_8bit, load_4bit, device=device, cache_dir=cache_dir)
    image_processor = processor['image']
    conv_mode = "llava_v1"
    conv = conv_templates[conv_mode].copy()
    roles = conv.roles

    image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values']
    if type(image_tensor) is list:
        tensor = [image.to(model.device, dtype=torch.float16) for image in image_tensor]
    else:
        tensor = image_tensor.to(model.device, dtype=torch.float16)

    print(f"{roles[1]}: {inp}")
    inp = DEFAULT_IMAGE_TOKEN + '\n' + inp
    conv.append_message(conv.roles[0], inp)
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()
    input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
    stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
    keywords = [stop_str]
    stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)

    with torch.inference_mode():
        output_ids = model.generate(
            input_ids,
            images=tensor,
            do_sample=True,
            temperature=0.2,
            max_new_tokens=1024,
            use_cache=True,
            stopping_criteria=[stopping_criteria])

    outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
    print(outputs)

if __name__ == '__main__':
    main()

视频推理

import torch
from videollava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
from videollava.conversation import conv_templates, SeparatorStyle
from videollava.model.builder import load_pretrained_model
from videollava.utils import disable_torch_init
from videollava.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria

def main():
    disable_torch_init()
    video = 'videollava/serve/examples/sample_demo_1.mp4'
    inp = '这个视频为何有趣？'
    model_path = 'LanguageBind/Video-LLaVA-7B'
    cache_dir = 'cache_dir'
    device = 'cuda'
    load_4bit, load_8bit = True, False
    model_name = get_model_name_from_path(model_path)
    tokenizer, model, processor, _ = load_pretrained_model(model_path, None, model_name, load_8bit, load_4bit, device=device, cache_dir=cache_dir)
    video_processor = processor['video']
    conv_mode = "llava_v1"
    conv = conv_templates[conv_mode].copy()
    roles = conv.roles

    video_tensor = video_processor(video, return_tensors='pt')['pixel_values']
    if type(video_tensor) is list:
        tensor = [video.to(model.device, dtype=torch.float16) for video in video_tensor]
    else:
        tensor = video_tensor.to(model.device, dtype=torch.float16)

    print(f"{roles[1]}: {inp}")
    inp = ' '.join([DEFAULT_IMAGE_TOKEN] * model.get_video_tower().config.num_frames) + '\n' + inp
    conv.append_message(conv.roles[0], inp)
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()
    input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
    stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
    keywords = [stop_str]
    stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)

    with torch.inference_mode():
        output_ids = model.generate(
            input_ids,
            images=tensor,
            do_sample=True,
            temperature=0.1,
            max_new_tokens=1024,
            use_cache=True,
            stopping_criteria=[stopping_criteria])

    outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
    print(outputs)

if __name__ == '__main__':
    main()

🗝️ 训练与验证

详见TRAIN_AND_VALIDATE.md。

👍 致谢

• LLaVA 本项目基础代码库，高效的大语言与视觉助手。
• Video-ChatGPT 贡献了评估代码与数据集。

🙌 相关项目

• LanguageBind 开源的五模态语言检索框架。
• Chat-UniVi 高效利用有限视觉token的框架。

🔒 许可协议

• 主体代码基于Apache 2.0许可。
• 服务仅供研究预览，非商业用途，需遵守LLaMA的模型许可、OpenAI数据使用条款及ShareGPT隐私政策。发现侵权请联系我们。

✏️ 引用

若论文或代码对您的研究有帮助，请引用并点亮星标:star:。

@article{lin2023video,
  title={Video-LLaVA: Learning United Visual Representation by Alignment Before Projection},
  author={Lin, Bin and Zhu, Bin and Ye, Yang and Ning, Munan and Jin, Peng and Yuan, Li},
  journal={arXiv preprint arXiv:2311.10122},
  year={2023}
}

@article{zhu2023languagebind,
  title={LanguageBind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment},
  author={Zhu, Bin and Lin, Bin and Ning, Munan and Yan, Yang and Cui, Jiaxi and Wang, HongFa and Pang, Yatian and Jiang, Wenhao and Zhang, Junwu and Li, Zongwei and others},
  journal={arXiv preprint arXiv:2310.01852},
  year={2023}
}

✨ 星标历史

🤝 贡献者