• [2024.10.22] Release checkpoints of VideoLLaMA2.1-7B-AV.
• [2024.10.15] Release checkpoints of VideoLLaMA2.1-7B-16F-Base and VideoLLaMA2.1-7B-16F.
• [2024.08.14] Release checkpoints of VideoLLaMA2-72B-Base and VideoLLaMA2-72B.
• [2024.07.30] Release checkpoints of VideoLLaMA2-8x7B-Base and VideoLLaMA2-8x7B.
• [2024.06.25] 🔥🔥 As of Jun 25, our VideoLLaMA2-7B-16F is the Top-1 ~7B-sized VideoLLM on the MLVU Leaderboard.
• [2024.06.18] 🔥🔥 As of Jun 18, our VideoLLaMA2-7B-16F is the Top-1 ~7B-sized VideoLLM on the VideoMME Leaderboard.
• [2024.06.17] 👋👋 Update technical report with the latest results and the missing references. If you have works closely related to VideoLLaMA 2 but not mentioned in the paper, feel free to let us know.
• [2024.06.14] 🔥🔥 Online Demo is available.
• [2024.06.03] Release training, evaluation, and serving codes of VideoLLaMA 2.

Basic Dependencies:

• Python >= 3.8
• Pytorch >= 2.2.0
• CUDA Version >= 11.8
• transformers == 4.40.0 (for reproducing paper results)
• tokenizers == 0.19.1

[Online Mode] Install required packages (better for development):

git clone https://github.com/DAMO-NLP-SG/VideoLLaMA2
cd VideoLLaMA2
git checkout audio_visual
pip install -r requirements.txt
pip install flash-attn==2.5.8 --no-build-isolation
pip install opencv-python==4.5.5.64
apt-get update && apt-get install ffmpeg libsm6 libxext6  -y

[Offline Mode] Install VideoLLaMA2 as a Python package (better for direct use):

git clone https://github.com/DAMO-NLP-SG/VideoLLaMA2
cd VideoLLaMA2
git checkout audio_visual
pip install --upgrade pip  # enable PEP 660 support
pip install -e .
pip install flash-attn==2.5.8 --no-build-isolation
pip install opencv-python==4.5.5.64
apt-get update && apt-get install ffmpeg libsm6 libxext6  -y

It is highly recommended to try our online demo first.

To run a video-based LLM (Large Language Model) web demonstration on your device, you will first need to ensure that you have the necessary model checkpoints prepared, followed by adhering to the steps outlined to successfully launch the demo.

• Launch a gradio app directly (VideoLLaMA2.1-7B-AV is adopted by default):

python videollama2/serve/gradio_web_server_adhoc_av.py

To facilitate further development on top of our codebase, we provide a quick-start guide on how to train a customized VideoLLaMA2 with VideoLLaVA dataset and evaluate the trained model on the mainstream video-llm benchmarks.

1. Training Data Structure: Follow the main branch(https://github.com/DAMO-NLP-SG/VideoLLaMA2/tree/main) of this VideoLLaMA2 codebase.
2. Command:

# VideoLLaMA2.1-audio pretraining
bash scripts/custom/pretrain_audio.sh
# VideoLLaMA2.1-audio finetuning
bash scripts/custom/finetune_audio.sh
# VideoLLaMA2.1-audio_visual finetuning
bash scripts/custom/va_joint.sh

3. Evaluation Data Structure: Follow the main branch(https://github.com/DAMO-NLP-SG/VideoLLaMA2/tree/main) of this VideoLLaMA2 codebase.

4. Command:

# ClothoAQA.sh evaluation
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/eval_audio_clothoAQA.sh
# TUT2017 evaluation
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/eval_audio_TUT2017.sh
# VocalSound evaluation
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/eval_audio_vocalsound.sh
# AVQA_music evaluation
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/eval_audio_video_AVQA.sh
# AVSD evaluation (need to set azure openai key/endpoint/deployname)
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/eval_audio_video_AVSD.sh
# AVSSD evaluation (need to set azure openai key/endpoint/deployname)
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash scripts/eval/eval_audio_video_AVSSD.sh

If you want to train a video-llm on your data, you need to follow the procedures below to prepare the audio/video/image sft data:

1. Suppose your data structure is like:

VideoLLaMA2
├── datasets
│   ├── custom_sft
│   |   ├── audio
│   |   ├── video
│   |   ├── image
|   |   └── custom.json

2. Then you should re-organize the annotated audio/video/image sft data according to the following format:

[
    {
        "id": 0,
        "audio": "audio/xxx.wav",
        "conversations": [
            {
                "from": "human",
                "value": "<audio>\nPlease describe the sound event within the audio."
            },
            {
                "from": "gpt",
                "value": "Loud television static dips in and out of focus."
            },
            ...
        ],
    }
    {
        "id": 1,
        "video": "images/xxx.jpg",
        "conversations": [
            {
                "from": "human",
                "value": "<image>\nWhat are the colors of the bus in the image?"
            },
            {
                "from": "gpt",
                "value": "The bus in the image is white and red."
            },
            ...
        ],
    }
    {
        "id": 2,
        "video": "videos/xxx.mp4",
        "conversations": [
            {
                "from": "human",
                "value": "<video>\nWhat are the main activities that take place in the video?"
            },
            {
                "from": "gpt",
                "value": "The main activities that take place in the video are the preparation of camera equipment by a man, a group of men riding a helicopter, and a man sailing a boat through the water."
            },
            ...
        ],
    },
    ...
]

3. Modify the scripts/custom/finetune_audio.sh:

...
--data_path datasets/custom_sft/custom.json
--data_folder datasets/custom_sft/
--pretrain_mm_mlp_adapter CONNECTOR_DOWNLOAD_PATH (e.g., DAMO-NLP-SG/VideoLLaMA2.1-7B-16F)
...

4. Modify the scripts/custom/va_joint.sh:

...
--data_path datasets/custom_sft/custom.json
--data_folder datasets/custom_sft/
--pretrain_mm_mlp_adapter CONNECTOR_DOWNLOAD_PATH (e.g., DAMO-NLP-SG/VideoLLaMA2.1-7B-16F)
...

Audio/Video-Audio Inference:

import sys
sys.path.append('./')
from videollama2 import model_init, mm_infer
from videollama2.utils import disable_torch_init
import argparse
def inference(args):
    model_path = args.model_path
    model, processor, tokenizer = model_init(model_path)
    if args.modal_type == "a":
        model.model.vision_tower = None
    elif args.modal_type == "v":
        model.model.audio_tower = None
    elif args.modal_type == "av":
        pass
    else:
        raise NotImplementedError
    # Audio-visual Inference
    audio_video_path = "assets/00000368.mp4"
    preprocess = processor['audio' if args.modal_type == "a" else "video"]
    if args.modal_type == "a":
        audio_video_tensor = preprocess(audio_video_path)
    else:
        audio_video_tensor = preprocess(audio_video_path, va=True if args.modal_type == "av" else False)
    question = f"Who plays the instrument louder?"
    # Audio Inference
    audio_video_path = "assets/bird-twitter-car.wav"
    preprocess = processor['audio' if args.modal_type == "a" else "video"]
    if args.modal_type == "a":
        audio_video_tensor = preprocess(audio_video_path)
    else:
        audio_video_tensor = preprocess(audio_video_path, va=True if args.modal_type == "av" else False)
    question = f"Please describe the audio:"
    # Video Inference
    audio_video_path = "assets/output_v_1jgsRbGzCls.mp4"
    preprocess = processor['audio' if args.modal_type == "a" else "video"]
    if args.modal_type == "a":
        audio_video_tensor = preprocess(audio_video_path)
    else:
        audio_video_tensor = preprocess(audio_video_path, va=True if args.modal_type == "av" else False)
    question = f"What activity are the people practicing in the video?"
    output = mm_infer(
        audio_video_tensor,
        question,
        model=model,
        tokenizer=tokenizer,
        modal='audio' if args.modal_type == "a" else "video",
        do_sample=False,
    )
    print(output)
if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('--model-path', help='', required=True)
    parser.add_argument('--modal-type', choices=["a", "v", "av"], help='', required=True)
    args = parser.parse_args()
    inference(args)

If you find VideoLLaMA useful for your research and applications, please cite using this BibTeX:

@article{damonlpsg2024videollama2,
  title={VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs},
  author={Cheng, Zesen and Leng, Sicong and Zhang, Hang and Xin, Yifei and Li, Xin and Chen, Guanzheng and Zhu, Yongxin and Zhang, Wenqi and Luo, Ziyang and Zhao, Deli and Bing, Lidong},
  journal={arXiv preprint arXiv:2406.07476},
  year={2024},
  url = {https://arxiv.org/abs/2406.07476}
}
@article{damonlpsg2023videollama,
  title = {Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding},
  author = {Zhang, Hang and Li, Xin and Bing, Lidong},
  journal = {arXiv preprint arXiv:2306.02858},
  year = {2023},
  url = {https://arxiv.org/abs/2306.02858}
}

The codebase of VideoLLaMA 2 is adapted from LLaVA 1.5 and FastChat. We are also grateful for the following projects our VideoLLaMA 2 arise from:

• LLaMA 2, Mistral-7B, OpenAI CLIP, Honeybee.
• Video-ChatGPT, Video-LLaVA.
• WebVid, Panda-70M, LanguageBind, InternVid.
• VideoChat2, Valley, VTimeLLM, ShareGPT4V.

This project is released under the Apache 2.0 license as found in the LICENSE file. The service is a research preview intended for non-commercial use ONLY, subject to the model Licenses of LLaMA and Mistral, Terms of Use of the data generated by OpenAI, and Privacy Practices of ShareGPT. Please get in touch with us if you find any potential violations.