README.md

🎥 VidChain Exercise: Chain-of-Tasks with Metric-based Direct Preference Optimization

📚 About This Repository

This repository contains the exercise materials and implementation for VidChain, a novel framework for Dense Video Captioning with VideoLLMs. VidChain combines Chain-of-Tasks (CoTasks) and Metric-based Direct Preference Optimization (M-DPO) to achieve superior temporal reasoning and coherence in video understanding.

🎯 Research Paper

VidChain: Chain-of-Tasks with Metric-based Direct Preference Optimization for Dense Video Captioning
Ji Soo Lee, Jongha Kim*, Jeehye Na, Jinyoung Park, Hyunwoo J. Kim†
AAAI 2025

🚀 Learning Objectives

By working through this exercise, you will:

• ✅ Reproduce baseline behavior of a video-language model (VTimeLLM, CVPR 2024 Highlight)
• 🔍 Observe limitations of existing approaches in temporal reasoning and coherence
• 🛠️ Implement and experiment with VidChain's improvements using M-DPO
• 🎬 Run inference on videos to generate dense temporal captions (Dense Video Captioning)
• 📊 Evaluate how preference alignment improves performance over baselines
• 💡 Discuss strategies for ensembling different reasoning paths of VidChain's CoTasks

📁 Repository Structure

VidChain-exercise/
├── README_HF.md              # This file - Hugging Face README
├── READ.md                   # Original exercise README
├── upload.py                 # Upload script for Hugging Face Hub
├── upload_single_file.py     # Single file upload utility
├── remove_file.py           # File removal utility
├── setup_hf_upload.py       # Setup script for HF upload
├── HF_UPLOAD_GUIDE.md       # Comprehensive upload guide
├── requirements.txt         # Python dependencies
├── app.py                   # Streamlit app for HF Spaces
├── asset/                   # Project assets and images
│   └── main.png            # Main framework diagram
└── VTimeLLM/               # VideoLLM implementation
    └── ...                 # VTimeLLM source code

🔧 Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Setup for VideoLLaMA2

# Clone the main repository
git clone https://github.com/mlvlab/VidChain.git
cd VidChain

# Install VideoLLaMA2 dependencies
conda create -n videollama python=3.10 -y
conda activate videollama
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121

cd VideoLLaMA2
pip install -r requirements.txt
pip install num2words datasets pycocoevalcap rich
pip install flash-attn==2.5.7 --no-build-isolation

3. Download Pre-trained Models

• VideoLLaMA2 checkpoints: Download from official repo
• VidChain checkpoints: Download from Hugging Face

🎯 Key Features

Chain-of-Tasks (CoTasks)

• Novel approach to video understanding through task decomposition
• Improves temporal reasoning capabilities
• Enhanced coherence in video captioning

Metric-based Direct Preference Optimization (M-DPO)

• Advanced training methodology for preference alignment
• Better performance over traditional baselines
• Improved temporal consistency

VideoLLaMA2 Integration

• State-of-the-art video-language model
• Support for ActivityNet and YouCook2 datasets
• Pre-extracted features for efficient training

📊 Dataset Support

This exercise supports two major datasets:

1. ActivityNet (301GB pre-extracted features)
2. YouCook2 (32GB pre-extracted features)

⚠️ Storage Warning: The pre-extracted features require significant storage space. Please ensure you have adequate disk space.

🔗 Related Resources

• Paper: VidChain: Chain-of-Tasks with Metric-based Direct Preference Optimization for Dense Video Captioning
• Main Repository: mlvlab/VidChain
• Dataset: simplecloud/VidChain-Data
• VideoLLaMA2: DAMO-NLP-SG/VideoLLaMA2

📝 Citation

If you find this work useful, please cite:

@inproceedings{lee2025vidchain,
  title={VidChain: Chain-of-Tasks with Metric-based Direct Preference Optimization for Dense Video Captioning},
  author={Lee, Ji Soo and Kim, Jongha and Na, Jeehye and Park, Jinyoung and Kim, Hyunwoo J},
  booktitle={AAAI},
  year={2025}
}

🤝 Contributing

This is an exercise repository for educational purposes. For contributions to the main VidChain project, please visit the main repository.

📄 License

This project is released under the same license as the main VidChain repository. Please refer to the main repository for license details.

---

Built with ❤️ for the AI research community

Part of the VidChain research project at AAAI 2025