Video Feature Extraction with Qwen Vision-Language Models

Overview

This repository contains two Python scripts to extract video-level features from MP4 files using Qwen multimodal (vision-language) models:

1. qwen-feature-vision-extraction.py
   • Uses Qwen2.5-VL-7B-Instruct
2. extract-feature-with-text-alignment.py
   • Uses Qwen2-VL-2B-Instruct

Both scripts process videos in two classes—plausible and implausible—and produce an HDF5 file containing feature vectors and class labels.

---

Requirements

• Python 3.8 or higher
• CUDA-enabled GPU (recommended)
• PyTorch
• transformers
• bitsandbytes
• tqdm
• h5py
• numpy

You also need the helper module qwen_vl_utils in your PYTHONPATH.

---

Installation

1. Clone this repository:

   git clone https://github.com/your-username/qwen-video-feature-extraction.git
   cd qwen-video-feature-extraction

2. Create a virtual environment and install dependencies:

   python -m venv venv
   source venv/bin/activate       # On Windows: venv\Scripts\activate
   pip install --upgrade pip
   pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
   pip install transformers bitsandbytes tqdm h5py numpy

3. Ensure qwen_vl_utils.py is accessible (either in the project root or installed as a package).