Multimodal Disaster Response

This project implements a multi-model disaster response prediction system using pre-trained SoTA Embeddings models (MTEB leaderboard) from Hugging Face. It provides a Gradio interface for easy interaction and prediction.

Installation

1. Clone the repository and navigate to the project folder:

   git clone https://github.com/s-smits/multimodal-disaster-response
   cd multimodal-disaster-response
   

2. Create a virtual environment named venv_disaster_response and activate it:

   • For macOS and Linux:

     python3 -m venv venv_disaster_response
     source venv_disaster_response/bin/activate
     

   • For Windows:

     python -m venv venv_disaster_response
     venv_disaster_response\Scripts\activate
     

3. Install the required dependencies:

   pip install -r requirements.txt
   

4. Set up your Hugging Face token as an environment variable:

   export HF_TOKEN=your_huggingface_token_here
   

   Make sure to enter your own Hugging Face token here. This token is necessary for downloading the pre-trained models used in this project.

Usage

Start the script:

python main.py

This opens a Gradio interface where you can:

1. Enter up to 10 disaster-related texts
2. Get predictions for each text, including:
   • Timeframe (Preparedness, Response, Other)
   • Transfer type (Request, Provide, Other)
   • Disaster response labels (with multiple thresholds)
   • Overall relevance

Models

The project uses three pre-trained models:

• Timeframe model: ssmits/best-timeframe-model-disaster-response
• Transfer type model: ssmits/best-transfer-type-model-disaster-response
• Disaster response model: ssmits/best-actionable-labelling-model-disaster-response

Requirements

See requirements.txt for a full list of dependencies. Key libraries include:

• gradio
• numpy
• transformers
• tensorflow
• huggingface_hub

License

MIT License