You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository accompanies an MSc thesis from the University of Victoria (2025). Access is granted for research and educational purposes.
Log in or Sign Up to review the conditions and access this model content.
Hyperbolic Flamingo
A vision-language model that leverages hyperbolic geometry for multimodal hateful meme detection.
Overview
This repository contains the implementation for research investigating whether hyperbolic embeddings can improve vision-language models (VLMs) for hateful meme classification. The architecture combines:
- Frozen vision encoders (CLIP or MERU)
- Frozen language model (GPT-2)
- Trainable Flamingo-style gated cross-attention
- Hyperbolic (Lorentzian) embedding space
The work explores native hyperbolic contrastive losses and documents numerical stability considerations for training hyperbolic VLMs.
Access
This repository has restricted access. To request access:
- Click the "Request Access" button above
- Fill out the brief form with your affiliation and intended use
- Access is typically granted within 24-48 hours
For questions, contact: rkwarner@uvic.ca
Requirements
- Python 3.8+
- PyTorch 2.0+ with CUDA
- 24GB+ GPU VRAM (48GB recommended)
Usage
# Install dependencies
pip install -r requirements.txt
# Run training (Euclidean baseline)
python hyperbolic_flamingo.py --config configs/hf_euclidean.yaml
# Run training (Hyperbolic)
python hyperbolic_flamingo.py --config configs/hf_lorentzian.yaml
Configuration files in configs/ control encoder choice, geometry mode, loss weights, and training hyperparameters. See config files for dataset path setup.
Citation
@mastersthesis{warner2025hyperbolic,
title={Hyperbolic Visual Language Models for Hateful Meme Classification},
author={Warner, Ryan},
school={University of Victoria},
year={2025}
}
License
MIT