Developed by Dr. Abdulmajeed Nomman & Athena AGI Research Lab – Stanford University
VGT-Core is a Python module that implements multi-head geometric attention over graphs extracted from visual data (e.g., images, scenes). It features flexible edge modeling, geometric disentanglers, and scalable architecture for learning structured patterns in vision.
- ✅ Multi-head attention with independent geometric edge modules
- 🧠 GeometryDisentangler for per-head pattern specialization
- 🔍 Modular utility toolkit: edge normalization, attention diversity loss, visualization
- ⚙️ Ready for PyTorch training pipelines
- 🔁 Designed for explainability and fine-grained pattern analysis
pip install vgt-coreOr from source:
git clone https://github.com/DataSoftcoAI/athena-labs
cd vgt_core_pkg
pip install .from vgt_core.core.vgt_core_v0_3 import VGTCoreV03
import torch
# Sample inputs
N, D, H = 12, 16, 4
node_features = torch.randn(N, D)
edge_features = torch.randn(N, N, 4)
adj_mask = torch.ones(N, N)
# Initialize model
model = VGTCoreV03(embed_dim=D, num_heads=H, disentangle=True)
# Forward pass
output, attn_weights = model(node_features, edge_features, adj_mask)| Module | Function |
|---|---|
EdgeNormalizer |
Normalize distance/angle edge features |
AttentionDiversityLoss |
Enforce head-level specialization |
GeometryDataSynthesizer |
Generate synthetic geometric test images |
AttentionVisualizer |
Render head-wise graph attention maps |
If used in research, cite as:
@inproceedings{nomman2025vgt,
title={Disentangled Geometric Attention for Structured Visual Reasoning},
author={Nomman, Abdulmajeed and Athena AGI Team},
year={2025},
note={Preprint: arXiv & CVPR 2026 Submission}
}
Apache 2.0 — Free for academic and commercial use with attribution.
“Every pattern in nature has geometry. Every geometry tells a story worth learning by machines.” — Athena
We provide a minimal package for experimentation and prototyping.
It includes:
- ✅ Core multi-head attention model
- ⚙️ Edge normalization utility
- 🧪 Minimal test script
Perfect for research notebooks, classroom demos, or prototyping new attention heads.
Athena labs(AGI)
Multi-Head Geometric Attention for Vision-based Graph Models
Developed by Dr. Abdulmajeed Nomman and Athena Research Team
VGT is a PyTorch module implementing a Multi-Head Geometric Attention mechanism. It is designed for graph-based computer vision models that represent images as spatial graphs (keypoints and geometric relations).
This package is part of the VGT-Core v0.2 project.
- Multi-head attention over image-based graphs
- Learnable edge embeddings (distance, angle, orientation)
- LayerNorm + Dropout for stability
- Modular design, easy integration with GNN pipelines
- Ideal for shape analysis, sketch recognition, or structural vision tasks
pip install vgtOr from source:
git clone https://github.com/DataSoftcoAI/athena-labs/VGT-Core-v0.2
cd VGT-Core-v0.2
pip install .from vgt.multihead_geometric_attention import MultiHeadGeometricAttention
model = MultiHeadGeometricAttention(embed_dim=16, num_heads=4)
output, attn_weights = model(
node_features, # Tensor [N, D]
edge_features, # Tensor [N, N, 4]
adj_mask # Tensor [N, N] (binary mask)
)node_features: Features for each keypoint/node in the image.edge_features: Geometric relations (distance, angle, cos, sin).adj_mask: Binary adjacency matrix representing graph connectivity.
- Structural image classification
- Sketch-based search and retrieval
- Scene graph modeling
- Robotics: object part understanding
- Human pose as graph with geometric links
If you use this package in your research, please cite:
@software{vgt2025,
author = {Abdulmajeed Nomman , Athena (AGI) Team},
title = {VGT - Vision Geometry Transformer},
year = {2025},
url = {https://github.com/DataSoftcoAI/athena-labs/VGT-Core-v0.2}
}
GitHub Repository → VGT-Core v0.2
"لكل عين صناعية... عدسة ترى من زاوية مختلفة، ومعًا نرى بوضوح أكثر." — أثينا