AbstractPhil / geovit-david-beans

🫘💎 DavidBeans: Unified Vision-to-Crystal Architecture

DavidBeans combines ViT-Beans (Cantor-routed sparse attention) with David (multi-scale crystal classification) into a unified geometric deep learning architecture.

Model Description

This model implements several novel techniques:

• Hybrid Cantor Routing : Combines fractal Cantor set distances with positional proximity for sparse attention patterns
• Pentachoron Experts : 5-vertex simplex structure with Cayley-Menger geometric regularization
• Multi-Scale Crystal Projection : Projects features to multiple representation scales with learned fusion
• Cross-Contrastive Learning : Aligns patch-level features with crystal anchors

Architecture

Image [B, 3, 32, 32]
       │
       ▼
┌─────────────────────────────────────────┐
│  BEANS BACKBONE                         │
│  ├─ Patch Embed → [64 patches, 512d]
│  ├─ Hybrid Cantor Router (α=0.3)
│  ├─ 8 × Attention Blocks (8 heads)
│  └─ 8 × Pentachoron Expert Layers
└─────────────────────────────────────────┘
       │
       ▼
┌─────────────────────────────────────────┐
│  DAVID HEAD                             │
│  ├─ Multi-scale projection: [256, 512, 768]
│  ├─ Per-scale Crystal Heads
│  └─ Geometric Fusion (learned weights)
└─────────────────────────────────────────┘
       │
       ▼
    [100 classes]

Training Details

Results

TensorBoard Logs

Training logs are included in the tensorboard/ directory. To view:

tensorboard --logdir tensorboard/

Usage

import torch
from safetensors.torch import load_file
from david_beans import DavidBeans, DavidBeansConfig

# Load config
config = DavidBeansConfig(
    image_size=32,
    patch_size=4,
    dim=512,
    num_layers=8,
    num_heads=8,
    num_experts=5,
    k_neighbors=32,
    cantor_weight=0.3,
    scales=[256, 512, 768],
    num_classes=100
)

# Create model and load weights
model = DavidBeans(config)
state_dict = load_file("model.safetensors")
model.load_state_dict(state_dict)

# Inference
model.eval()
with torch.no_grad():
    output = model(images)
    predictions = output['logits'].argmax(dim=-1)

Citation

@misc{davidbeans2025,
  author = {AbstractPhil},
  title = {DavidBeans: Unified Vision-to-Crystal Architecture},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/AbstractPhil/geovit-david-beans}
}

License

Apache 2.0