Atom: Text-Guided Geometric Relation Parsing

This repository contains the paper-facing code for a lightweight geometric relation parser built on a derived Ext-PGDP5K protocol.

The public package is intentionally clean:

• no private logs
• no training outputs
• no full PGDP5K data dump
• no remote-server credentials or deployment helpers

Instead, it includes a tiny runnable demo subset so the project can be cloned and executed immediately.

Repository Layout

• src/: model, data loader, logic regularization, metrics
• train.py: training entry point
• scripts/: protocol and evaluation utilities
• docs/ext_pgdp5k_protocol.md: derived protocol notes
• demo_data/PGDP5K_demo/: small metadata-only demo split
• run_demo.ps1: one-click Windows demo
• run_demo.bat: one-click Windows batch wrapper

Quick Start

1. Create a Python environment.
2. Install dependencies:

pip install -r requirements.txt

3. Run the demo:

.\run_demo.ps1

Or:

run_demo.bat

The script trains for one short epoch on the included demo subset and writes outputs to outputs/demo_run.

Demo Data

The included demo_data/PGDP5K_demo subset is a tiny metadata-only package prepared from a few samples for reproducibility checks. It is not intended for reporting final paper numbers.

The full PGDP5K dataset and the full derived Ext-PGDP5K protocol should be prepared separately for real experiments.

Example Manual Command

python train.py ^
  --data-root demo_data/PGDP5K_demo ^
  --ext-root demo_data/PGDP5K_demo/Ext-PGDP5K ^
  --epochs 1 ^
  --batch-size 2 ^
  --device cpu ^
  --output-dir outputs/demo_run

Notes

• The demo package is designed for smoke testing and code review.
• Final paper experiments should be run on the full dataset.
• The current active relation setting contains four evaluated labels: intersect, parallel, perpendicular, and bisect.