This repo provides a prototype pipeline to handle strictly-Manhattan binary IC layout images: extract skeletons, vectorize skeletons into Manhattan polylines, train a skeleton-generation model, and expand skeletons back to images.

Project Structure

scripts/: Core processing scripts.
- skeleton_extract.py: Binarize & extract skeleton PNG.
- vectorize_skeleton.py: Trace skeleton PNG into JSON polylines.
- expand_skeleton.py: Rasterize polyline JSON back to a PNG.
models/: PyTorch models (e.g., SkeletonVAE).
train/: Training scripts.
datasets/: Dataset building helpers.
out/: Output directory for generated data and logs (ignored by git).

Setup

This project uses uv for dependency management.

Install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

Install dependencies:
```
uv sync
```

Usage

1. Prepare Data

Ensure your dataset is available. For example, link your data folder:

ln -s ~/Documents/data/ICCAD2019/img ./ICCAD2019/img

(Note: The workspace already contains a link to ICCAD2019 if configured).

2. Run the Pipeline

You can run individual scripts using uv run.

Extract Skeleton:

uv run scripts/skeleton_extract.py path/to/image.png out/result_dir --denoise 3

Vectorize Skeleton:

uv run scripts/vectorize_skeleton.py out/result_dir/image_sk.png out/result_dir/image_vec.json

Expand (Reconstruct) Image:

uv run scripts/expand_skeleton.py out/result_dir/image_vec.json out/result_dir/image_recon.png

3. Build Dataset

Batch process a folder of images:

uv run datasets/build_skeleton_dataset.py ICCAD2019/img out/dataset_processed

4. Train Model

Train the VAE on the processed skeletons:

uv run train/train_skeleton_vae.py out/dataset_processed --epochs 20 --batch 16

Debugging

A debug script is provided to run the full pipeline on a few random images from the dataset:

uv run debug_pipeline.py

Results will be in out/debug.