Netlistify

Netlistify is a modular deep learning framework for converting analog and mixed-signal (AMS) circuit schematics into HSPICE-compatible netlists. It integrates YOLOv8 for component detection, ResNet for orientation classification, and a modified DETR Transformer for precise wire detection and connectivity analysis. With provided preprocessing scripts, synthetic datasets, and pre-trained models, Netlistify enables accurate, end-to-end schematic interpretation and netlist reconstruction out-of-the-box.

Publications

• Chun-Yen Huang, Hsuan-I Chen, Hao-Wen Ho, Pei-Hsin Kang, Mark Po-Hung Lin, Wen-Hao Liu, Haoxing Ren, “Netlistify: Transforming Circuit Schematics into Netlists with Deep Learning“, ACM/IEEE International Symposium on Machine Learning for CAD (MLCAD), Santa Cruz, CA, September 8–10, 2025

Prerequisites

• Python 3.10.14: [Download and install Python 3.10.14] for your operating system.

Installation

1. Navigate to the project directory:

2. Install the project dependencies:

   pip install -r requirements.txt

Usage

To run the main.py script to train, execute the following command:

python main.py

Adjusting Training Parameters

You can modify training parameters (e.g., learning rate, batch size, number of epochs) in main_config.py before running the script.

Output file

The weights of the model will be saved to "runs/FormalDataset"

Inference

1. Select the trained weights in main_config.py: Open main_config.py and set the path to the trained model weights.

2. Refer to testing.py for an example implementation of how to use the trained model for line detection. You can follow the function calls in testing.py to integrate inference.

⭐ Inference Out-of-the-Box (using pretrained model)

If you want to run inference out-of-the-box without training:

1. Download the pretrained model from the provided Pretrained Model.

2. Place the weights into the directory specified in main_config.py.

3. Run:

   python inference.py
   

This will execute the full inference pipeline using the pretrained model.

Dataset

Our dataset contains 100,000 schematic images, designed to train and evaluate deep learning models for schematic interpretation. The dataset used in this project is available on Hugging Face:Schematic images.

Note: The dataset is provided in .zip format. Users need to manually download and extract it before use.

Acknowledgments

The authors would like to express their sincere gratitude to NVIDIA, Inc., and to the National Science and Technology Council (NSTC) of Taiwan, under Grant No. NSTC 111-2221-E-A49-137-MY3, for their generous support of this research.

Contributors

Name	Affiliation	profile
Chun-Yen Huang	NYCU
Hsuan-I Chen	NYCU
Hao-Wen Ho	NYCU
Pei-Hsin Kang	NYCU
Mark Po-Hung Lin	NYCU
Wen-Hao Liu	NVIDIA Research	link
Haoyu Yang	NVIDIA Research	link
Haoxing Ren	NVIDIA Research	link

Contact

For technical questions, please contact hank2397.ai11@nycu.edu.tw or deathate.ai12@nycu.edu.tw