CktGNN

CktGNN is a two-level GNN model with a pre-designed subgraph basis for the analog circuit (DAG) encoding. CktGNN simultaneously optimizes circuit topology and device features, achieving state-of-art performance in analog circuit optimization.
For more information, please check our ICLR 2023 paper: 'CktGNN: Circuit Graph Neural Network for Electronic Design Automation' (https://openreview.net/forum?id=NE2911Kq1sp).

OCB: Open Circuit Benchmark

• OCB is the first open benchmark dataset for analog circuits (i.e. operational amplifiers (Op-Amps)), equipped with circuit generation code, evaluation code, and transformation code that converts simulation results to different graph datasets (i.e. igraph, ptgraph, tensor.) Currently, OCB collects two datasets, Ckt-Bench-101 and Ckt-Bench-301 for the general-prpose graph learning tasks on circuits.

• Ckt-Bench-101 (directory: /OCB/CktBench101): Ckt-Bench-101 is generated based on the dataset used in our ICLR 2022 paper. Ckt-Bench-101 contains 10000 different circuits, and it eliminates invalid circuits in the datasets used in our ICLR 2023 paper and replace them with new valid simulations.

• Ckt-Bench-301 (directory: /OCB/CktBench301): Ckt-Bench-301 contains 50000 circuits. Circuits in Ckt-Bench-301 and in Ckt-Bench-101 have different device features and circuit topologies. This benchmark dataset is proposed to perform the Bayesian optimization, as it might be hard to implement simulation code on the circuit simulators without relevant expertise.

• Source code (directory: /OCB/src): The source codes enable users to construct their own analog circuit datasets of arbitrary size. The directory /OCB/src provides simulation codes for circuit simulators. /OCB/src/circuit_generation.py generates circuits and writes them in .txt file. /OCB/src/utils_src.py includes functions (train_test_generator_topo_simple) that convert the txt circuits and relevant simulation results to igraph data.

• Tutorial (directory: /OCB): The tutorial /OCB/Tutorial.pdf provides guidance of understanding the source code and implementing the performance simulation on circuit simulators.

Environment

• Tested with Python 3.7, PyTorch 1.8.0, and PyTorch Geometric 1.7.2
• Other required packages are listed in requirements.txt
• The experiment codes are basically constructed upon D-VAE.

Experiments on Ckt-Bench-101

• Run variatioal auto-encoders on Ckt-Bench-101 to train different circuit/DAG encoders (CktGNN, PACE, DVAE, DAGNN...), and test the decoding ability of the decoder in the VAE (proportion of valid DAGs, valid circuits, novel DAGs generated from the latent encoding space of the circuit encoder, proportion of accurately reconstructed DAGs.)

python main.py --epochs 300 --save-appendix _cktgnn --model CktGNN --hs 301

User can select different models (e.g. DVAE, DAGNN ..) and uses the corresponding save appendix (e.g. --save-appendix _dvae, --save-appendix _dagnn ) to store the results.

• Run SGP regression to test whether the circuit encoder can generate a smooth latent space w.r.t. circuit properties. There are many circuit properties: FoM, Gain, Bw, Pm, while FoM is the most critical one. In circuit optimization, FoM characterizes the circuit quality and the objective is to maximize circuit's FoM.

python sgp_regression.py --checkpoint 300 --save-appendix _cktgnn --model CktGNN --hs 301

Make sure that --save-appendix --model --hs are consistent with previous experiment, and set --checkpoint equals to the --epoch.

Experiments on Ckt-Bench-301

• Two bayesian optimization methods are provided in directory /search_methods: DNGO and Bohamiann. These codes are based on pybnn

• Implement Bayesian optimization on Ckt-Bench-301 to find the optimal circuits with the higest FoM.

First, run python inference.py --checkpoint 300 --save-appendix _cktgnn --model CktGNN --hs 301 to generate the embeddings in the latent space (by the selected circuit encoder).

Second, run python bo_search.py --save-appendix _cktgnn --model CktGNN --search-method bo to search the optimal circuit in Ckt-Bench-301. Expctation improvement searching is used, and users can use --search-method to choose the Byesian optimization model (bo = Bohamiann, dngo = DNGO).

Please leave an issue if you have any trouble running the code or suggestions for improvements.

Reference

If you find OCB and the source code are useful, please cite our paper.

@article{dong2023cktgnn,
title={CktGNN: Circuit Graph Neural Network for Electronic Design Automation},
author={Dong, Zehao and Cao, Weidong and Zhang, Muhan and Tao, Dacheng and Chen, Yixin and Zhang, Xuan},
journal={arXiv preprint arXiv:2308.16406},
year={2023}
}

Zehao Dong, Washington University in St. Louis

[email protected]

Aug 31 2023