This repository contains some of the results files from running various graphgym experiments on spice netlists.
Each experiment contains a few different numbers - one for each execution of the experiment w/ a fixed random seed. Within an individual execution, there are statistics collected on the training data and validation data (saved under train/ and val/) as well as model checkpoints in ckpt. The experiment configuration is available under config.yaml.
If we would like to view the final training/validation accuracy, we can simply look at the last line of the stats.json files in the respective files. With jq, this looks like this:
tail -n 1 train/stats.json | jq '.accuracy'
First, clone and install our fork of graphgym. Then you can run experiments by running the following in the run/ directory of the repository:
python main.py --cfg configs/omitted_classification/ctrl_embeddings.yaml --repeat 3
In this case, ctrl_embeddings.yaml is the experiment configuration and it is repeated with 3 different random seeds.