Skip to content

Latest commit

 

History

History
65 lines (56 loc) · 3.38 KB

README.md

File metadata and controls

65 lines (56 loc) · 3.38 KB
Raw

ReG-NAS

1 Required Environment

cuda version: 11.1
python: 3.7
torch: 1.11.0
pyg (for conda) /torch_geometric (for pip): 2.0.4
tensorboardX: 2.8.0

To check whether your environment match the requirement, please run

bash run_single.sh

If no errors occurs, then you can use ReG-NAS. ReG-NAS itself contains torch_geometric and torch.nn for we modified some files in it. However, you still need to install torch_geometric and torch in your environment

2. How to use

2.1 To run a single experiments

RegNAS offers several training pipelines for experiments. As listed below:

main.py: Groundtruth training pipeline for ogbg-molhiv or other classification based groundturth task.
main_ZINC.py: Groundtruth training pipeline for ZINC or other regression based groundtruth task.
main_proxy_le.py: Proxy training pipeline using Laplacian Matrix's eigenvectors as proxy task.
main_proxy_pm.py: Proxy training pipeline using Poorest Model generated vectors as proxy task.
main_proxy_rm.py: Proxy training pipeline using Randomly-Selected Model generated vectors as proxy task. (also applicable for GM based pipeline)
main_proxy_rv.py: Proxy training pipeline using Random generated vectors as proxy task

For example, if we want to run main.py for 3 times with the configuration your_configs.yaml (stored in configs/pyg), simply use

python main.py --cfg configs/your_configs.yaml --repeat 3 # graph classification

The result will be stored in results/, where you can analyze the training results

2.2 To run experiments in a search space

First you can define your search space in grids/, in which we offered some example files in it.

After defined you search space, the shell files listed below are designed for managing experiments in a search space:

run_batch.sh: Classification-based groundtruth training
run_batch_ZINC.sh Regression-based groundtruth training
run_batch_proxy.sh: Proxy training

In these file, you can define your training configurations (hyper-parameter settings) in CONFIG, specify your search space file in GRID, define training pipeline in MAIN, and define max parallel jobs in ``MAX_JOBS''. Then simply use

bash run_batch.sh

to run the experiment.

2.3 Results analysis

All the functions for result analysis is in proxy_groundtruth_analysis.py, which contains:

proxy_groundtruth_analysis: Analyzes proxy ranking and classification based groundtruth ranking. Computes rho and tau, draws scatter figure.
proxy_groundtruth_analysis_ZINC: Analyzes proxy ranking and  regression-based grdountruth ranking. Computes rho and tau, draws scatter figure.
proxy_groundtruth_analysis_epoch: Analyzes proxy ranking and classification based groundtruth ranking at each epoch. Computes rho and tau, draws scatter figure.
proxy_groundtruth_analysis_ZINC: Analyzes proxy ranking and regression-based grdountruth ranking at each epoch. Computes rho and tau, draws scatter figure.
similarity_analysis: Analyzes Ranking Stability for groundtruth ranking or proxy ranking. Computes rho and tau, draws scatter figure.
similarity_analysis_epoch: Analyzes Ranking Stability at each epoch. Computes rho and tau, draws scatter figure.

Simply add these function below if __name__ == '__main__': when you want to analyze experiment results. Before using, please change the file path contained in the function.