dgcnz/relaxed-equivariance-dynamics

Code for "Effect of equivariance on training dynamics"

Relaxed equivariance dynamics

Our blogpost has been published at GRaM Workshop @ ICML 2024.

Our extended blogpost can be found here.

Package contributions

To maximize convenience, reproducibility and encourage usage of our modules (models, datasets, tools), we've package some of them separately.

• gconv: (Relaxed) Regular Group Convolutions package
• JHTDB HuggingFace Dataset

Setup

Make virtual environment and install dependencies:

make setup_env

Source your virtual environment:

source .venv/bin/activate

Local Usage

Reproduce the training results from a given experiment:

python -m src.train experiment=wang2024/rgcnn_triple

SLURM Usage

Pre-requisites

• ssh into snellius
• Move into the project root

  cd ~/development/dl2

Train model based on experiment config

Let's say you want to run the experiment at configs/experiment/wang2022/equivariance_test/convnet.yaml. You can make use of the shortcut slurmtrain as follows:

make strain experiment=wang2022/equivariance_test/convnet

If you need to modify anything, the script is at scripts/slurm/train.sh.

Print slurm logs given job id

make slurmcat id=6246500

The logs are stored at scripts/slurm_logs/slurm_output_{id}.out.