galaxyMoCov3

Overview

In this study, I evaluated the effectiveness of supervised training with MoCov3 for linear probing on Galaxy10 DECals.

How to setup on Oscar

1. Create a conda env using csci2952_mocov3.yml.

module load miniforge/23.11.0-0s
source /oscar/runtime/software/external/miniforge/23.11.0-0/etc/profile.d/conda.sh
mamba env create -f csci2952_mocov3.yml
conda activate csci2952_mocov3

4. Now install pytorch.

interact -q gpu -g 1 -f ampere -m 20g -n 4
module load cudnn cuda
pip install torch torchvision torchaudio

How to run

1. Pretraining

sbatch run_pretrain.sh

2. Posttraining

sbatch run_posttrain.sh

3. Evaluatino

sbatch run_eval.sh

Note: Edit the bash files to change training parameters as needed.

Results

The results are stored in logs/800.

Acknowledgement

I use the following sites/tools for the evaluation.

• Original MAE implementation
  https://github.com/facebookresearch/moco-v3
  https://github.com/facebookresearch/mae

• ChatGPT

  • How to use a custom dataset
  • How to use args in a bash file
  • Debugging