COMP6258-Reproducibility-Challenge

Attempt at reproducing the paper found in: Openreview link

Minimal code found, but CosNetUnit taken from the appendix of the paper. The rest is our own interpretation of the paper's instructions.

How to run

1. Setup virtualenv

1. python3.11 -m venv .venv
2. source .venv/bin/activate (or .venv/bin/Activate.ps1 on windows)
3. pip install -r requirements.txt

2. Dataset setup (Iridis)

On Iridis, ImageNet ILSVRC2012 can be loaded from the university's shared dataset cache:

module load common-dataset-loader
common-dataset-loader load ImageNet/ImageNetILSVRC2012

This makes the dataset available at ~/hpc-shared-data/ImageNet/ImageNetILSVRC2012. Pass this path as --dataset_root with --dataset iridis100 or --dataset iridis1k.

3. Run training session

1. Decide on model, options can be found by running python3 train.py --help | grep model
2. Decide on dataset, options can be found by running python3 train.py --help | grep dataset
3. Now run the training session on the desired system python3 train.py --model MODEL_CHOICE --dataset DATASET_CHOICE
4. Other options can be adjusted, such as batch size for the dataset, optimiser learning rate and number of epochs for training.