gojkoc54 / augmix

Implementation of the AugMix paper

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Reconstruction and ablation study of the AugMix paper

(Reproducibility Challenge, Fundamentals of Inference and Learning EE-411, EPFL)

Structure of the project

augmix/
    |
    |-- src/
    |   |
    |   |-- scripts/
    |   |   |-- run_params_ablation.sh
    |   |   |-- stage_logs.sh
    |   |
    |   |-- snapshots/
    |   |   |-- {dataset}_{model}_{width}_{depth}
    |   |
    |   |-- main.py
    |   |-- models.py
    |   |-- utils.py
    |   |-- results_visualization.xlsx
    |   
    |-- environment.yml
    |
    |-- report.pdf
    |
    |-- README.md
  • main.py
    • script containing the main pipeline of the project - training, testing, checkpointing, ...
    • DISCLAIMER - most of the code in this script was borrowed from the official implementation of Augmix by Google (https://github.com/google-research/augmix)
  • scripts/run_params_ablation.sh
    • bash script used for running all the experiments in our project
    • runs all desired combinations of AugMix hyperparameters for all the models of interest and datasets
  • models.py
    • python file containing implementations of the architectures used in this project
    • some of the models are inherited from the original papers, and some were added by us in order to test the method more broadly (Wide ResNet and AlexNet)
  • utils.py
    • python file containing functions that are used all over the project (mostly AugMix-related transformations)
  • results_visualization.xlsx
    • Excell file containing the visualization process of our results; used to generate all the figures in our report
  • snapshots
    • directory containing the results (training and testing logs) of all the experiments
    • it's generated during training and later used for the visualization process

Running the code

DISCLAIMER - the good old 'it worked on my machine' ... We hope it works on your as well, but can't guarantee :)

  • Set up the conda environment (GPU required!)
conda env create -f environment.yml
conda activate augmix
  • Run the ablation study bash script
cd src/scripts
chmod 777 run_params_ablation.sh
./run_params_ablation.sh

Team info

About

Implementation of the AugMix paper


Languages

Language:Python 95.3%Language:Shell 4.7%