Implementation of the paper "Towards Learning Convolutions From Scratch"

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

DISCLAIMER - There is no public implementation of this paper, so the whole code base in this repository was designed from scratch (as the convolutions in the paper... dad joke, I know)

Structure of the project

convolutions_from_scratch/
    |
    |-- src/
    |   |
    |   |-- scripts/
    |   |   |-- run_train.sh
    |   |   |-- run_figure_3.sh
    |   |
    |   |-- models.py
    |   |-- optim.py
    |   |-- train.py
    |   |-- utils.py
    |   |-- results_visualization.ipynb
    |   
    |-- checkpoints/
    |   |-- cifar10/
    |   |-- svhn/
    |   |-- figure_3/
    |
    |-- plots/
    |
    |-- environment.yml
    |
    |-- report.pdf
    |
    |-- README.md

• results_visualization.ipynb
  • pre-compiled jupyter notebook containing the visualization process of our results
  • used to generate all the figures in our report
• optim.py
  • python file containing the implementation of the $\beta$-LASSO optimizer
• scripts/run_train.sh, scripts/run_figure_3.sh
  • bash scripts used for running all the experiments in our project
• models.py
  • python file containing implementations of the architectures used in this project (S-FC, S-LOCAL, ...)
• utils.py
  • python file containing functions that are used in the training pipeline of the project - training, testing, checkpointing, ...
• train.py
  • main script for training; it combines all the necessary functions and performs the whole training pipeline
• checkpoints
  • directory containing the pre-trained models (checkpoints) and metric trackers for all experiments
  • it's generated during training and later used for the visualization process

Running the code

• Ready-to-use notebook for visualizing our results src/results_visualizations.ipynb

• Set up the conda environment (GPU required!)

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

conda env create -f environment.yml
conda activate convs_from_scratch

• Run the training scripts

cd src/scripts
chmod 777 run_train.sh run_figure_3.sh
./run_train.sh
./run_figure_3.sh

Team info

• Team members:
  • Gojko Cutura - gojko.cutura@epfl.ch
  • Soroush Mehdi - soroush.mehdi@epfl.ch
  • Khashayar Najafi - khashayar.najafi@epfl.ch