Skip to content

A set of scripts and experiments making it easier to analyze deep learning empirically.

Notifications You must be signed in to change notification settings

csinva/dnn-experiments

Repository files navigation

understanding how deep learning works

this repo contains code for running a variety of different experiments attempting to understand deep learning via empirical experiments

organization

  • each folder contains a readme with code documentation, as well as comments in the code
  • the vision_fit and vision_analyze folders detail a number of experiments on multilayer perceptrons and convolutional neural networks using various datasets including MNIST, CIFAR, and custom datasets
  • the sparse_coding folder contains code for running and analyzing sparse coding on different sets of images
  • the mog folder contain code examples for fitting synthetic datasets generated as mixtures of Gaussians
  • the poly_fit folder contains code for fitting simple 1D polynomials
  • the scripts folder contains scripts for launching jobs on a slurm cluster
  • the eda folder contains minimum working examples for simple setups with various pytorch and scikit-learn functions

requirements

  • the code is all tested in python3 and pytorch 1.0

running

  • the scripts folder contains sample slurm scripts for launching jobs on a cluster
  • most of the experiments are time-consuming and should be parallelized over many machines
  • to do so, ssh into one of the scf nodes (e.g. legolas) and run module load python
  • set the parameters you want to sweep as lists in one of the submit*.py files
  • then run this file and it will automatically launch slurm jobs for each set of parameters