WGAN

Authors: Tim Whiting & Evan Peterson

An implementation of a Wasserstein Generative Adversarial Network (WGAN) in Julia for Advanced Machine Learning class.

Dependencies

Needed dependencies for running the project locally (run this from the Julia REPL):

import Pkg; Pkg.add.(["Flux", "Images", "ImageMagick", "NNlib", "BSON", "Plots", "Juno", "FileIO"])

If running on GPU with your local machine, also include:

import Pkg; Pkg.add.(["CuArrays"])

Leveraging TPU

Here is a link to the JuliaTPU repository which tells how to run compile Julia to run on TPU's and gives instructions how to do it on google's colab: https://github.com/JuliaTPU/XLA.jl

For smallNORB

Get the datasets from here: https://cs.nyu.edu/~ylclab/data/norb-v1.0-small/ Clone this python repository that converts them all to images: https://github.com/ndrplz/small_norb Follow the instructions in the repository to convert all of the files to images

TODOs

• Run for a long time, see if works
• Implement Experiment(s), Options:
  • Bidirectional latent encoder
  • Coordinate convolutions
  • Validate on a hold-out set from the same distribution as the train set
  • Anything else cool!