cyclegan

Preface

This repository contains the code to running cyclegan on the 2021 Bass Connections team's data. It is mostly adapted from [https://www.tensorflow.org/tutorials/generative/cyclegan].

Data downloading:

The folder colab-cyclegan-data, containing training images, has been zipped and uploaded to box in case we need it for google colab. To get it, run

!wget -O data.zip https://duke.box.com/shared/static/5yfb0hgw6dphe3p9oexml8vfqncvjo2j -O data.zip ; unzip data.zip ; rm data.zip

in one of the code blocks in the iPython notebook. The exclamation mark in the front tells the Jupyter notebook that this whole thing is a Linux shell command and should be interpreted as such. This should extract and make a folder called colab-cyclegan-data

IMPORTANT: the predict script is not working. We used the ipynb file manually for predicting.

Running cyclegan non-interactively

(Download colab-cyclegan-data through the method above if you haven't already)

conda env create -f cyclegan.yml

conda activate cyclegan

python train_wrapper.py

python predict_wrapper.py cyclegan.ipynb

TODO

1. Make jitter cyclegan images