search

gauravmm / cifar-gan

Our start with GANs.
1 stars 0 forks source link

readme

Introduction to GANs

This is our quick-and-dirty introduction to GANs using TensorFlow.The aim of this is to figure out the technical challenges in semi-supervised learning and representation learning using GANs before we apply it to a a larger dataset like ImageNet.

We support both the CIFAR-10 dataset and the LUNA16 + KDSB17 lung cancer datasets.

The Goal

Our goal is to develop a semi-supervised classifier from a lot of data without labels, and a little data with labels.

Installation

This codebase is targeted at Python 3.5

  1. Set up Ubuntu 16.04 (or higher) on a machine.
  2. Follow the installation instructions here to install TensorFlow.
  3. Install the prerequisites with sudo -H pip3 install numpy urllib3.

To get started with development:

  1. git clone git@github.com:gauravmm/cifar-gan.git
  2. git tag -l to list the tags
  3. git checkout tags/<tag_name> -b <branch_name> to create a new branch named <branch-name> starting from tag <tag_name>.
    Remember to name the branch after yourself!

To run the code

  1. ./experiment.sh
    Output will be logged to train_logs/ including loss, accuracy, and generated images; and weights will be saved to weights/.

Coding Standards

We will use git for distributed version control. Feel free to submit pull requests with partial/incomplete work. This is research code, and so there are only two coding standards:

  1. You must only commit to branches that you create. (Read from everyone's branches, only write to your own.)
  2. Any code you use to generate results from must be easily reproducible. This means that:
    1. It should be self-contained in a single commit. (You should indicate that commit hash in the Google spreadsheet we are using to aggregate results.)
    2. We must be able to replicate that on any machine by running experiment.sh.
    3. You must set your random seed explicitly so that all runs are identical.

Bugs & Features

Submit bug reports and feature requests on the issues page. Feel free to submit pull requests for bugs or features. DO NOT MERGE INTO PRODUCTION -- submit a pull request and I will handle the merging.