/improved_gan_pytorch

Pytorch implementation of semi-supervised DCGAN based on "Improved Techniques for Training GANs"

Primary LanguagePythonMIT LicenseMIT

improved_gan_pytorch

Pytorch implementation of semi-supervised DCGAN based on "Improved Techniques for Training GANs".

Feature matching and semi-supervised GAN have be reimplemented.

So far, other improved techniques haven't been added.

Prerequisites

Pytorch Version: 2.0.3 and Python 2.7

Usage

Run file： python improved_GAN.py

BTW, in my example, my classifer is for CIFAR10 dataset,

and labeled input : unlabeled input : generated fake input = 1 : 1 : 1

Users also can change the settings according to my program's comments.

P.S.

For Generator Loss, it is also equal to -loss_unlabled_fake + loss_feature_matching.

For Labeled Loss, it is also equal to -loss_target + log_sum_exp(before_softmax_labeled_output)

To do

  1. to average input labeled data over 10 classes subset.
  2. to adjust the network structure for high accuracy classification
  3. to reimplement other techniques in improved GAN
  4. to reimplement "Bad GAN" paper

Semi-supervised + Feature matching CIFAR10 Classification

image

image

300th epoch