Unbiased Auxiliary Classifier GANs with MINE

This repo contains code for the CVPR 2020 workshop paper Unbiased Auxiliary Classifier GANs with MINE. [pdf] [arXiv]

The code is largely based on the PyTorch implementation of AC-GAN

To reproduce results in the paper:

CIFAR10

To run TAC-GAN (reimplemented):

python main.py --dataset cifar10 --dataroot datasets/cifar10 --cuda --nz 100 --ny 10 --num_classes 10 --batchSize 256 --samplerBatchSize 256 --niter 200 --outf results/cifar10/tac --loss_type tac --label_rotation --visualize_class_label 0 --netD_model snres32 --manualSeed 0 --use_onehot_embed

To run UAC-GAN (with MINE):

python mine.py --dataset cifar10 --dataroot datasets/cifar10 --cuda --nz 100 --ny 10 --num_classes 10 --batchSize 256 --samplerBatchSize 256 --niter 200 --no_ac_on_fake --outf results/cifar10/mine --use_shared_T --netD_model proj32 --loss_type mine --label_rotation --visualize_class_label 0 --netT_model proj32 --use_cy --lambda_mi 0.1 --manualSeed 0 --use_onehot_embed

1-D MoG

python mog_1d.py --distance 4 --num_runs 100 --gan_loss bce