How good is my GAN?

This is code release for our paper "How good is my GAN?" published on ECCV 2018.

Requirements

Code is written for Python 3.5 and TensorFlow 1.8 (might require minor modifications for more recent versions). You are also expected to have normal scientific stack installed: NumPy, SciPy, Matplotlib. Among auxillary packages install easydict

Datasets

To launch experiments on CIFAR10/100 you should download and convert it first. Run the following command (it will download both CIFAR10 and CIFAR100):

python data/download_and_convert_cifar10.py

Then you need to download Inception and compute Inception statistics:

python evaluation/inception.py --dataset=cifar10

With ImageNet it is a bit more complicated, you have to register to download images and then reconvert them to tfrecord first if you want fast loading (related functions are collected in data/imagenet.py), but I don't remember all details now. Alternatively, you can use a function data/imagenet.py:load_dataset() instead of converting to tfrecord (it reads individual files directly), but it will still require minor modifications of input pipeline.

Training

First you need to train a vanilla classifier (replace cifar10 by cifar100 if needed):

python classifier.py --run_name=cifar10_classifier_ms_decay --dataset=cifar10 --image_size=32

To train SNGAN and evaluate it, run the following command:

python gan.py --action=train_gan,generate,train_all_classifiers --dataset=cifar10 --run_name=sngan --sum_pooling --gen_depth=256 --gen_linear_dim=1024 --gan_loss=hinge --spectral_normalization --conditional_bn --projection --acgan_dw=0 --acgan_gw=0 --gradient_penalty=0 --learning_rate=2e-4 --adam_beta1=0.0 --adam_beta2=0.9 --num_discriminator_steps=5 --arch=resnet --lr_decay --bn_decay=0.9 --batch_size=64

To train WGAN-GP and evaluate it, run the following command:

python gan.py --action=train_gan,generate,train_all_classifiers --dataset=cifar10 --run_name=wgangp --gan_loss=wasserstein --acgan_dw=1.0 --acgan_gw=0.1 --gradient_penalty=10 --learning_rate=2e-4 --adam_beta1=0.0 --adam_beta2=0.9 --num_discriminator_steps=5 --arch=resnet --lr_decay

To train DCGAN and evaluate it, run the following command:

python gan.py --action=train_gan,generate,train_all_classifiers --dataset=cifar10 --run_name=dcgan ---gan_loss=classical --acgan_dw=1.0 --acgan_gw=0.1 --gradient_penalty=0 --learning_rate=2e-4 --adam_beta1=0.5 --adam_beta2=0.999 --num_discriminator_steps=1 --arch=dcgan --batch_size=100 --max_iterations=50000 --noise_dim=100