README.md

Codes for the experiments in "Subsampling Generative Adversarial Networks: Density Ratio Estimation in Feature Space with Softplus Loss"

If you use this code, please cite

@ARTICLE{Ding_2020,  
    author={Ding, Xin and Wang, Z. Jane and Welch, William J.},
    journal={IEEE Transactions on Signal Processing},  
    title={Subsampling Generative Adversarial Networks: Density Ratio Estimation in Feature Space With {Softplus} Loss},   
    year={2020},  
    volume={68},  
    number={},  
    pages={1910-1922}
}

1. To do list

• The 25 2-D Gaussians Simulation
• The Reduced MNIST dataset
• The CIFAR-10 dataset
• The CelebA dataset
• The STL-10 dataset

2. Sample Usage

2.1 Simulation

# DRE-SP+RS with lambda=0
python3 main.py --NSIM 3 --DRE DRE_SP --Sampling RS --epoch_DRE 5 --base_lr_DRE 1e-3 --lambda_DRE 0 --DR_final_ActFn ReLU --batch_size_DRE 512 --epoch_gan 5

2.2 CIFAR-10

# pre-train a ResNet34 for feature extraction
python3 PreTrainCNN.py --CNN ResNet34 --isometric_map --transform

# pre-train an InceptionV3 for evaluation
python3 PreTrainCNN.py --CNN InceptionV3 --transform

# DCGAN, DRE-F-SP+RS with lambda=0.01
python3 main.py --GAN DCGAN --DRE DRE_F_SP --Sampling RS --lambda_DRE 0.01 --samp_nfake 50000 --DR_final_ActFn ReLU --epoch_DRE 200 --base_lr_DRE 1e-4

2.3 Reduced MNIST

First, run 'data_processing.py' to reduce the size the training set to 5000 and store data in a h5 file.

Then,

# Pre-train ResNet34 for feature extraction
python3 PreTrainCNN.py --CNN ResNet34 --N_TRAIN 5000 --isometric_map --transform --batch_size_train 512 --base_lr 0.01 --epochs 200


# pre-train InceptionV3 for computing IS and FID
python3 PreTrainCNN.py --CNN InceptionV3 --transform --batch_size_train 32

# DCGAN: DRE-F-SP+RS with lambda=0.01
python3 main_unsupervised.py --N_TRAIN 5000 --GAN DCGAN --epoch_gan 500 --DRE DRE_F_SP --Sampling RS --lambda_DRE 0.01 --samp_nfake 50000 --DR_final_ActFn ReLU --epoch_DRE 500 --epoch_pretrainCNN 200 --base_lr_DRE 1e-4  --comp_ISFID --samp_round 3

2.4 CelebA (64x64)

First, download (http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) and uncompress the CelebA dataset. Put the folder 'img_align_celeba' under './data'.

Second, run './data/preprocess_celeba.py' to prepare the data. A h5 file will be generated. Images are resized to 64x64 instead of cropping.

Third,

# Pre-train a ResNet34 for feature extraction
python3 PreTrainCNN.py --CNN ResNet34 --isometric_map --epochs 100 --batch_size_train 256 --batch_size_test 128 --base_lr 1e-3 --num_classes 6 --transform

# SNGAN: DRE-F-SP+RS with lambda=0.01
python3 main.py --GAN SNGAN --epoch_gan 100 --DRE DRE_F_SP --Sampling RS --PreCNN_DR ResNet34 --epoch_pretrainCNN 100 --DR_Net MLP5 --lambda_DRE 0.01 --samp_nfake 50000 --DR_final_ActFn ReLU --epoch_DRE 100 --base_lr_DRE 1e-4 --comp_ISFID --samp_batch_size 1000 --samp_round 1

2.5 STL-10 (64x64)

# fine tune the fully connected layers attached to ResNet34
python3 PreTrainFC.py --CNN ResNet34 --epochs 50 --batch_size_train 64 --batch_size_test 64 --base_lr 0.01 --transform --resume_epoch 0

# SNGAN: DRE-F-SP+RS with lambda=0.01
python3 main.py --GAN SNGAN --epoch_gan 500 --DRE DRE_F_SP --Sampling RS --PreCNN_DR ResNet34 --epoch_fc 50 --DR_Net MLP5 --lambda_DRE 0.01 --samp_nfake 50000 --epoch_DRE $EPOCH_DRE --base_lr_DRE 1e-4 --KS_test --comp_ISFID --samp_batch_size 1000

3. Some Results

3.1 25 2-D Gaussians

3.2 CIFAR10

Inception Score versus epoch of DCGAN training. The grey line stands for epoch 60.

Frechet Inception Distance versus epoch of DCGAN training. The grey line stands for epoch 60.

3.3 Reduced MNIST

3.4 CelebA