GANs with spectral normalization and projection discriminator

This is an unofficial PyTorch implementation of sngan_projection

Miyato, Takeru, and Masanori Koyama. "cGANs with projection discriminator." arXiv preprint arXiv:1802.05637 (2018).

Dependencies:

• PyTorch1.0
• numpy
• scipy
• tensorboardX
• tqdm
• torchviz pip install torchviz and graphviz sudo apt-get install graphviz

Usage:

There are two ways to run the training script:

• Run the script directly (We recommend this way): python3 main.py or python main.py. In this way, the training parameters can be modified by modifying the parameter.py parameter defaults.

Parameters

Parameters	Function
--version	Experiment name
--train	Set the model stage, Ture---training stage; False---testing stage
--experiment_description	Descriptive text for this experiment
--total_step	Totally training step
--batch_size	Batch size
--g_lr	Learning rate of generator
--d_lr	Learning rate of discriminator
--parallel	Enable the parallel training
--dataset	Set the dataset name,lsun,celeb,cifar10
--cuda	Set GPU device number
--image_path	The root dir to training dataset
--FID_mean_cov	The root dir to dataset moments npz file

Results

We have reproduced the FID (in Cifar-10, best result is FID=17.2) result reported in the paper.

The convergence curve of FID is as follows:

CIFAR10 results

200K:

500K:

600K:

800K:

1000K:

Acknowledgement

• sngan_projection
• BigGAN-PyTorch
• pytorch.sngan_projection