pytorch-generative-model-collections

2019-09-11 |

196 |

0 |

pytorch-generative-model-collections

pytorch-generative-model-collections

Original : [Tensorflow version]

Pytorch implementation of various GANs.

This repository was re-implemented with reference to tensorflow-generative-model-collections by Hwalsuk Lee

I tried to implement this repository as much as possible with tensorflow-generative-model-collections, But some models are a little different.

This repository is included code for CPU mode Pytorch, but i did not test. I tested only in GPU mode Pytorch.

Dataset

GAN, LSGAN, WGAN, WGAN_GP, DRAGAN, EBGAN, BEGAN used dataset of the torchvision.
infoGAN, ACGAN, CGAN used dataset downloaded from MNIST (http://yann.lecun.com/exdb/mnist/) and fashion-MNIST (https://github.com/zalandoresearch/fashion-mnist).

Difference from tensorflow-generative-model-collections

This repository does not have Variational Auto-Encoders.
DRAGAN : The DRAGAN loss calculation method is different.
ACGAN : The learning procedure is different.
EBGAN : MSE is used instead of L2 norm and the last layer of discriminator does not use sigmoid activation function.
BEGAN : L1 norm is used instead of L2 norm and the last layer of discriminator does not use sigmoid activation function.

Generative Adversarial Networks (GANs)

Lists (Table is borrowed from tensorflow-generative-model-collections)

图片.png

Variants of GAN structure (Figures are borrowed from tensorflow-generative-model-collections)

Results for mnist

Network architecture of generator and discriminator is the exaclty sames as in infoGAN paper.
For fair comparison of core ideas in all gan variants, all implementations for network architecture are kept same except EBGAN and BEGAN. Small modification is made for EBGAN/BEGAN, since those adopt auto-encoder strucutre for discriminator. But I tried to keep the capacity of discirminator.

The following results can be reproduced with command:

python main.py --dataset mnist --gan_type <TYPE> --epoch 25 --batch_size 64

Fixed generation

All results are generated from the fixed noise vector.

Conditional generation

Each row has the same noise vector and each column has the same label condition.

Name | Epoch 1 | Epoch 10 | Epoch 25 | GIF :---: | :---: | :---: | :---: | :---: | CGAN |

InfoGAN : Manipulating two continous codes

All results have the same noise vector and label condition, but have different continous vector.

Name | Epoch 1 | Epoch 10 | Epoch 25 | GIF :---: | :---: | :---: | :---: | :---: | infoGAN |

Loss plot

Name | Loss :---: | :---: | GAN |

Results for fashion-mnist

Comments on network architecture in mnist are also applied to here.
Fashion-mnist is a recently proposed dataset consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. (T-shirt/top, Trouser, Pullover, Dress, Coat, Sandal, Shirt, Sneaker, Bag, Ankle boot)