Skip to content

Latest commit

 

History

History
95 lines (60 loc) · 1.77 KB

README.md

File metadata and controls

95 lines (60 loc) · 1.77 KB
Raw

GAN.pth

a pytorch gan framework

Features

  • As GAN develop so quickly, keeping up to date matters

  • Providing GAN specific features as possible i.e modular,but keep Flexible as possible

    • AdaIn, Spectral normalization, skip connections, resnets
    • Out-of-box Models
    • Trainer:train_generator, train_discriminator
    • GAN Loss
    • Metrics
    • Demo

  • Documentation

  • Production oriented, Research Friendly

    • distributed training(multiple gpus and multiple machines)
    • easy to deploy on production(performance optimized for mobile platform and server side)

  • Custom Datasets & Pretrained Models provided in GANHub

  • Easy to extension For Example

    • do config through args file for each model
    • train_gene
    • well structured code(object oriented programming)

  • well tested, as close as possible to official sota effect

Inspired by following frameworks

torchgan

PytorchGANZoo

PyTorch-GAN

StarGAN_v2-Tensorflow

Classical Models

StyleGAN/StyleGAN2

StarGAN/StarGAN2

FUNIT

Components

Discriminator

progressive growing (pggan) self-attention

Generator

style-based generator(stylegan)

GAN Loss

Non-Saturating Loss + R1/R2 WGAN+GP

Metrics

IS(Inception Score) FID(Inception Distance) LPIPS PPL

Datasets

face aligner

Utils

video interpolation reporter: tensorboard reporter

Demos

use GAN.pth framework to develop following apps as demos

Demo GANS

  • Vanilla GAN
  • DCGAN

Faceswap

AI Stylist

AI Portraits

Please access [GANHub]https://github.com/habout632/GANHub) for more demos, datasets, pretrained networks.

Documentation built with readthedocs