DYZhang09 / CycleGAN

CycleGAN Implementation

Simple CycleGAN Implementation

Introduction

This repository is a simplified CycleGAN implementation aiming for image translation such as horse-to-zebra(may has some bugs).

The official implementation is here

file structure

.
├── dataIO              // the code of dataset I/O module 
│   ├── __init__.py
│   ├── datasets.py
│   ├── preprocess.py
│   └── utils.py
├── model               // the code of definition of CycleGAN model
│   ├── CycleGAN.py
│   ├── __init__.py
│   ├── basic_blocks.py
│   ├── networks.py
│   └── utils.py
├── weights             // the pretrained weight files
├── test.py             // the code for testing
└── train.py            // the code for training

how to train

use python train.py --dataroot ./data/horse2zebra [OPTIONS]

some of the options are listed below, you can use python train.py -h for more information.

    --dataroot      the root path of the dataset, must have 4 subfolders: trainA, trainB, testA, testB
    --weight_file   where to store the temporary weights
    --res_layers    the num of Resnet blocks in CycleGan
    --batch_size    the batch size of input images
    --shuffle       whether to shuffle input images
    --crop_size     the crop size
    --crop_pos      the crop position
    --flip          whether to flip input images horizontally
    --grayscale     whether to read images as gray images
    --resize        the size of images after resizing
    --lr            the learning rate
    --gpu           whether to use gpu to train
    --epochs        the num of training epochs

how to test

use python test.py --dataroot ./your/dataset [OPTIONS] and the result will reside in dataroot/fake/.

most options are same as training options except --lr --epochs, use python test.py -h for more information.