3D-GANs-pytorch

A PyTorch implementation of 3D GANs

Introduction

In this implementation aims to build general framework for 3D-GANs using Pytorch, I have implemented or will implement several Genertators,Discriminators and losses.

The basics model is SRGAN, which is a super resolution network originally publish in Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

Train and Test

You can used python pretrain.py to pretrain the genertator. This file can also use to do super resolution without GAN.

run python train.py and python test.py
The adjustable parameter is write in ./options
You must specify --img_width --img_height --img_depth --file_extension according to your dataset.

Train your own datasets

The folder structure is below:

3D-GANs-pytorch
├── datasets
├ ├── train
├ ├── test
├── option.py
├── train.py
├── test.py
├── pretrain.py
├── models
├── utils
├── data

Your own datasets need to be put into datasets/train and datasets/test.

Dataset in this implementation using Iseg, nibabel.load(image_file) is used to load the image_file. If your data format is different, you may need to rewrite the CustomDataset in data/customdataset.

NOTE: You may need to rewrite a normalize method according to the mean and variance of your data set.

Completed

Genertators

• √ ResnetGenertator
• × 3DUnet
• × ResUnet

Discriminators

• GAN
• Patch GAN
• Pixel GAN

Losses

• Gradient difference Loss
• ...

Example

High resolution / Low resolution
Genertated with GAN / Genertated with CNN