Class-conditioned ODI generator (PyTorch)

Simple PyTorch implementation of class-conditioned ODI generator based on pix2pix(cGAN).

Paper(arXiv): Omni-Directional Image Generation from Single Snapshot Image by Keisuke Okubo and Takao Yamanaka.

Outline of ODI generation system with class-conditioned generator

Requirements

• numpy
• Torch
• Pillow

Getting started

Installation

• Clone or download this repository.
• Download sample dataset(Google Drive).
• Unzip the zip file and place the dataset folder in the same location as the implementation folder.

class-conditioned-ODI-generator-pytorch/
  +- implementation/
  |    +- train.py
  |    +- networks.py
  |    +- ...
  +- dataset/
       +- train/
       |    +- base/
       |    +- label/
       +- test/
            +- base/
            +- label/

• Move to the implementation directory: cd implementation

Run

• For GPU, run the following command

python train.py --cuda

• For CPU, run the following command

python train.py

Use padding method

• Run with this argument: --padding

Optional arguments

# of epoch: --niter <int> + --niter_decay <int>

• --niter <int>: # of iter at starting learning rate. Default: 100

• --niter_decay <int>: # of iter to linearly decay learning rate to zero. Default: 0

• --g_ch <int>: Generator channels in first conv layer. Default: 128

• --d_ch <int>: Discriminator channels in first conv layer. Default: 128

• --save_interval <int>: Interval epoch of network weight saving. Default: 10

• --graph_save_while_training: If save current loss graph while training. Default: False

Acknowledgments

We thank to these developers:

• Implementation: pix2pix-pytorch by mrzhu-cool.
• Sample dataset: Using part of the SUN360 dataset by Jianxiong Xiao, Krista A. Ehinger, Aude Oliva, and Antonio Torralba (Massachusetts Institute of Technology).