Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks(DCGAN)

---

Introduction

• The code is adapted from the PyTorch documentation examples:
  https://github.com/pytorch/examples/tree/master/dcgan
  

• The file download_lsun.py comes from a nice repository for downloading LSUN dataset:
  https://github.com/fyu/lsun
  

• I have added massive comments for the code. Hope it beneficial for understanding GANs/DCGANs, especially for a beginner.

Environment & Requirements

• CentOS Linux release 7.2.1511 (Core)
  
• python 3.6.5
  
• pytorch 1.0.0
  
• torchvision
  
• argparse
  
• os
  
• random
  
• subprocess
  
• urllib

Usage

Train DCGAN with MNIST:

python3 main.py --dataset mnist --cuda

Two folders will be created, i.e., ./data & ./results. The ./data folder stores dataset.
The ./results folder stores the generated images and the trained models.
You can also use cifar10, lsun, imagenet, randomly generated fake data, etc.

Download lsun dataset:

python3 download_lsun.py --category bedroom 

Download data for bedroom and save it to ./data.
By replacing the option of --category, you can download data of each category in LSUN as well.
python3 download_lsun.py
Download the whole data set.

NOTE

• The DCGAN architecture is a relatively primary version. Now there exists some new modifications.
  
• The batch_size, size of feature maps of both G and D are all set to 64, different from that in the paper (128).
  
• With above hyperparameters set to 128, the model confronts gradient vanishing. Hope someone help me with it.

References

1. PyTorch documentation
2. https://github.com/pytorch/examples/tree/master/dcgan
   
3. https://github.com/fyu/lsun