DCGAN demo

This folder contains an example implementation of DCGAN [1] in MatConvNet. The example trains on the CELEB-A data [2].

Data

First download and extract the aligned face images (img_aligin_celeba.zip) to data/celeba by using the link http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

There are two entry-point scripts:

• dcgan_train.m: trains a new model from scratch.
• dcgan_generate.m: generates images by using the trained model.

Train

To use the training code using a gpu on your system, use something like:

opts.train.gpus = 1 ;
dcgan_train(opts) ;

Generate

% load trained generative network from the last epoch
d = dir(fullfile(opts.expDir,'net-epoch-*.mat'));
load(fullfile(opts.expDir,d(end).name));
netG = dagnn.DagNN.loadobj(netG);
opts.network = netG ; 
opts.gpu = 1 ;
dcgan_generate(opts) ;

References

1. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, Alec Radford, Luke Metz, Soumith Chintala, 2016.

2. Deep learning face attributes in the wild., Liu, Z., Luo, P., Wang, X., & Tang, X. Proceedings of the IEEE International Conference on Computer Vision. 2015.