nashory / gans-collection.torch

Torch implementation of various types of GAN (e.g. DCGAN, ALI, Context-encoder, DiscoGAN, CycleGAN, EBGAN, LSGAN)

gans-collection.torch

Torch implementation of various types of GANs (e.g. DCGAN, ALI, Context-encoder, DiscoGAN, CycleGAN, EBGAN). Note that EBGAN and BEGAN implementation is still not stable yet. I am working on this.

Contents

• DCGAN
• Context-encoder
• ALI
• DiscoGAN
• CycleGAN
• Energy-Based GAN (EBGAN)
• LSGAN

Prerequisites

• Torch7
• python2.7
• cuda
• other torch packages (display, hdf5, image ...)

Usage

1. download training data:

python download.py --datasets <dataset>
(e.g) python run.py --datasets celebA

---------------------------------------
The training data folder should look like : 
<train_data_root>
                |--classA
                        |--image1A
                        |--image2B ...
                |--classB
                        |--image1B
                        |--image2B ...
---------------------------------------

2. run GANs training:
   Note that you need to change parameter options in "script/opts.lua" for each GANs.

python run.py --type <gan_type>
(e.g) python run.py --type dcgan

Display GUI : How to see generated images in real-time?

step by step instruction:

1. set server-related options(ip, port, etc.) in "script.opts.lua"
2. run server (python server.py --type <gan_type>)
3. open web browser, and connect. (https://<server_ip>:<server_port>)

you will see like this:

Results

training	Final

Acknowledgement

• brought dataloader code from (DCGAN)
• referenced the code from (Context-encoder)

Author

MinchulShin, @nashory
Will keep updating other types of GANs.
Any insane bug reports or questions are welcome. (min.stellastra[at]gmail.com) :-)

gans-collection.torch's Issues

instability when training ALI.

It seems the network learns to ignore the encoded noise z_hat at the end of the training.
Need to configure.

differences between cyclegan and discogan

(https://www.quora.com/What-is-the-difference-between-CycleGAN-and-DiscoGAN-They-both-seem-to-be-the-same-thing)

• CycleGAN has a single cycle-consistency loss. Also, when comparing F(G(y)) with y, it seems that the two papers use different forms of distance measures (MSE, hinge-loss).
• CycleGan has an addition hyperparameter to adjust the contribution of reconstruction/cycle-consistency loss in the overall loss function.
• the generator structures used in DiscoGan and CycleGan are somewhat different.