This repository contains implementation of WGAN and DCGAN in Julia using Knet. Here is a detailed report about WGAN.
Just run julia train.jl with the following arguments. Note that if procedure is gan, discriminator weights won't get clipped.
usage: train.jl [--gpu GPU] [--dn DN] [--type TYPE] [--data DATA]
[--procedure PROCEDURE] [--zsize ZSIZE]
[--epochs EPOCHS] [--report REPORT]
[--batchsize BATCHSIZE] [--lr LR] [--clip CLIP]
[--opt OPT] [--leak LEAK] [--out OUT] [-h]
WGAN Implementation in Knet
optional arguments:
--gpu GPU GPU ID if -1 don't use GPU (type: Int64,
default: 0)
--dn DN Train discriminator n times (type: Int64,
default: 1)
--type TYPE Type of model one of: [dcganbn (regular
DCGAN), mlpg (Generator is MLP), mlpgd
(Both MLP), dcgan (Generator has no BN and has
constant filter size)] (default: "dcganbn")
--data DATA Dataset dir (lmdb) (default: "/home/cem/lsun")
--procedure PROCEDURE
Training procedure. gan or wgan (default:
"gan")
--zsize ZSIZE Noise vector dimension (type: Int64, default:
100)
--epochs EPOCHS Number of training epochs (type: Int64,
default: 20)
--report REPORT Report loss in n iterations (type: Int64,
default: 500)
--batchsize BATCHSIZE
Minibatch Size (type: Int64, default: 64)
--lr LR Learning rate (type: Float64, default: 0.0002)
--clip CLIP Clip value (type: Float64)
--opt OPT Optimizer, one of: [adam, rmsprop] (default:
"adam")
--leak LEAK LeakyReLU leak. (type: Float64, default: 0.2)
--out OUT Output directory for saving model and
generating images (default: "./models")
-h, --help show this help message and exit
Images below are the plots of the outputs of the implementation in this repo. Original outputs can be found here
