Skip to content
A torch implementation of "Pixel-Level Domain Transfer"
Branch: master
Clone or download
Latest commit d227256 Nov 1, 2017
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
LICENSE Create LICENSE Nov 1, 2017
README.md add display Nov 27, 2016
data.lua bug fix Nov 27, 2016
epoch7.jpg add display Nov 27, 2016
gan.jpg Add files via upload Nov 26, 2016
main.lua bug fix Nov 27, 2016
prepare_data.ipynb bug fix Nov 26, 2016

README.md

Pixel Level Domain Transfer

A torch implementation of "Pixel-Level Domain Transfer". based on dcgan.torch.

Dataset

The dataset used is "LookBook", from Donggeun Yoo.

Training

To train the model, put the LOOKBOOK dataset under repository, resize images to 64*64. Prepare the dataset using prepare_data.ipynb. Then run

th main.lua

You can tune the parameters, such as number of filters, optimizer, etc.

Example results

Example results on LOOKBOOK dataset(top), left is input, right is generated clothes. Results on a similar dataset (bottom). More results will be added soon.

Results

Monitor the performance

  • Install display package with: luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec
  • Start the server with: th -ldisplay.start
  • Open this URL in your browser: http://localhost:8000

Below shows the results after 7 epochs, each 3*1 block is generated cloth, true cloth, input image. Errors of G, D, and A network will be plotted.

epoch 7

You can’t perform that action at this time.