This application is designed to be used with python > 3.5 on linux.
Pip requeriements are available in requirements.txt at the root of this repository.
This file can be used as it follows :
pip3 install -r requirements.txt
Additionnal libraries may be needed.
The dataset must be a folder containing 3 folders ['test','train','val'] containg images (jpg).
Dataset1 architecture:
├── dataset1/
│ ├── train/
| ├── test/
│ └── val/
The dataset folder must be present into the 'data' folder. Its name is latter used to start a training on it.
Using the folowing command line will start a traning over the dataset: 'dataset1'
python pix2pix.py --dataset_name 'dataset1'
Other parameters can be changed, such as : n_epochs, batch_size, lr, among others (see pix2pix.py for further information on parameters).
Dataset can be generated using node js scripts like 'vertical.js' in data folder.
To generate a dataset, use the following command line.
node vertical.js
Some node packages may need to be installed as described in the following.
npm install d3
npm install d3-gridding
npm install jsdom
The outcome will be in the output folder.
Parameters, such as the dataset size can be changed directly in the script.
Once the SVGs are generated, they can be transformed into JPGs (format that the GAN uses), using 'svgtojpg.sh' as it follows :
./svgtojpg.sh
Some node packages may need to be installed as described in the following.
brew install ImageMagick
The outcome will be in the output folder.
Finally, labels and inputs are concatened and formatted as the model expects using the python script 'preprocess.py'
python preprocess.py
The outcome will be in the 'res' folder.