This repo is an adapted version of the BMVC 2018 paper: SPG-Net by Song et al.

Prepare the data

Download the Cityscapes datasets and the masks from PartialConv. Create a directory data and put the downloaded data under data.

Training

If you want to use the synchronized batchnorm, you can set it up following the steps (credits: SPADE):

cd models/networks/
git clone https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
cp -rf Synchronized-BatchNorm-PyTorch/sync_batchnorm .
cd ../../

Train with the following command:

python main.py --gpu_ids 0 --batch_size 2

and check the results at logs.

This code borrows heavily from SPADE and pix2pixHD. I also adapt the code from EdgeConnect for processing the masks for training. Thanks for their amazing works!