Chainer implementation of "Toward Multimodal Image-to-Image Translation".
This is a faithful re-implementation of the official PyTorch implementation.
git clone --recursive https://github.com/wkentaro/chainer-bicyclegan.git
cd chainer-bicyclegan
conda install -c menpo -y opencv
pip install .
cd examples/bicyclegan
# ./train.py <dataset> --gpu <gpu_id>
./train.py edges2shoes --gpu 0
It shows that the BicycleGAN output has more variety with smaller batchsize.
- batchsize=2,epoch=60 (left)
- batchsize=60,epoch=60 (right)
- batchsize=2,epoch=5 (left)
- batchsize=60,epoch=15 (right)