chainer-pix2pix

chainer implementation of pix2pixhttps://phillipi.github.io/pix2pix/

The Japanese readme can be found here.

Example result on CMP facade dataset

From the left side: input, output, ground_truth

usage

1. pip install -r requirements.txt

2. Download the facade dataset (base set) http://cmp.felk.cvut.cz/~tylecr1/facade/

3. python train_facade.py -g [GPU ID, e.g. 0] -i [dataset root directory] --out [output directory] --snapshot_interval 10000

4. Wait a few hours...

• --out stores snapshots of the model and example images at an interval defined by --snapshot_interval

• If the model size is large, you can reduce --snapshot_interval to save resources.

Using other datasets

• Gather image pairs (e.g. label + photo). Several hundred pairs are required for good results.

• Create a copy of facade_dataset.py for your dataset. The function get_example should be written so that it returns the i-th image pair a tuple of numpy arrays i.e. (input, output).

• It maybe necessary to update the loss function in updater.py.

• Likewise, make a copy of facade_visualizer.py and modify to visualize the dataset.

• In train_facade.py change in_ch and out_ch to the correct input and output channels for your data.