Image-to-Image Translation (Chainer)

Chainer implementation for Image-to-Image Translation Using Conditional Adversarial Networks, it's transplanted from pix2pix.

Result

Step by Step

Download Datasets --- Cityscapes

Sign up, Log in & Download gtFine_trainvaltest.zip & leftImg8bit_trainvaltest.zip from cityscapes

Split Dataset into train, val & test set

    python split_data_sets.py --root [Image Folder] --list [List Folder]  

Note: Run python split_data_sets.py -h for more details.

Combine A & B into a single image

    python combine_A_and_B.py --list [List Path] --save_dir [Save Folder]

Generate list file for train, val & test dataset

	ls train > train.txt
	ls val > val.txt
	ls test > test.txt

Note: Run commands above in folder containing train, val & test subfolder

Train

	python train.py

Note: Run python train.py -h for more options.

Test

	python test.py

Note: Run python test.py -h for more options.

Network Architecture

Generator

U-net

Discriminator

PatchGAN, FOV is determined by n_layers.

Acknowledgments

Code borrows heavily from Chainer: DCGAN.