pix2pix implemented by keras

This is a keras implementation of paper Image-to-Image Translation with Conditional Adversarial Networks (pix2pix). I learned a lot from tdeboissiere's code.

Requirements

• python 3.6
• keras 2.1.5
• tensorFlow 1.4
• opencv 3.4.1

Getting Started

1. Preparing your data

All your training data and validation data should be processed into A2B form as shown in figure. Then moving the training images (3 channels) into folder ./data/train , and validation images into folder ./data/val .

2. Training

python train.py

• Model pictures and generated validation pictures will be saved in folder ./figures, and weights will be saved in folder ./weights/pix2pix.

• If you want to use perceptual loss, please replace loss = [l1_loss, 'binary_crossentropy'] (line44) by loss = [perceptual_loss, 'binary_crossentropy'] (line45) in train.py.

3. Testing

You need to change the path of folder testset_dir and save_dir , and run

python test.py

Have fun!