Pix2Pix

Reference

• Implementation of "Image-to-image translation with conditional adversarial networks"
• https://arxiv.org/abs/1611.07004
• https://github.com/tensorflow/docs/blob/master/site/en/r2/tutorials/generative/pix2pix.ipynb
• The code is written with tensorflow eager execution (2.0)

Train

• Download "facades.tar.gz" dataset and unzip it. https://people.eecs.berkeley.edu/~tinghuiz/projects/pix2pix/datasets/
• Modify the BASE_PATH variable in pix2pix.py to the data.
• run python pix2pix.py to train the model, the weights will stored in weights/, and generated image will saved in generated_img/

Test

run python test.py to load weights from weights/ and generate 5 images.

Network Architecture

• The Generator is a U-NET and defined in PixGenerator.py. The output is tanh activated.
• The Discriminator is a PatchGAN network with size 30*30.

Data

• The dataset use random_jitter and Nomalized to [-1,1]
• The process in written in data_preprocess.py

Hyper-parameter

• The DropOut is used in training and testing, according to the paper. There is no random vector z as input like original GAN. The random of input is represented by dropout.
• In training, BATCH_SIZE = 1 obtains better results. However, if the generator is a naive 'encoder-decocer' network, you should use base_size > 1. The U-NET used here can prevent the activations of bottle-neck layer to become 0.
• In the original paper, the author called 'instance batch normalization'.

Generated image

• Epoch 5
• Epoch 30
• Epoch 70
• Epoch 120
• Epoch 150