Pix2Pix Tensorflow

This is a general purpose implementation of the pix2pix algorithm for image to image translation. This algorithm is based on pix2pix by Isola et al.

Code in this repo has been heavily borrowed from this implementation of the pix2pix tensorflow. But the linked code is not easy to use as I have experienced it first hand while working on a related project.

This repo is an attempt to make the pix2pix implementation easily approachable for training and testing.

A guide on how to use this code for image to image translation is provided.

---

Dependencies

You should have the following dependencies installed in your system to run this code.

• python 3
• tensorflow==1.10.0

Getting Started

1. Training

The goal in image to image translation is to convert the input image A to target image B.

Example:

You can also specify the mapping AtoB or BtoA in config.py.

For training, you should generate such images and put all the training images in the train_data folder in the root directory.

Alternatively, you can also use --input-dir flag to set your custom input directory.

e.g

python train.py --input_dir <training_images_folder>

For more control over the training, refer to config.py which contains all the configurable settings to be used. You'll find comments there to help you out.

2. Generating samples

For generating output samples using your trained model, you should follow this pattern.

• Create folder inputs inside folder test_data inside the project root. (Note: You can change this in the config.py file)
• Put your input images inside the inputs folder and run test.py.
• You can specify the --checkpoint flag to point to the folder where model checkpoints are saved.

Contributors

1. Pranav Gajjewar
2. Omkar Thawakar