Unsupervised Attention-guided Image-to-Image Translation

A Tensorlayer implementation of the paper "Unsupervised Attention-guided Image-to-Image Translation"[1].

Network architecture

Results on Horse-to-zebra translation

The images within one row are: original input $s$, the attention mask $s_a$, the masked original image $s * s_a$, and the translated image $s'$ respectively.

• horse2zebra:

• zebra2horse:

  

  

  

requirements

• Ubuntu 16.04.6 LTS

• Python 3.7

• Tensorflow-gpu 2.0.0

• Tensorlayer 2.1.1

• CUDA 10.0

• anaconda

Install Tensorflow-gpu 2.0.0 and Tensorlayer with anaconda

conda create --name project python=3.7
conda install tensorflow-gpu
conda activate project
pip install tensorlayer

usage

Configure your training and evaluation

Before training and evaluation, you have to configure your dataset and hyperparameters in config/exp_xx.json, where xx is your experiment name.

You do not need to download the dataset by yourself: the dataset will be automatically downloaded (a zip file) into your folder. See data_loader.py for details.

Currently we only support the following datasets:

horse2zebra, apple2orange, lion2tiger, summer2winter_yosemite

You can specify your training and testing dataset via editing "dataset_name" in config/exp_xx.json.

• For example, in training mode, the dataset_name can be horse2zebra_train
• In testing mode, the dataset_name can be horse2zebra_test

Training and restoring from the specified checkpoint

To start training, run the train.sh:

bash train.sh

You have to modify train.sh to specify your configuration file name and where to store all the outputs (checkpoints, tensorboard events, and images).

To restore from the previous checkpoint, run the restore.sh

bash restore.sh

Here you have to modify restore.sh to specify your previous checkpoint.

Evaluation

Run test.sh for testing. You can see the results in your log_dir.

bash test.sh

Reference

[1]. Unsupervised Attention-guided Image-to-Image Translation