Training UGATIT on your own dataset.

UGATIT is a state-of-art of image-to-image technologies.

*Paper

U-GAT-IT: Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization…

*GitHub Project Page.

taki0112/UGATIT

UGATIT Image2Image examples.

You can train this model on your own dataset.

For this training, I recommend Google Colaboratory notebook that you can use free GPU. Because UGATIT needs strong computing power.

1,Clone from GitHub project page above.

git clone https://github.com/taki0112/UGATIT.git
cd UGATIT

2,Install TensorFlow1.14.(If you don’t have TensorFlow1.Because this model is made by TensorFlow1 , not TensorFlow2.0.)

pip install tensorflow-gpu==1.14

3,Make your own dataset. I recommend using 6200 images ( TrainA(DomainA):3000,TrainB(DomainB):3000,TestA(DomainA):100,TestB(DomainB):100).Because the selfie2anime dataset for original project has this amount of images. Image size does not matter.UGATIT utils automatically resize your images. Make dataset directory, and make directories for each domain in it.

Name dataset directory name (e.g. “selfie2anime”). And put dataset directory in UGATIT directory.

4, Run the train scripts. You have to specify your own dataset name in “— dataset” argument.

python main.py --dataset your_dataset_name --phase train

Please follow my Twitter. https://twitter.com/JackdeS11 And please clap your hands 👏.

Any question?