[PYTHON] Train UGATIT with your own dataset
UGATIT is a state-of-the-art machine learning image converter.
* Paper
* GitHub project page
You can train this model on your own dataset.
For this training, we recommend a Google Colaboratory notebook with a free GPU. This is because UGATIT requires powerful computing power.
1, clone from the GitHub project page above.
git clone https://github.com/taki0112/UGATIT.git
cd UGATIT
- Install TensorFlow 1.14 (without TensorFlow1 because this model was created with TensorFlow1 instead of TensorFlow2.0).
pip install tensorflow-gpu==1.14
- Create your own dataset. We recommend using 6200 images (TrainA (DomainA): 3000, TrainB (DomainB): 3000, TestA (DomainA): 100, TestB (DomainB): 100). This is because the selfie2anime dataset in the original project has this amount of images. The size of the image is not important. UGATITutils will automatically resize the image. Create a dataset directory and create a directory for each domain in it.
Specify the directory name of the dataset (for example, “selfly2anime”). Then place the dataset directory in the UGATIT directory.
4, run the train script. You must specify your own dataset name in the “— dataset” argument.
python main.py --dataset your_dataset_name --phase train
Training starts, and the result image and checkpoint are output.
Any question?
Recommended Posts