Training Style Transfer with ml5js | aiarts.week10

Overview

This week, I trained a ml5js style transfer model with devCloud services. I chose the engraving artwork “Thangka” from artist Wenna as the target style for this model.

Target Style Image: Thangka by Wenna
Target Style Image: Thangka by Wenna

Training

With proper editing to the training script, the training process ran smoothly on the devCloud server. I logged on devCloud and checked on the status of the task a few times through the process with the “qstat” command. Following is the check-in feedback after a few hours.

check phase 2
check-in #3

The entire training process took about less than a day. After the training task is done, a success report can be seen, and model is saved to the specified model folder as shown in the screenshots below.

Result #1
Result #1
Result #2
Result #2

Testing

I transferred the output model to my local machine, and attached it to a ml5js web application. Following are some of the style transferring results on webcam input.

Test Output #1
Test Output #1
Test Output #2
Test Output #2
Test Output #3
Test Output #3

Reference

Wenna – 搜神记, Search of Gods, 一席, Yixi.

Aven – train_style_transfer_devCloud, GitHub.

Leave a Reply