Arbitrary Style Transfer in TensorFlow js
This is an implementation of Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization on Tensorflow 2 and Tensorflow js. Demo website : https://cryu854.github.io/ArbitraryStyle-tfjs/
The model runs purely on your browser, so your data will not be leaked.
REFLECT_PADDING=True to use reflection padding in both encoder and decoder to avoid border artifacts, but the model will not be able to be deployed on the browser.
Use main.py to stylize a content image to arbitrary style.
Stylization takes 29ms per frame(256x256) on a GTX 1080ti.
Example usage:
python main.py inference --content ./path/to/content.jpg \--style ./path/to/style.jpg \--alpha 1.0 \--model ./path/to/pre-trainind_model
Use --alpha to adjust the stylization intensity. The value should between 0 and 1 (default).
Use main.py to train a new style transfer network.
Training takes 2.5~3 hours on a GTX 1080ti.
Before you run this, you should download MSCOCO and WikiArt dataset.
Example usage:
python main.py train --content ./path/to/MSCOCO_dataset \--style ./path/to/WikiArt_dataset \--batch 8 \--debug True \--validate_content ./path/to/validate/content.jpg \--validate_style ./path/to/validate/style.jpg
Use tensorflow-js converter to generate a web friendly json model.
If you use reflection padding in encoder or decoder, the converter will not work properly because the current version of tensorflow-js does not support the mirrorpad operator.
Example usage:
tensorflowjs_converter --input_format=tf_saved_model --saved_model_tags=serve models/model models/web_model