Real-time Performance RNN in the Browser

Thanks to the recently-released deeplearn.js environment, you can now enjoy real-time Performance RNN piano performances in the browser.

deeplearn.js is an open-source Javascript library that enables GPU-based training and evaluation of models in the browser. It includes tools for porting TensorFlow models, which we were able to use to translate a Performance RNN checkpoint into a format that deeplearn.js could use. We then translated our model architecture to use the deeplearn.js TypeScript libraries and added a user interface.

The source code for the new model is available at the deeplearn.js GitHub repository.

We hope you enjoy playing with Performance RNN in the browser and that this serves as an example of how easy it is to translate TensorFlow models to the web. Let us know if you try porting other Magenta models to the browser!