LSTM with Attention for Music Generation

Our post on the technical implementation

This was our final project for Data Science Lab. We were scared from the start. How do beginner level data science students with near 0 practical experience come up with an innovative project? Short answer: By shelling $400.

For real, this was an amazing project. We came up with several ideas but settled on this one because our professors are in a band. We figured they would be able to help with something music related. They were the ones who helped us refine the idea into what it is. Music generating software based on Recurrent Neural Networks already exist, but we improved it by adding Attention. Attention allows the model to concentrate on the overall progression/melodic structure of the music. The melodic structure for output sounded pretty incredible.

To train and test this thing, we needed some beefy resources. The paper space servers costed a lot because we were training all the time for almost 2 straight weeks. We ended up securing the highest possible score and the awe of our classmates so it was all worth it. I enjoyed the project but I felt like I didn't understand enough about Attention and RNNs on a technical level. I want to spend more time learning about the lower abstractions of the implementations of different deep models so I can build incredible applications.

I should apply to Spotify!