Vectorization, step 1 of infinity

I achieved the goal of generating “music” in under one minute. And as expected, it sounds like shit. There’s a lot to do… a lot of questions to answer.

I really want the next step to be iterating on the network architecture. But I don’t think I can get there yet. To get any benefit from using recurrent architectures I need a lot of data.

My first experiment uses one song! An I purposefully picked that song to have a very simple rhythmic structure. In fact I can assume it is a sequence of non concurrent quarter notes. It is a bad assumption, but it makes vectorization super simple.

So before I get to hack on fancier networks I have to sort out a more musical vectorization. Ideally the initial vectorization will be a light as possible. I want to push the thinking into the network. What is a discrete package of music? What is the equivalent of a pixel?

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.