Make a RoboJam model
In this exercise you’ll try making your own MDRNN. The example code for RoboJam and the Kanji RNN are below.
Notebook: MDN Robojam Touch Generation (Github) (open in Colab)
- Start by stepping through the code for either the Kanji or RoboJam MDRNN. These systems both use 3D RNNs: 1 dimension for each of the x and y directions, and the pen state (down or up) for the Kanji RNN, and time for RoboJam.
- Think about some continuous-valued data you would like to model. It
could be 3D like the Kanji, or RoboJam RNN, or have more or less
dimensions (it’s easy to change the dimension of the RNN). Try to
find some data in form that is easy to open with
numpy. Try to load the data into the notebook and plot sections of it.
- Try training a new MDRNN with your data! It might take a while to start getting useful results, but keep track of what you did so that you can improve your MDRNN next time.
- Can your generated data be incorporated into an artwork, composition or instrument? Think about how the different dimensions of the data are represented. Are they related or independent?