Week 5 — Artificial Pianist
Last week we mentioned different approaches on how to evaluate the results, the music files generated by the model. We have come to the conclusion that there should be 2 different approaches, one of them should be objective and the other one should be subjective.
This week we completed training the model and find some results.
First Notes from The Artificial Pianist
The Artificial Pianist is still very novice and a little shy but improving itself day by day. There are some music pieces played by The Artificial Pianist.
For subjective approach we made a simple user qualitative study on the results and we reach a decision that we should improve the model. We searched different researches on the topic and the transformer model.
Multi-Head Attention
Let’s assume that we are trying to estimate the next piano chard using the previously used one piano chard. The outcome will be annoying for the listener because the music should be created by considering the harmony of piano chards. Therefore, we have decided to use “multi-head self-attention layer” in our model. This layer allows the model to jointly attend to information from different representation subspaces at several positions. With this approach, we can define a model that can produce music artistically.
