Questions

1. If you want to generate a musical score, what do you train your model to do?
2. What are the properties that are interesting in RNNs concerning music prediction?
3. Given an RNN hidden layer with the notation h(t + 2), what two inputs is the hidden layer getting?
4. Given the following parameters for the generation, --num_steps 32 and --qpm 80, how long will the generated MIDI be in seconds? How many bars will it be?
5. What happens if you increase --branch_factor and increase --temperature during the generation phase?
6. How many nodes will the beam search algorithm go through at the last iteration for a generation of 3 steps with the --branch_factor 4 and --beam_size 2 parameters?
7. What is the Protobuf Message class that is used in Magenta to represent a sequence of MIDI notes? (NoteSequence)
8. Using the one-hot encoding described in the encoding section, what is the encoded vector...