Music generation and tuning using convolutional networks and reinforcement learning

Required Packages:

• PyTorch
• Mido (install with: pip install mido)

Our demo allows you to generate MIDI samples with a given policy derived from a pre-trained Q-network. The model is primed with a random starting state then generates a drum composition with 32 sixteenth notes for a total of two measures. To play the MIDI, use a MIDI-enabled media player such as WMP or VLC, or use https://www.midieditor.org/ .

example: python Demo.py 0.05 10

• "Note_CNN" - supervised model
• "0.01" - Q-network trained with epsilon = 0.01 greedy action selection
• "0.05" - epsilon = 0.05
• "0.1" - epsilon = 0.1
• "0.3" - epsilon = 0.3
• "0.5" - epsilon = 0.5

• integer, the number of MIDI samples you wish to generate

• num_samples MIDI samples with filenames "demo_song-#.mid"