2017/18 CC CS Senior Project on Generative Adversarial Networks (GANs) - Eli, Calvin, Aidan, Nikhil

Music, perhaps more so than any other human activity, is regarded as great expression of creativity and emotion. Nonetheless, recent developments in machine learning have expanded the boundaries of computer creativity. One of the most promising developments is the Generative Adversarial Network (GAN). By training two competing networks, a GAN learns to output data resembling its input; one common application is the production of realistic images from a dataset of photographs. By reducing the problem of music generation to that of image generation, we show that the same technique can be used to create music as well as images. Our network learns to generate and visualize original songs in the style of its training input. Using a two-dimensional convolutional GAN and the MIDI file format, we create multi-track songs resembling the human-generated input.

https://www.dropbox.com/s/h3g01tdjh8yaiv6/renaissance-gan-convolutional.pdf?dl=0

Run with:

python3 rgan.py --input [Input .mid file, hdf5 file, or directory] --output [Ouput directory to be created]

Optional arguments:

--epochs [How many epochs to stop after, default 12000] 
--batch [Batch size, default 5, 50 recommended for images]
--save-every [How often (in epochs) to save the image, default 100, 5 recommended for images]
--plot-every [How often (in epochs) to plot the image, default 100]
--no-display (Don't generate live display, necessary if X server is not set up.)

By default, GANiel will display a live feed of his generation, and, in the case of music, play generated files live.