GANs work by pitting two AI programs against eachother - one to spot true and fake examples of a certain type of media or object. It is the discriminator. The other creates fake objects to try to trick the discriminator. It is the forger. Eventually, these two can theoretically reach an equilibrium. The majority of the program is located in gan.py
Recursively iterates over a tensor generated from a 15-second .wav file - looking for similar characteristics that it can find that signify a musical clip, returning a single digit classifier. 1 for yes, 0 for no. It also has methods to evaluate but not train, mostly for the purpose of analyzing the forger's work
Starts with random inputs and generates what could be a music file, passes it to the discriminator. Eventually over enough iterations, it optimizes to maximize the output of the discriminator. In other words, it learns how to make files that the discriminator considers music.
Also in this file are other dependencies such as audiotorch-using methods I made to convert .wav files into RNN-acceptable tensors. Another method I have is one that divides the testing and validation sets up randomly by a preconceived ratio so we can evaluate our model after we train it.
Most of these files don't sound like music. But if you want to listen to them there will be a directory with a few examples in them.
I got a lot of help and support from Asya Shklyar, Pomona ITS and HPC, Professor Joseph Osborn on questions regarding Cuda, GPUs, PyTorch, ffmpeg, data sources, and etc. If I hope to continue this project I will probably seek their help again! Additionally, I must also mention that the skeleton for the RNN in both classes was based on https://pytorch.org/tutorials/intermediate/char_rnn_classification_tutorial.html this tutorial from PyTorch.