nikhil-kotecha / generating_music

Trained on Bach, used LSTMs to generate new music. (Detailed writeup: "Final Writeup - Generating Music using an LSTM Network.pdf")

Generating_Music

Detailed writeup: "Final Writeup - Generating Music using an LSTM Network.pdf"

The Main.ipynb file is divided into 4 cells to individually run:

1): import Python libraries
	INSTRUCTIONS:
		-run cell to import all the necessary Python libraries (REQUIRED)

2): import .MIDI files
	INSTRUCTIONS:
		- run cell to read in the current working directory and specified MIDI files (REQUIRED)
	USER INPUT:
		- specify list of folders to extract music from
		- specify max # time steps required to retain each MIDI file
		- specify # pieces to set aside for independent validation during training

3): Build model graph
	INSTRUCTIONS:
		- run cell to build the Input Kernel, timewise LSTM, notewise LSTM, loss, and optimizer graphs (REQUIRED)
	USER INPUT:
		- specify hidden sizes for timewise and notewise LSTM stages (a list for each)
		  num_t_units and num_n_units

4): Train model
	DESCRIPTION:
		- runs a specified number of batches for training
		- every 100 batches it runs a validation batch and records validation loss in addition to training loss
		- every 500 batches it
	USER INPUT: 
		- specify # epochs (batches)
		- specify batch size (number of Note State Matrices)
		- specify # time steps to train on (must be less than max #time steps)
		- specify model name to restore (or None)
		- specify name to store model to
		- specify keep_prob = 1 - drop out rate
	
5): Generate Averaged Test and Validation likelihoods

6): Music Generation
	DESCRIPTION:
		- runs a for a specified amount of time steps, generating samples of the form Note_State_Batch, storing them, and feeding
			them back to the input of the next time step
		
	USER INPUT:
		- length of song in 16th note steps
		- batch size of generated music
		- specify keep_prob = 1 - drop out rate

All models, loss graphs, and generated midi files are automatically saved to a folder i.e. Generating_Music/Output/Save_Model_Name where "Generating_Music" is the working directory and the subdirectory "Output" is reserved for all the generated files. models are restored from the same corresponding directories

Latest .MIDI files are under Ouptput/Piano_Midi2