ocarina

Music Generator using LSTM

The midi files in ocarina_midi_data/ are over dataset.
We have converted these files into notes and chords and save it into data/ocarina.txt through a music21 library in convert_mid_to_notes.py. master Branch Nearby chords are in the same string and seprated by .. This makes our Hash dictionary to have 286 Keys. nodot Branch Nearby chords are in diffreant string and in text file start with >. This makes our Hash dictionary to have 106 Keys.
This is now used in model.py which has CUDA support on.
This input data is tokenized using a random function to give int values
The tokenized data is then converted into a tensor [a , b ,c]
- c is decided through n, this value indicates the number of notes that belong to same music file are going to help genrate next node.
- b is n//2, which is the number of data under the same label. more information required
- a indicates number of possible such labels more information required
LSTM model as ofknow has some input output issues, explained by error in line 157 and hence line 157 to 165 is just to check for if 1 output is available or not.
Learning Rate is chosen 0.1
Loss criterion is Cross Entropy Loss
SGD is used for backporpagation
Training part is commented out, until Model is fixed.
playMusic function will later have the ability to read the output file and use music21 to convert this back to midi file, to be played back our own generated melody.

batch
n_iters
num_epochs
inpu_dim
hidden_dim
output_dim
seq_dim

BlackDChase/ocarina