- for more information about the project visit : https://github.com/fatemetkl/music-generator-Tensorflow-/blob/master/Music%20Composer-presentation.pdf
- the music composer implementation using keras is available at https://github.com/fatemetkl/Music-composer-keras
“What I cannot create, I cannot understand.”- Richard Feynman
• Project’s dataset is a collection of about 175,000 MIDI files (we used The Lakh MIDI Dataset v0.1) • We used pretty_midi to help us with reading and writing. MIDI file structure looks like this:
• As mentioned in the chart , we see that each song has multiple instruments • But we chose the instrument with longest bars (so in each song, the instrument with greatest number of notes were kept) • Drum based instruments were left out because they have different annotation system than other instruments like piano, violin, etc.
• Each note has this values:
Start time End time Pitch Velocity • We categorized each note due to its values. • Then we hot-encoded notes due to categories
• Instead of start time and end time we used the difference between them (duration) and we already had velocities and pitches. • We calculated max value for each of the three variables mentioned above and splitted the notes according to their distribution of all notes in that song.
• Instead of loading all the training dataset at once, training data are loaded in sets of specific size (definition of batch) • We have a BatchProcessor class that generates batches for each iteration. • Each batch’s notes are one-hot-encoded together.
• We needed an architecture that generates new songs. • First of all, because we needed to generate new songs, we used encoder-decoder architecture. • Second of all, because music and songs are sequential, we needed to use recurrent networks; so we chose LSTM.
• It can be difficult to train standard RNNs to solve problems that require learning long-term temporal dependencies because of gradient vanishing • LSTM has memory; it can maintain information in its memory for long periods of time. • A set of gates is used to control when information enters the memory, when it's output, and when it's forgotten.
• We used TensorFlow. • We needed to have control over our network and TensorFlow gave us more control than Keras • For example, in TF, we have a computation graph that we can do any computation with any input that we want. • Also TF is more used in researches and building new networks as well. • TF’s syntax is harder than Keras and might be more confusing.
