This Rust executable converts MIDI files into PyTorch tensors using the performance representation described in the PerformanceRNN paper [1]. It was written specifically to pre-process the files in the MAESTRO dataset [2] and I haven't yet tested it only other MIDI files.
I originally wrote this in Python to reproduce PerformanceRNN but this Rust version is about ~50x faster, which will make it easier to pre-process larger MIDI datasets.
It uses midly to read MIDI files and tch to write to PyTorch tensor files.
cargo run [input_directory] [output_directory]
Example: cargo run ~/Downloads/maestro-v3.0.0 output/maestro-v3.0.0
This will recursively traverse the input directory, convert all MIDI files into tensors, and then write them out to the output directory in the same file system structure. They can then be read in PyTorch like this:
import torch
events = torch.jit.load('file.midi.pt').state_dict()['0']
Each output file is an i16 1D tensor, where each value has the following meaning:
- 0-127 :
NOTE_ON - 128-255 :
NOTE_OFF - 256-355 :
DELTA_TIME - 356-387 :
VELOCITY_CHANGE
The NOTE_ON and NOTE_OFF events are the same as MIDI note values, so an element with value 60 indicates a NOTE_ON event turning on Middle C (C4), and an element with value 128 + 60 = 188 indicates a NOTE_OFF event turning C4 off.
DELTA_TIME events are discretized with 10 ms granularity. For example, an element with value "300" indicates a time delay of (300 - 256 + 1) * 10 ms = 450 ms. Time delays that are greater than 1 second are split up into multiple consecutive DELTA_TIME events, so a delay of 2450 ms is represented as three events: [355, 355, 300].
VELOCITY_CHANGE events indicates volume/emphasis extracted from the MIDI note events, then divided by 4.
[1] Ian Simon and Sageev Oore. "Performance RNN: Generating Music with Expressive
Timing and Dynamics." Magenta Blog, 2017.
https://magenta.tensorflow.org/performance-rnn
[2] Curtis Hawthorne, Andriy Stasyuk, Adam Roberts, Ian Simon, Cheng-Zhi Anna Huang,
Sander Dieleman, Erich Elsen, Jesse Engel, and Douglas Eck. "Enabling
Factorized Piano Music Modeling and Generation with the MAESTRO Dataset."
In International Conference on Learning Representations, 2019.