/Unsupervised-Symbolic-Music-Decipher

Primary LanguageJupyter Notebook

Prepare Repo

  • Use conda to install "./environment.yaml"
    • Fix CUDA version to your native.
  • Copy "./example.env" to "active.env".
  • Download the LA Midi Dataset and set LA_MIDI_PATH in "./active.env" to the midi dir (ending with .../Los-Angeles-MIDI-Dataset-Ver-4-0-CC-BY-NC-SA/MIDIs)
  • Install fluidsynth.
  • Is your fluidsynth configured interestingly like NYUSH?
    • The NYUSH HPC fluidsynth is peculiar in the following known ways:
      • The output is always s16le PCM.
      • Output file has missing/incorrect metadata.
        • It is stereo, not mono. The SR is not always 44100.
    • Given your fluidsynth, either keep or remove the is_fluidsynth_nyush flag in "./hpc/prepare_datasets_template.sbatch".
  • Obtain "MuseScore_Basic.sf2" and put it in "./assets".
    • Instructions are in "./assets/acknowledge.md".
  • Run cd ./hpc
  • Run python3 ./midi_process_parallel.py and wait for completion.
  • Run python3 ./prepare_datasets_parallel.py --stage cpu and wait for completion.
  • Run python3 ./prepare_datasets_parallel.py --stage gpu and wait for completion.

Train

From the outermost dir,

  • Run cd ./hpc
  • Run python3 ./sched_piano.py to submit a slurm job.
    • It uses train_piano_template.sbatch which has some NYUSH HPC-specific setups. You may want to review it and make edits.
    • Under the hood it runs ../main_train_piano.py. Feel free to run it directly.