pop2piano / preprocess /README.md
sweetcocoa's picture
initial test
88490a8
|
raw
history blame
1.28 kB

Preprocess Scripts


  • Note : the order of these scripts is IMPORTANT.
  • the preprocessing step is easy. but environment setting is not. please understand.
  • If you encounter any problems, please do not hesitate to email me or open an issue to the github.
  1. Transcribe piano wavs to midi
  • You should transcribe {piano_cover_file.wav} -> {piano_cover_file.mid}

  • I recommend you to use original codes from this repo : High-resolution Piano Transcription with Pedals by Regressing Onsets and Offsets Times

  • Instead, you can also you my docker script.

    docker run -it --gpus all --rm -v /DIRECTORY_THAT_CONTAINS_PIANO_WAV/:/input -v /DIRECTORY_THAT_MIDI_OUTPUT/:/output jonghochoi/piano_transcribe:bytedance1
    
  • If you are using GPU RTX 30XX or higher, this script may not work properly. It's because the version of pytorch is too low(1.4).

  • then upgrade the version of pytorch in the docker..

  1. Estimate Pop's beats
python bpm_quantize.py DATA_DIR 
  1. synchronize midi
python pop_align.py DATA_DIR
  1. get separated vocal track
python split_spleeter.py DATA_DIR
  1. caculate melody chroma accuracy
python melody_accuracy.py DATA_DIR