search

Natooz / BPE-Symbolic-Music

Code of the paper "Byte Pair Encoding for Symbolic Music" (EMNLP 2023). Better and faster music generation
https://natooz.github.io/BPE-Symbolic-Music/
38 stars 2 forks source link

The usage on Orchestra music #4

Open MHZ9825 opened 4 months ago

MHZ9825 commented 4 months ago

Thanks for your Briliiant work !!! However, I'm trying to use it on Orchestra, which means I have to get the hidden features from a multi-inst, if the work get the way to get the hidden features? It seems I can only get the tokens without inst.

Natooz commented 4 months ago

Thank you for your message! I am not sure to understand exactly what you are trying to accomplish. Is Orchestra a specific dataset? Or do you mean you are using multiple instruments/tracks? In any way it is possible to use byte pair encoding (or unigram or wordpiece) with multiple instruments, whether you tokenize all tracks all at once (config.one_token_stream_for_programs enabled) or separately. I invite you to take a look at MidiTok's documentation for more details as this repository is not up-to-date anymore :)

MHZ9825 commented 3 months ago

Hi again! Thanks for the reply~ I have checked the documentation and have conveniently used MIDITOK to get the MIDI Token ids. however I still haven't found out how to compress the token sequence using BPE. May I ask how this step should be handled. Can I assume that I need to build my own dataset and then train a Tokenizer of my own using BPE's method?

Natooz commented 3 months ago

This is done automatically once the tokenizer has been trained. :) You can see the documentation of the encode (to tokenize files) and encode_token_ids (to apply BPE/Unigram to token ids and reduce the sequence). The latter is automatically called by encode unless you specify it to not.

MHZ9825 commented 3 months ago

Thanks！I 'll try it~