Edit model card

license: gpl-3.0

MusicLang Predict model

MusicLang logo

This is the ONNX version (compatible with transformers.js) of MusicLang-4k.There is a unquantized and quantized version. MusicLang Predict is a model for creating original midi soundtracks with generative AI model.

It can be used for different use cases :

  • Predict a new song from scratch (a fixed number of bars)
  • Continue a song from a prompt
  • Predict a new song from a template (see examples below)
  • Continue a song from a prompt and a template

To solve template generation use cases, we provide an interface to create a template from an existing midi file.

To make the prediction we have an inference package available here : MusicLang Predict which is based on the musiclang language : MusicLang.

Installation

Install the musiclang-predict package with pip :

pip install musiclang-predict

How to use ?

  1. Create a new 8 bars song from scratch :
from musiclang_predict import predict, MusicLangTokenizer
from transformers import GPT2LMHeadModel

# Load model and tokenizer
model = GPT2LMHeadModel.from_pretrained('musiclang/musiclang-4k')
tokenizer = MusicLangTokenizer('musiclang/musiclang-4k')
soundtrack = predict(model, tokenizer, chord_duration=4, nb_chords=8)
soundtrack.to_midi('song.mid', tempo=120, time_signature=(4, 4))
  1. Or use an existing midi song as a song structure template :
from musiclang_predict import midi_file_to_template, predict_with_template, MusicLangTokenizer
from transformers import GPT2LMHeadModel

# Load model and tokenizer
model = GPT2LMHeadModel.from_pretrained('musiclang/musiclang-4k')
tokenizer = MusicLangTokenizer('musiclang/musiclang-4k')

template = midi_file_to_template('my_song.mid')
soundtrack = predict_with_template(template, model, tokenizer)
soundtrack.to_midi('song.mid', tempo=template['tempo'], time_signature=template['time_signature'])

See : MusicLang templates For a full description of our template format. It's only a dictionary containing information for each chord of the song and some metadata like tempo. You can even create your own without using a base midi file !

  1. Or even use a prompt and a template to create a song
from musiclang_predict import midi_file_to_template, predict_with_template, MusicLangTokenizer
from transformers import GPT2LMHeadModel
from musiclang import Score

# Load model and tokenizer
model = GPT2LMHeadModel.from_pretrained('musiclang/musiclang-4k')
tokenizer = MusicLangTokenizer('musiclang/musiclang-4k')
template = midi_file_to_template('my_song.mid')
# Take the first chord of the template as a prompt
prompt = Score.from_midi('my_prompt.mid', chord_range=(0, 4))
soundtrack = predict_with_template(template, model, tokenizer, 
                                   prompt=prompt,  # Prompt the model with a musiclang score
                                   prompt_included_in_template=True  # To say the prompt score is included in the template
                                   )
soundtrack.to_midi('song.mid', tempo=template['tempo'], time_signature=template['time_signature'])

Contact us

If you want to help shape the future of open source music generation, please contact us

License

The MusicLang predict package (this package) and its associated models is licensed under the GPL-3.0 License. The MusicLang base language (musiclang package) is licensed under the BSD 3-Clause License.

Downloads last month
199
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.