Read this in [EN|ZH]
Try the demo:
Download pre-trained model on 🤗Hugging Face
Qyin is a model for music score generation and it's based on Qwen2-7B model.
It generates the music score in ABC notation and then can be converted into MIDI files by abcmidi, then into any sound format that is available on this planet.
MidiCaps was used for training dataset for this model. Preprocess scripts will be shared later in this repo.
This model is trained using 3*A100 80G. 5*L20 48G also works with a lower batch size and deepspeed enabled.
It's recommonded to have minimum 16GB video memory to run inference using this model.