Fine-tuning pipeline?

[antchris2]

Hello,
The docs mention the possibility of fine-tuning on bespoke datasets. They also describe a pipeline for training notochord models, via:

python notochord/train.py --data_dir /path/to/data/storage --log_dir /path/for/tensorboard/logs --model_dir /path/for/checkpoints --results_dir /path/for/other/logs train

Is there an already set-up pipeline for similarly fine-tuning an existing notochord model? If not, is there any recommended guidance for doing so?

[victor-shepardson]

hey thanks for the interest! the process for this is a bit hairy right now because the preprocessing script isn't included in the PyPI package. In the current release (0.7.5) you can train with notochord train --data_dir /path/to/data/storage --log_dir /path/for/tensorboard/logs --model_dir /path/for/checkpoints --results_dir /path/for/other/logs train, however newer versions of pytorch might refuse to load a checkpoint for finetuning. I will fix both of these issues very shortly, in the next release.

Finetuning can be done by passing a --checkpoint argument to the train command with the base model. You can get the base model here: https://github.com/Intelligent-Instruments-Lab/iil-python-tools/releases/download/notochord-v0.4.0/notochord_lakh_50G_deep.pt or find it using notochord files after running notochord homunculus once to download it.

What kind of material are you planning to train on, and how much? If you have a very small amount of MIDI data, you might need to add --epoch-size 100 --min-valid 1 --min-test 0. You might also try --freeze-rnn if it overfits very quickly. If you have a bunch of General MIDI files like those in the Lahk MIDI dataset, you can use the existing preprocessing. If you have something not General MIDI-like, you might want to roll your own preprocessing.

I will get back to you once I've made the release and updated the docs.

[victor-shepardson]

hi, I've just published with notochord 0.8.0, see the readme for training instructions. please let me know if you run into any difficulty.