Training
Hi!
I didn't find the Issue tab in your GitHub repository, so I'll ask here.
Is there a guide on how to start a workout, how exactly to train, and which dataset is needed? I would also like to try to start a workout.
You will need to prepare a large training dataset consisting of:
- WAV files containing audio data
- PV files containing MIDI at 10 ms, which corresponds to a hop length of 160
You can use the MIR-1K dataset provided here:
https://huggingface.co/datasets/AnhP/Mir-1k-use-DJCM-training/resolve/main/dataset-10ms.zip
After downloading and extracting the dataset:
Replace the dataset path in the code from "hybrid" to your own dataset path.
Run train.py to start training.
If you want to experiment, you can modify the default training parameters to suit your needs.
Thanks! Is there a benchmark code that you used to create your comparison tables?
Thanks! Is there a benchmark code that you used to create your comparison tables?
https://github.com/lars76/pitch-benchmark
If it's not a secret, did you train on this same dataset or a different one, and approximately how many hours of data are there?
No matter how much I try, my results are far from what you achieved. I even added 2 hours of my own data, and it got better, but it's still far off.
If it's not a secret, did you train on this same dataset or a different one, and approximately how many hours of data are there?
No matter how much I try, my results are far from what you achieved. I even added 2 hours of my own data, and it got better, but it's still far off.
If you want, you can try a mix of
Mir-1K + PTDB TUG + Vocadito
or
Mir-1K + PTDB TUG + 3,903 files from M4Singer, with half extracted using PM and the other half using RMVPE, like I did.
Note: you may not achieve the expected results, as I continuously adjusted the parameters during training.