How to create a non-computational visual domain data

[LeighSu]

Hi Yoshitomo,
This is an excellent tool. And Thanks for your shareing.
I am planning to use this toolkit run KD on voice-related tasks, suck like keyword spotting, voiceprint recognition.
I can find the self define loss and model example from other discussion, but I am not sure how to create a new data set, In particular, it's not compatible with torchvision.
I have found a interest annonation, @register_dataset, but I can not find any example or reference about it?

[yoshitomo-matsubara]

Hi @LeighSu

If you're using torchdistill=v0.3.3 (which I'd suggest at this time), you can register your own dataset like this

from torchdistill.datasets.registry import register_dataset


@register_dataset
class YourOwnDataset(...):
    ...

and specify the dataset name and kwargs in your yaml file