Hello,
I’m the maintainer of a datasets/transforms library for event-based data called Tonic: GitHub - neuromorphs/tonic: Publicly available event datasets and transforms.
I modeled the API after torchvision to a great extent, which was so helpful! One thing though that always bugged me is that I relied on a copy of the download utils to be able to download and check datasets. These days I’m working on migrating Tonic to depend on torchdata and datapipes, which I’m really looking forward to. My hope was that I could rely on torchdata functionality to download and check data. Unfortunately I saw that torchvision relies on its own download utils scripts. I was wondering, why not move those scripts to torchdata directly? At least Tonic, but likely also audio/text could benefit from that code.
Thank you so much!
I think your idea is great and am sure the code owners would be interested in discussing it further. Would you mind creating a GitHub issue in the vision or data repository and would you be interested in working on an implementation?