A few weeks ago, when I started on a project with PyTorch, I wanted to be able to drop certain samples from my dataset as part of a Transform, but found no easy way to go about it.
So I ended up creating nonechucks - a library for PyTorch that provides wrappers for datasets, samplers, and transforms to allow for dropping unwanted samples dynamically. Apart from allowing Transforms to act as Filters, it also allows you to iterate through a dataset containing samples that might be erroneous.
Check it out at https://github.com/msamogh/nonechucks.
If you have any feedback, please do let me know - either down below at the comments or at my email: firstname.lastname@example.org. I’d appreciate it very much. Thank you.