Before I open a tiny PR, I am going to post in the forums first
I would like to add a
max-scale parameter to the
RandomResizedCrop(GPU) init function.
Then the only other change that has to be made is to sample from
min_scale, max_scale and not like it currently is from
Why? In the self-supervised domain, multi-crop augmentation pipelines considerably improve performance. (see here)
In the multi-crop augmentation strategy, one generates high and low-resolution crops.
The low-resolution crops must be known for some methods and the crop method should therefore have a
One could use the touch vision transform as an alternative, but I am working with low-resolution images, and the single interpolation strategy from fastai sounds ideal for this scenario.
By setting the default value to 1. there would be no breaking changes and the PR would be quite small.
But if it is possible, I would still like to create a PR and give something back.