How to process an extremely uneven number of sequences?

I have an uneven training sequence for classification.
For example, the image of the first category has 1000 tiles.
The second category has 10 tiles.
These tiles are the same size, e.g. 1024*1024.
Is there the best training strategy?

The main problem is that I cannot pad all sequences to 1000 because of my limited GPU mem size.