Data augmentation with albumentations ["transform_config"]

[SprAJR]

I'm training a custom dataset of 1000px square on Efficient_AD. I'm using the albumentations yaml output as an argument in transform_config for eval and val. Does this transform_config actually transform the anomaly mask?

Unless I specify resize 256px (the size of my config.yaml for Effecient_AD) in albumentations, I get a dimension mismatch on training. When I do specify the resize 256px in albumentations, I get a image AUROC of about 50% versus 70-80% without albumentations.

So in addition to wondering whether transform_config transforms the masks, I'm also wondering if resizing at the albumentation level rather than at the model layer is causing an issue.

EDIT: Sorry, my initial description and questions were not well phrased.