large memory footprint during finalize_feature

[JD-ETH]

I was able to get first scores for my trained segmentation network! However, when trying to parallelize, I noticed that my memory usage shoots up in htop (50G memory, 2048 dim, ~52k samples).

I know the API says this step is memory intensive, but looking through the code, I have a hard time understanding why ---- we are just doing a (n, p) @ (p, p) operation here, and a full NxP would only take ~0.2G in memory. I tried to change this line #59 but it has no effect.

Would someone have more insights here?

[JD-ETH]

I see now it's the logger.debug lines that floods the memory. updated the PR and removed those lines.