aaroh-dell/Solar-Flare-Prediction ? reverse-engineered prompt

Build me a notebook that predicts solar flare severity from NASA SDO magnetogram images.

I want it to use 128 by 128 single channel inputs, with a shared ResNet 18 style backbone and two outputs, one for flare class, B, C, M, or X, and one for magnetic flux regression. Please handle the class imbalance properly so the model cares more about missing M and X flares, not just overall accuracy. I’d like training, evaluation, and simple plots for loss and performance, plus clear metrics that focus on recall for the rare high intensity flares.

The notebook should run in a normal local setup too, so please replace any Kaggle only file paths with something easy to change. If needed, look up the current PyTorch and torchvision docs online and use a clean modern approach.