SRNet training crashes on Kaggle OOM

This looks less like a model failure and more like a data-pipeline trap: padding every batch to the largest sample means one outlier can dominate memory usage and take the whole job down.

• –Pre-scan image dimensions and either resize, bucket by aspect ratio, or drop pathological outliers before training
• –Checkpointing matters here; if the run crashed without saved state, resuming means restoring the last good checkpoint, not trying to revive the dead notebook
• –Kaggle memory limits are tight, so smaller batches, mixed precision, and offline preprocessing are the practical fixes
• –If SRNet is being used for steganalysis or another research workload, this is a good example of why variable-resolution training needs guardrails from day one