Proposed question: improve accuracy of a vanilla CNN model

[samhubs]

When I completed the m2 problem, I also wanted to check if there were ways to improve accuracy from ~69% to higher. The idea that I explored was to creage multiple iteration of m2 solution CNN model (let's call it vanilla) without major changes: added BatchNorm, Dropout, Weight initialization, Scheduler, added additional Conv and Linear layers. I am convinced there should be a question where we provide a base vanilla CNN model and let the learners play with pytorch options and try to improve the accuracy. I am not clear on how to frame the question and the solution because I tried 8-10 iterations and the solution notebook kept on growing without much improvement :). I know using a models.ResNet50 would do the trick but I wanted to extract the most juice from the vanilla CNN model.

Thanks,