How to improve error rate in practice problem? (image classification - identify apparel)

Analytics Vidhya run a free competition to identify clothing. I found it to be a nice practice problem for beginners.

I used resnet 50, trained for 8 epochs before unfreezing, 14 epochs after unfreezing until error rate remained constant. I threw out data augmentation and normalisation because they didn’t make sense to me for this data set and produced worse results when I tried.

My highest accuracy with their test set is 0.91 and I’m wondering how could this be improved, since the top results in the leaderboard are in the 0.98 range. Transfer learning is not an option since the images are tiny (28x28).

Here is my notebook for reference.

Many thanks in advance for advice!