I had a project where I was meant to get the highest accuracy possible on this set with PyTorch, no convolution, no pooling, not even a learning rate scheduler, just linear neural networks. Ask me how frustrating that was! The highest I could get was 52%, after HOURS and HOURS