Hi,
in the solutions notebook of lab 5 the confusion matrix is totally off as it represents/shows a much lower accuracy than what expected after the training.
I had the same issue and resolved it by applying the normalization transforms on the data before feeding it to the model for evaluation. This made me recover exactly the same accuracy as given during the training/fitting run.
Hope this helps. Best regards