Dear all,
thanks to a notification from a kind colleague, we have discovered a bug in the GitHub repo. Namely, the MNIST test set downloaded via the `mnist.py` script was in reality still the training set. This was due to a bad copy-paste operated by me (I wrongfully set the `train` flag to True also for the test set).
Hopefully I haven't created any problem to all of you. Fortunately, all the home exercises (besides the Keynote) did not require a competitive test set accuracy on MNIST, so, no harm should've been done.
In case I had, please kindly accept my excuses.
Marco