cuda: false

[jesper-olsen]

There is a "cuda":true flag in mnist.json.

I notice the code breaks if that flag is flipped to false - I guess you didn't try running without?

There are a few places that have to be edited.

Running the code on my Macbook (M1, no cuda), I notice CPU utilisation
only seems to be ~120% (over 8 cores), and I wonder if this could be
improved. Maybe a more "vectorised" implementation is not possible?