Provide some estimates of memory limitation
We should try and provide some estimates of memory limitations and how, number of training structures and number of parameters affect these things.
For reference both @srinivasan.mahendran and @catherine4hu have recently had some memory errors so this could be quite useful.
Note to self 5/2 - 2020, the expression @freeriks uses for estimating memory usage for numpy arrays of size NxM is pretty much all you need to get a decent estimate of fitting memory usage.
Edited by Erik Fransson