Tony Morris

you can read that IMHO http://nlp.seas.harvard.edu/NamedTensor  ... it's recent and shows a few issues... tensor are just matrix with dimension batch x dim1 x dim2 x ... and all matrix computations aren't possible depending on those dims... a type-dependent system could help but python is farfar

What manual process and/or tools do you currently execute to check your matrix dimensions?

model.compile(...) in tensorflow

but imagine you had a kind of "compiler" that would check in real-time your python code and infer dimensions and check you have no inconsistencies in your dimensions...

Do you have a mechanical process by which you (the programmer) can currently do this? i.e. a method of program analysis, however manual?

no rely on TF/pytorch checking capabilities... It doesn't help coding robust, you must run "model.compile/model.fit" to know if it works & if you have branch if/then/else in which you've made a dimension error, your training can crash after 5h when it finds this condition

Is pytorch the thingo that does the dimension checking?

pytorch is just another DeepLearning lib... it has a different approach than tensorflow but it has same tensor concept... it separates less a static phase "compile" vs "fit" but it's not so far despite what pytorch people say...