Struggling with SVCCA in machine learning. Any tips are highly appreciated #pytorch #machinelearning https://discuss.pytorch.org/t/svcca-on-lstm-output/58096
@semordnilap don't have experience with SVCCA but from other Python stuff it's often matrix shapes, or other incorrect input.
(ranting a little)Imho libraries should try catch it earlier.. (also lower the LOC.. cant the documentation be separated..)
@jasper The weird thing is that matrices randomly generated with the same sizes don't return this error. It must be something in the preprocessing, probably the svd or even earlier in the preprocess.
@semordnilap hmm searching https://duckduckgo.com/?q=DLASCL+parameter&t=h_&ia=web it might be there is a NaN in your input or generated somewhere?
Also see there is a `verbose` option https://github.com/google/svcca/blob/master/cca_core.py#L217 it seems possible it'll print something.
More paranoidly... presumably it behaves the same if one of the `a`s is a copy.
Linux geeks doing what Linux geeks do...