Symbolic shapes

[hgaspar]

Modern software (e..g pytorch) use symbolic shapes, where the word symbol here is in the sense of sympy (i.e. supporting symbolic manipulations, and symbolic propagation.

For example:

Product of two matrices of symbolic shapes (m,k) and (k,n) is m,n.

Second example: concat of two 1D tensors with symbolic shapes (legths) of "x" and "y" respectively. Obviously the length of an explicitly concatenated 1D tensor is x+y, as a symbol. Strides etc can also be similarly inferred.