tensordot would be a useful function to have for general contractions between tensors. It looks like all the basic functionality is already present, but not a numpy-matching api. One nice gain if this was present would be that pytorch could be used as a backend by opt_einsum for order-optimized tensor contractions.

Two simple and purely python options would be:

1. a dot/matmul based method - like this gist (adapted from numpy's implementation itself I think)
2. an einsum based method - like in this PR (not for pytorch but similarly translates tensordot calls to einsum calls)