Alot of these networks are still too heavy for cpu inference would you be open to a PR to add these models ?