Keras adds virtually no overhead over torch actually… the torch backend numbers are much worse simply because Keras + torch is not compiled by default, whereas the native torch models were compiled with the best compilation config we found. Keras + torch has essentially the

→ View original post on X — @fchollet