Current implementation does not work for new training procedure with layer replacements and training continuation