Something like a loto game (I don’t know all torchvision.models), but the network will be set by the good result versus bad.
The all Relu formula did not allow because is too simple and errors will follow a wrong way.
N-dimensional network will be more precise into training mode.