Why the non-linearity in the last layer?

[marcelodiaz558]

Hi,
I don't understand why developers decided to include a swish non-linearity in the architecture right after the fully connected classifier layer, I don't think that it's good choice for an image classification task since swish will throw out some information that might be important, just like ReLU does with negative values. Most likely I'm wrong and there's some good reason to do this, can anyone please explain this to me?

This is how the network forward propagation ends:

(_fc): Linear(in_features=1280, out_features=1000, bias=True) (_swish): MemoryEfficientSwish()

Thanks