MLP in Allegro

[kudoryutaro]

Why does allegro's MLP use fx.Graph() and fx.proxy.GraphAppendingTracer(graph)?
Any reason for using fx.Graph() instead of the usual torch.nn.Linear?
Is it faster to use fx.Graph()?

[Linux-cpp-lisp]

Hi @kudoryutaro ,

torch.fx is an API for building torch.nn.Modules that can use any PyTorch operations to do the computations: https://pytorch.org/docs/stable/fx.html. In practice, I don't think there is any user-relevant difference between our MLP implementation and the more basic one you are describing.