Torch.nn.functional.xxx vs. torch.xxx

#1

What is the current and future best practice regarding the use of torch.nn.functional.XXX or torch.XXX?

For instance I see that torch.conv{1,2,3}d exist, but only torch.avg_pool1d … Should these functions be already used instead of their torch.nn.functional equivalent? If yes where are the missing average pooling?

#2

If you check source code functionals come from torch.xxx

Altough im not dev, i guess functionals have several benefits for autograd etcetera… therefore I would recommend you to use nn functionals

#3

You use the documented interfaces. They are good for you.
You can use the non-documented interfaces, they will break for you and you get to keep the pieces.
The functions will be mostly identical in function at the moment.

So the back story is:

  • ATen native (i.e. non-TH… legacy) doesn’t know about namespaces at this point,
  • this causes, at the moment, all “public” ATen functions to show up as torch.XXX,
  • because there also is a C++ interface, you cannot rename everything with “_” as one might think of doing,
  • so things showing up in torch.XXX when they’re documented to live in torch.nn.XXX accidental and a small bug likely to be fixed at some point of time.

Best regards

Thomas