Linear layer requires 2D tensor. Does it mean that I got to use ‘view’ function to reshape the previous output every time I call it?
But for some experiments BN and RELU are both needed afterwards. They both require 4D, for which I need to call reshape function again and again…
and I can’t find one corresponding reshape function for nn.Sequence()? Is there any plan to add it?
What about BatchNorm1d (ref.)?
And of course, a Linear Module is simply an affine transformation, hence it requires vectors (1D) or stack of vectors, i.e. matrices (2D).
ReLU() is a point-wise operator, so it should work with any input dimensionality.
No, we don’t plan to add any modules for reshaping, that could be put into a Sequential. You might want to take a look at how torchvision models are implemented.
thank you for replying me! I’m so sorry for not noticing BatchNorm1d in the document. Now my confusion is perfectly solved~
that’s reasonable enough~ Thanks for your replying!