Currently network layers know only their size, which is basically the number of channels or variables in their outputs. For feedforward and most recurrent models this is sufficient, but for convolution models this quickly becomes painful to deal with since the user needs to compute the sizes of the image patches at each step. theanets should permit each layer to know its shape so that we can simplify some of this for the end user.
Currently network layers know only their
size
, which is basically the number of channels or variables in their outputs. For feedforward and most recurrent models this is sufficient, but for convolution models this quickly becomes painful to deal with since the user needs to compute the sizes of the image patches at each step.theanets
should permit each layer to know itsshape
so that we can simplify some of this for the end user.