The size of pads, strides, and dilations attributes depends on the
size of the input tensor, but our default values for these attributes
uses a fixed size (four for pads, two for strides and dilations),
which causes the code to crash when the input tensor is 5D.
We, of course, cannot determine the shape of the input tensor in the
__init__ function of ConvNode (since the input tensor isn't
available), but based on the fact that we only handle up to 5D input
tensors, this patch sets the default value of the attributes to work with
the maximum input size.
The size of
pads
,strides
, anddilations
attributes depends on the size of the input tensor, but our default values for these attributes uses a fixed size (four forpads
, two forstrides
anddilations
), which causes the code to crash when the input tensor is 5D.We, of course, cannot determine the shape of the input tensor in the
__init__
function ofConvNode
(since the input tensor isn't available), but based on the fact that we only handle up to 5D input tensors, this patch sets the default value of the attributes to work with the maximum input size.