Expand support for grouped convolutions
nickfraser opened this issue · 0 comments
Details
As discussed in #883, grouped convolutions (i.e., groups>1
) are only supported when in_channels == out_channels == groups
. LowerConvsToMatMul
fails with a reshape error.
New behavior
Ideally, it would be great to expand support for arbitrary grouped convolutions. In particular, it could be low hanging fruit to expand current support to allow in_channels == groups
and out_channels = n*groups
(where n is an integer >1) as this would simply result in individual input channels being broadcasted to multiple output channels.
Motivation
Current support handles the common case of standard depthwise-separable convolutions well, but grouped convolutions are becoming more popular as a general optimization strategy to reduce the amount of weights / ops in a given layer.
Parts of FINN affected
Certainly the LowerConvsToMatMul
transform in the FINN compiler is affected. It's likely that some changed would need to be made to FINN-HLSLIB
as well.