The errors in using LowerConvsToMatMul ()

[sansi-zhang]

Using Finn's LowerConvsToMatMul() in networks with normal and depthwise separable grouped convolutions Incorrectly:
Cannot reshape array of size 648 into shape (72,81).

I looked at the shape of the weights I was using and I found the error:
The error is because the first layer has been grouped convolved, but the "LowerConvsToMatMul()" "is still ungrouped.

In other words, it does not seem to recognize the grouping Settings, but I looked at the onnx corresponding node information, which contains the grouping information.

So how do I solve this problem?

I hope you can give me some advice and guidance.

Thanks.

[sansi-zhang]

@maltanar
Hi, on the above issues, I made some verification and comparison, and more specifically found the problem:
An error occurred when I tried to implement the channel expansion operation using grouped convolution, which does not seem to support grouped convolution with unequal input and output channels.

Is there a suitable solution in Finn to handle the grouped convolution of this channel transformation?

I hope you can give me some guidance and advice.

Thanks.