Hi,
First of all, thanks for your great work! I have 2 related questions, that I post here.

1st)
I have been able to deploy several networks in FINN and I was exploring more advanced configurations. I tried activations per channel in Brevitas, with following setup for QuantIdentity and QuantReLU:

QuantReLU(act_quant=Uint8ActPerTensorFloat,bit_width=act_bit_width,per_channel_broadcastable_shape=(1,oup,1,1),scaling_stats_permute_dims=(1,0,2,3),scaling_per_output_channel=True)

I tried many absorb/reorder transformations to convert the remaining Mul node, with no success, as activations QuantNodes are not scalar/1D anymore, but tensors. Is activation per channel supported by FINN? If so, which is the sequence of transformations to streamline the model and absorb those Mul nodes, as it happens when activation is per tensor?

2nd)
I have observed that BatchNorm sometimes holds values really close to zero. After Multithreshold absorb them and layers are converted to fpgadataflow, the weightDataType of the Thresholding is very high to keep the results, increasing a lot the LUT usage.

Somehow, I tried the per channel activation to solve it, but it did not work. Is there any way to limit the decimals of those values? Using BatchNorm2dToQuantScaleBias in Brevitas? Limiting in Pytorch? It worked for me configuring QuantReLU as relu6, with min and max val and fixed scale, but causing undesirable accuracy drop.