I am debugging my network so I have a chance to look at the source code of NCS installed at
/usr/local/bin/ncssdk. I do not quite understand the logic of
parse_tensor in which it iterates on each ops and attach to the
network object. However at some layers there are local attachments, for example
Conv2D attached with
SeparableConv2dNative, and then
Conv2D is attached again to the network which may create cyclic paths I guess?
@dpvo You can try using mvNCProfile to check if this is indeed the case with your model. mvNCProfile creates a visual representation of the model from input to output by creating a output_report.html file and a output.gv.svg image of the network layers.