I was recently trying to deploy the network on an FPGA using VTA, but I ran into this problem when deploying mobilenetV2. When executing nn.conv2d, I set groups to the following format:
groups = int(in_planes/16)
return nn.Conv2d(in_planes, out_planes, kernel_size=3, groups=groups ,stride=stride ,padding=1, bias=False)
Although such setting can be achieved, the final result is inconsistent with the result obtained by software training. I don’t know why such a result occurs. Have you tried to implement Mobilenet and verified the correctness of the result?