vkverma01
commented
4 years ago I think it is not increasing layer, please refer to line no-60 of the code (self.gwc(x) + self.pwc(x)). Here the pwc is not operating on the gwc (Unlike in MobileNet where both are applied sequentially) both are separate and addition operation can be done in parallel without increasing the layer. You can consider as two parallel layers of gwc and pwc and add them.
Please correct me if I am wrong.
Dear irvinxav: Thanks for your code. In you paper writted “In our model, two convolutional layers (depthwise convolutional layer and pointwise convolutional layer) is replaced by one HetConv convolutional layer which reduces the latency from one to zero.” But actually, the Fps be worse because you increase the layer of the network.