GSConv code

[stoor2kushidai]

In the implementation of GSConv, the second layer of convolution is not the DWConv pointed in the paper, but a standard convolution. What is the reason for this?

[alanli1997]

In the implementation of GSConv, the second layer of convolution is not the DWConv pointed in the paper, but a standard convolution. What is the reason for this?

It looks like you are ignoring the "groups" argument in the second layer. In fact, we defined "groups=c_" in the Conv block, which is the same as the DWConv.

    self.cv1 = Conv(c1, c_, k, s, None, groups=g, act=act) # groups=g=1
    self.cv2 = Conv(c_, c_, 5, 1, None, groups=c_, act=act)  # the "c_" is the channels number, so this actually is a DWConv layer.