About Qmodes

[Siamw]

how can i use kernel-wise quantization ? (Qmodes) Does it work at your code?

[hustzxd]

There is currently no support, but you can add kernel-wise LSQ as follows:

w_reshape = self.weight.reshape([self.weight.shape[0], -1]).transpose(0, 1)
# quantize w_reshape
# ...
w_q = w_reshape_q.transpose(0, 1).reshape(self.weight.shape).detach() + self.weight - self.weight.detach()

[Siamw]

thanks! :D