Does that mean you're training a lower bit-rate(?) network or you are training a full network to 'know' it will eventually be running under quantization?

I'd imagine there's differences in the two approaches?