In PTQ, is possible to quantize activation in per channel mode?

quantization
1

As far as I know, we can quantize weight in per-channel mode by using default_per_channel_weight_observer. But I did not see any per_channel related configuration for activation.
Is possible to do it for activation?

2

Hi @111357,

Sorry for the late reply. You should be able to set this by changing the qconfig. See here for more details.
I think you should be able to use default_per_channel_weight_observer, from the implementation here it looks like it’s just calling PerChannelObserver with dtype=int8.