How can I apply drift_analog_weights in the forward path during training?

[matifali]

I am trying to train a network where I want to apply the drift_analog_weights in the forward path but only want to use it for computing the forward pass and not update(drift) the weights for update.

Can it be done with the current setup?

[maljoras]

This is not possible directly, but that is the reason of the rpu_config.modifier settings (see here). There a noise can be defined that is applied during forward (and backward) pass, but not on the updated weights. This noise is drawn again for each mini-batch. Currently only a programming noise polynomial can be given, but not the exact drift (however, we plan to update this in future).