Hi,
there could be a number of different things going on. To check the role of the adapter, maybe you could compare the outputs of the configurator vs the adapter for the same data to see if they are identical, or where the outputs differ. Changes in the defaults of the networks for example could play a role as well, that would be my suspicion. I think for us it would be best if you post the recovery plot you obtain with v1 (and maybe a few more details on the networks you used), and (if possible) a self-contained script with the v2 code.
Regarding the one-hot, you might have to ensure that resp, and stim have the datatype integer and not float, and that they run from 0 to (num_classes - 1). If this doesn’t fix it, please reach out.
We have seen a few estimation problems due to our defaults for the DeepSet recently (see for example here). If you are using that, this might also be a problem here. But I don’t know the exact changes between v1 and v2 in this regard.
This would be my first thoughts, I think with a few more details we might be able to give more hints on how to resolve this.