You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, pypesto.predict.amici_predictor.AmiciPredictor.__init__ requires a pypesto.objective.amici.amici.AmiciObjective.
Constructing an AmiciPredictor from an AmiciObjective is admittedly convenient. However, requiring an AmiciObjective renders it unnecessarily complicated to make ensemble predictions for anything other than the original parameter estimation problem (e.g., prediction for unseen experimental conditions).
The only things that AmiciPredictor should require, are an amici model, a (potentially optional) list of amici.ExpData, an amici.Solver, and a parameter mapping. Therefore, I'd suggest to refactor AmiciPredictor to be constructable from only those inputs, while preserving the possibility to construct it from an AmiciObjective.
EDIT: Not so easy to decouple in the case of hierarchical optimization, since the inner parameters aren't included in the ensemble. However, it's questionable if the current state, i.e. recomputing the inner parameters is desirable. I'd say that mostly ones wants the optimal inner parameters computed from the training data and not from any unseen data.
The text was updated successfully, but these errors were encountered:
Currently,
pypesto.predict.amici_predictor.AmiciPredictor.__init__
requires apypesto.objective.amici.amici.AmiciObjective
.Constructing an
AmiciPredictor
from anAmiciObjective
is admittedly convenient. However, requiring anAmiciObjective
renders it unnecessarily complicated to make ensemble predictions for anything other than the original parameter estimation problem (e.g., prediction for unseen experimental conditions).The only things that
AmiciPredictor
should require, are an amici model, a (potentially optional) list ofamici.ExpData
, anamici.Solver
, and a parameter mapping. Therefore, I'd suggest to refactorAmiciPredictor
to be constructable from only those inputs, while preserving the possibility to construct it from anAmiciObjective
.EDIT: Not so easy to decouple in the case of hierarchical optimization, since the inner parameters aren't included in the ensemble. However, it's questionable if the current state, i.e. recomputing the inner parameters is desirable. I'd say that mostly ones wants the optimal inner parameters computed from the training data and not from any unseen data.
The text was updated successfully, but these errors were encountered: