PyTorch models are fit with propensity estimators regardless of weighting strategies used

[kamilest]

Example (executed regardless of the value in the previous if-block at L174) :

CATENets/catenets/models/torch/slearner.py

Line 181 in bb4da46

log.info("Fit the PyTorch po_estimator with the propensity estimator")

This might appear in other places as well—maybe worth checking (@bcebere?)

[bcebere]

Hello @kamilest

The if-block at 174 returns self when it's done

        if self._weighting_strategy is None:
            # fit standard S-learner
            log.info("Fit the PyTorch po_estimator")
            self._po_estimator.train(X_ext, y)
            return self

        # use reweighting within the outcome model
        log.info("Fit the PyTorch po_estimator with the propensity estimator")
        self._propensity_estimator.train(X, w)

The propensity estimator is created/fitted only if self._weighting_strategy is not None.

There is a bit of spaghetti in the code, but otherwise, is this not the expected behaviour?
The PropensityNet class takes the weighting_strategy as parameter.

[kamilest]

Ah yes my bad, didn't check enough times. Thank you! 😅