✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 92bd2a6. Link to the linter CI: here}

I think we need to be careful before requiring non-empty steps to a Pipeline. We never know how users use the estimators. For instance you can create an empty pipeline and set the steps afterwards using set_params.

That being said, checking for non-empty steps in _validate_steps seems legit since

• it currently fails with a non-informative error message
• it's a fit time only check
  But maybe we'd like it to work instead of failing, a bit like if it was an implicit "passthrough" ? Because I find the following a bit odd:

Pipeline(steps=[]).fit(X,y)  # fails
Pipeline(steps=[("a", "passthrough")]).fit(X,y)  # works

I'd like to have the opinion of @glemaitre, @adrinjalali, or @ogrisel before making a decision here.

I think the "empty is passthrough' semantics feel quite natural.

Since right now fitting an empty pipeline fails anyway, I'm happy with this PR since it makes the error message much more user friendly and doesn't really change any behavior.

Right now you can't use a Pipeline without steps, so there isn't anyone out there that relies on this. This means we can change the behaviour of __sklearn_is_fitted__ to indicate that it is not fitted and raise a nicer error.

In a new PR we can, if we decide to do it, start allowing empty pipelines.

Right know, this works in jupyter:

pipe = Pipeline([])
      
pipe # to visualise it

If we change the behaviour of __sklearn_is_fitted__ (by raising), the visualisation won't work.
Is that OK @betatim? It seems to me that it contradicts what Jérémy said above:

We never know how users use the estimators. For instance you can create an empty pipeline and set the steps afterwards using set_params.

I am not sure I understand why we should raise an exception from __sklearn_is_fitted__? Can't we return False instead?

I am not sure I understand why we should raise an exception from sklearn_is_fitted?

I thought that it was what you meant (in the issue). 🫣

I don't think __sklearn_is_fitted__ has to be modified. We currently consider a pipeline containing only "passthrough" steps to be fitted. If later we decide to make pipelines with no steps equivalent to an implicit passthrough, then it should be considered fitted for consistency.

I let you decide

I don't need one. But maybe you were asking the others?