support multi treatment in meta learners #141
Conversation
heimengqi
commented
Nov 7, 2019
heimengqi
commented
- extend the meta learners support multiple treatments
- remove DRLearner
- change tests and notebook accordingly
…microsoft/EconML into mehei/metalearnermultitreatment
|
Still need to write a more comprehensive test includes testing multi Y, array Y or column Y, like the test for DML. |
econml/metalearners.py
Outdated
| for model in self.final_models: | ||
| taus.append(model.predict(X)) | ||
| taus = np.column_stack(taus).reshape((-1, self._d_t - 1,) + self._d_y) # shape as of m*d_t*d_y | ||
| if self._d_y: | ||
| taus = transpose(taus, (0, 2, 1)) # shape as of m*d_y*d_t | ||
| return taus |
There was a problem hiding this comment.
Looks like very similar logic to this shows up in a few places. Would it be worthwhile to create a common base class so that the logic doesn't need to be repeated?
econml/metalearners.py
Outdated
| @@ -150,39 +149,42 @@ def fit(self, Y, T, X, inference=None): | |||
| self : an instance of self. | |||
| """ | |||
| # Check inputs | |||
| if X is None: | |||
| X = np.ones((Y.shape[0], 1)) | |||
There was a problem hiding this comment.
[minor]
I think that in this case, the default could be a 0-column array rather than a column of ones (the columns from T will still be there):
| X = np.ones((Y.shape[0], 1)) | |
| X = np.empty((Y.shape[0], 0)) |
There was a problem hiding this comment.
On the other hand, maybe it's silly to even allow X=None because there is no W (unlike DML)
There was a problem hiding this comment.
Yes. In the setting of Slearner, X = None is the same with learning the diff of mean(Y) in each class. I will keep it for now.
