[MRG+1] Multioutput bagging #4848
Conversation
|
|
|
bump |
| proba[k] += all_proba[j][k] | ||
|
|
||
| for k in range(self.n_outputs_): | ||
| proba[k] /= self.n_estimators |
There was a problem hiding this comment.
Do we actually need a loop here?
There was a problem hiding this comment.
It's needed to get a probability. Do you have a suggestion?
There was a problem hiding this comment.
proba /= self.n_estimators
unless proba is not a Numpy array?
There was a problem hiding this comment.
it's a list of the proba for each output.
|
Other than my comment, +1 for merge |
|
Thanks for the review @glouppe ! |
|
appveyor is working good! |
|
Can you please add a new entry in |
sklearn/ensemble/bagging.py
Outdated
| proba_est = estimator.predict_proba(X[:, features]) | ||
| if n_outputs == 1: | ||
| proba_est = [proba_est] | ||
| except AttributeError: |
There was a problem hiding this comment.
why not use hasattr as previously?
|
@ogrisel I have taken into account your comments. |
|
bump fix merge conflict |
|
just out of curiosity: what is your application for multi-output? |
|
I am still pretty torn on multi-output stuff. It adds so much complexity, breaks the API for |
I admit I have the same issue. I dont know if I would implement multi-output in RFs if I had to redo it from scratch... |
|
I am doing a lot of multi-label learning and multi-output regression. Using trees or variations as base learners, I am stuck with the current tree api and also often face this issue #2451. |
|
so downgrading trees to multi-label would help you? |
|
I think that we will see more and more usage of multi-label in the Not to say that I have an opinion on this actual PR: I haven't looked at |
|
@GaelVaroquaux I am totally for multi-label, the question is more multi-output multi-class, which doesn't seem to be a common setting. |
|
@GaelVaroquaux I am totally for multi-label, the question is more multi-output
multi-class, which doesn't seem to be a common setting.
I am with you, completely. My remark was more in the sense of
multi-output trees. But we should wait a bit: Jacob is experimenting with
some rewrites of the tree modules that seem to be giving a factor 2 speed
up in RFs.
|
Given that it's an interface problem and that I am already working around it, it wouldn't help me. |
|
Maybe instead of having a list of array for each proba. We should have a tri-dimensional array of shape (n_samples, n_classes, n_outputs). If the number of classes is not identical, we could set the probability to zero. At least, we would get fast numpy operations and preserve the whole feature. For |
|
Closing as superseded by #8547. |
This pull request brings multi-output support (#3449) to the bagging meta estimators.
It's different of #3449 since the implementation to make the averaging is shared for single-output and multi-output data.
I haven't implemented multi-output decision function as no base estimator currently support this.