No warning when LogisticRegression does not converge #10866
Comments
|
If you use verbose=1 in your snippet, you'll get plenty of ConvergenceWarning. from sklearn.datasets import load_breast_cancer
from sklearn.linear_model import LogisticRegressionCV
data = load_breast_cancer()
y = data.target
X = data.data
clf = LogisticRegressionCV(verbose=1)
clf.fit(X, y)
print(clf.n_iter_)I am going to close this one, please reopen if you strongly disagree. |
|
I also think that It does not make sense to expect the users to set the verbosity in order to get the warning. Also other methods don't do this: e.g. MLPClassifier may raise ConvergenceWarnings so does MultiTaskElasticNet or LassoLars. The culprit is this line that only affects the |
|
After more thoughts, the case of However, from sklearn.datasets import load_breast_cancer
from sklearn.linear_model import LogisticRegression
data = load_breast_cancer()
y = data.target
X = data.data
clf = LogisticRegression(solver='lbfgs')
clf.fit(X, y)
print(clf.n_iter_)silently fails to converge, and this is a bug IMO. This was also recently relevant in #10813 |
|
A quick git grep seems to confirm that ConvergenceWarning are generally issued when verbose=0, reopening. The same thing happens for solver='liblinear' (the default solver at the time of writing) by the way (you need to decrease max_iter e.g. set it to 1). There should be a test that checks the ConvergenceWarning for all solvers. |
|
I see that nobody is working on it. Can I do it ? |
|
Sure, thank you @AlexandreSev |
|
Unless @oliverangelil was planning to submit a PR.. |
|
You can reuse test_max_iter indeed to check for all solvers that there has been a warning and add an assert_warns_message in it. If you do that, test_logistic_regression_convergence_warnings is not really needed any more. There is no need to touch test_lr_liblinear_warning. The best is to open a PR and see how that goes! Not that you have mentioned changing the tests but you will also need to change sklearn/linear_model/logistic.py to make sure a ConvergenceWarning is issued for all the solvers. |
|
yes I think there are a few cases where we only warn if verbose. I'd rather
consistently warn even if verbose=0
|
…ion. For now, convergence warnings are silent when verbose is set to 0 for lbfgs and liblinear solver, whereas they are not for others. Fix scikit-learn#10866.
Modify slightly test_max_iter to check that a warning is sent. See scikit-learn#10866.
…ion. For lbfgs and liblinears solvers, the convergence warnings appeared only when verbose was greater than 0, whereas they appeared with verbose = 0 with other solvers. See scikit-learn#10866
Modify test_max_iter to also check for the warning, and remove a test which becomes redundant with this modification. See scikit-learn#10866.
…re_warn With the add of convergence warning in logistic regression, this test failed. See scikit-learn#10866
…ion. For lbfgs and liblinears solvers, the convergence warnings appeared only when verbose was greater than 0, whereas they appeared with verbose = 0 with other solvers. See scikit-learn#10866
Modify test_max_iter to also check for the warning, and remove a test which becomes redundant with this modification. See scikit-learn#10866.
…re_warn With the add of convergence warning in logistic regression, this test failed. See scikit-learn#10866
For lbfgs and liblinears solvers, the convergence warnings appeared only when verbose was greater than 0, whereas they appeared with verbose = 0 with other solvers. Create test to check the convergence warning in logistic regression and in linear svm. Update `test_search` to ignore this convergence warning. Fixes scikit-learn#10866
For lbfgs and liblinears solvers, the convergence warnings appeared only when verbose was greater than 0, whereas they appeared with verbose = 0 with other solvers. Create test to check the convergence warning in logistic regression and in linear svm. Update `test_search` to ignore this convergence warning. Fixes scikit-learn#10866
|
C:\ProgramData\Anaconda3\lib\site-packages\sklearn\linear_model_logistic.py:762: ConvergenceWarning: lbfgs failed to converge (status=1): Increase the number of iterations (max_iter) or scale the data as shown in: |
|
please solve this issue with right algorothms |
Uh oh!
There was an error while loading. Please reload this page.
Description
I've run LogisticRegressionCV on the Wisconsin Breast Cancer data, and the output of clf.n_iter_ was 100 for all but 1 of the variables. The default of 100 iterations was probably not sufficient in this case. Should there not be some kind of warning? I have done some tests and ~3000 iterations was probably a better choice for max_iter...
Steps/Code to Reproduce
Expected Results
Some kind of error to be shown. E.g: "result did not converge, try increasing the maximum number of iterations (max_iter)"
Versions
The text was updated successfully, but these errors were encountered: