Closed
Description
It looks like there is a regression in LinearSVC
. I ran the bench_mnist.py
script in #3939 and compared it against ##3204, where the main difference is that #3939 tracks master.
The fourier_approx_svm results are much worse on master and take longer.
Classifier train-time test-time error-rate
fourier_approx_svm 139.248281956 1.81456899643 0.0487 (old)
fourier_approx_svm 231.988574028 1.94094800949 0.072 (master)
On the old branch I get
sklearn/svm/base.py:731: ConvergenceWarning: Liblinear failed to converge, increase the number of iterations.
"the number of iterations.", ConvergenceWarning)
which I don't get on master...