assert GPU CPU intercept_ equal when fit_intercept is false in cuml.LogisticRegression #5567
Conversation
lijinf2
added
breaking
lijinf2
force-pushed
the
lrsg_intercept
branch
from
2800f6c to
18acacf
Compare
|
/ok to test |
| if fit_intercept is False: | ||
| assert np.array_equal(culog.intercept_, sklog.intercept_) |
There was a problem hiding this comment.
I thought the point of this change was to ensure that intercept_ is equal in both cases. Why are we only testing for equality when fit_intercept is False ?
There was a problem hiding this comment.
When fit_intercept is True, intercept_ can be a list of non-zero. It is possible that culog.intercept_ and sklog.intercept_ are different, given their implementations are not exactly the same.
When fit_intercept is False, intercept_ is a list of zero. The test case is able to assert equal the culog.intercept_ and sklog.intercept_.
There was a problem hiding this comment.
But they would still be similar, no? You can use the cuml.testing.utils.array_equal function like here.
There was a problem hiding this comment.
Had tried assert array_equal(culog.intercept_, sklog.intercept_, 1e-3, with_sign=True).
Some tests failed:
FAILED tests/test_linear_model.py::test_logistic_regression[column_info0-1000-2-float32-none-1.0-True-1.0-0.001] - AssertionError: assert <array_equal: [0.60477465] [0.6086404] unit_tol=0.001 total_tol=0.0001 with_sign=True>
FAILED tests/test_linear_model.py::test_logistic_regression[column_info0-1000-10-float32-l1-1.0-True-1.0-0.001] - AssertionError: assert <array_equal: [-0.15404558 0.10340142 0.12519604 ... -0.13840534 0.13014889 -0.01927175] [-0.14051172 0.11931933 0.14109698 ... -0.12780005 0.14333196 -0.0074189 ] unit_tol=0.001 t...
There was a problem hiding this comment.
Mean accuracies of culog and sklog are the same though. It seems the models converged to different optima.
lijinf2
force-pushed
the
lrsg_intercept
branch
from
18acacf to
c251293
Compare
|
/ok to test |
lijinf2
force-pushed
the
lrsg_intercept
branch
from
c251293 to
60fd4ec
Compare
lijinf2
force-pushed
the
lrsg_intercept
branch
2 times, most recently
from
d75f99e to
8e69f1b
Compare
|
/ok to test |
lijinf2
force-pushed
the
lrsg_intercept
branch
from
8e69f1b to
f284d4f
Compare
| if fit_intercept is False: | ||
| assert np.array_equal(culog.intercept_, sklog.intercept_) |
Also adopted the code structure of the SG class to prepare for future PRs. This PR depends on and has included [PR 5567](#5567) Authors: - Jinfeng Li (https://github.com/lijinf2) Approvers: - Dante Gama Dessavre (https://github.com/dantegd) URL: #5558
|
Closed this PR because it has been merged as part of the follow-up PR: #5558 |
There is a discrepancy in intercept_ between cuml.linear_model.LogisticRegression and sklearn.linear_model.LogisticRegression. When fit_intercept=False and n_classes > 2, GPU intercept_ has shape (1, ) while CPU intercept_ has shape (n_classes,). This PR revises QN class to resolve the discrepancy.