TST Change expected result type np.int64 -> np.int #18089
Conversation
There was a problem hiding this comment.
Thanks @ckastner !
We be good to run this test in CI, likely by specifying the version of pandas to install so it is not skipped
scikit-learn/azure-pipelines.yml
Line 148 in 6e0ae8a
sklearn/datasets/tests/test_base.py
Outdated
| @@ -186,12 +186,12 @@ def test_loader(loader_func, data_shape, target_shape, n_target, has_descr, | |||
|
|
|||
|
|
|||
| @pytest.mark.parametrize("loader_func, data_dtype, target_dtype", [ | |||
| (load_breast_cancer, np.float64, np.int64), | |||
| (load_breast_cancer, np.float64, np.int), | |||
There was a problem hiding this comment.
np.int is being deprecated in numpy. Does the following work?
| (load_breast_cancer, np.float64, np.int), | |
| (load_breast_cancer, np.float64, np.int_), |
There was a problem hiding this comment.
Actually, my use of np.int was mistaken anyway. The data-loading function uses plain (builtin) int, not np.int as I originally claimed:
scikit-learn/sklearn/datasets/_base.py
Line 267 in 8f09e33
I propose switching to that (PR updated), so that the data-loading function and the test are consistent. If you'd prefer np.int_, please let me know. Sorry about the confusion.
On a side note: grep -R 'np\.int[^_1368a-z]' sklearn/* gives me ~100 occurrences of np.int in the codebase which will probably need to be updated at some point, along with the other deprecated types.
|
As for the (I hope we do not run into 32 bit + pandas issues) |
The function being tested uses int, so match that in the test. The test otherwise fails on 32-bit architectures. Closes: scikit-learn#18084
To add a data point, the packages we build for Debian are built on a number of 32-bit architectures (x86, ARM, MIPS), and all builds use pandas. All but very few tests pass, and IIRC the few failing tests have indications other than pandas. |
|
Thanks for the information! Could you still enable pandas in the 32 bit build? It's not currently in the diff.. |
|
Sorry about that, I thought that comment wasn't meant for me. Enabled and pushed. Just for my own understanding: I assume this enabling is limited to this particular PR and won't be merged? Because if the intention is to merge, then it should probably also be enabled for the 64-bit build, where it is currently disabled. |
|
The change did not have an effect on CI, the tests involving pandas were skipped again. I assume that deploying this change involves further steps? |
| @@ -50,7 +50,8 @@ elif [[ "$DISTRIB" == "ubuntu" ]]; then | |||
|
|
|||
| elif [[ "$DISTRIB" == "ubuntu-32" ]]; then | |||
| apt-get update | |||
| apt-get install -y python3-dev python3-scipy python3-matplotlib libatlas3-base libatlas-base-dev python3-virtualenv | |||
| apt-get install -y python3-dev python3-scipy python3-matplotlib libatlas3-base libatlas-base-dev python3-virtualenv python3-pandas | |||
There was a problem hiding this comment.
We use numpy from apt-get thus we need to use pandas from there as well.
| @@ -1200,7 +1200,7 @@ def test_check_fit_params(indices): | |||
| def test_check_sparse_pandas_sp_format(sp_format): | |||
| # check_array converts pandas dataframe with only sparse arrays into | |||
| # sparse matrix | |||
| pd = pytest.importorskip("pandas") | |||
| pd = pytest.importorskip("pandas", minversion="0.25.0") | |||
There was a problem hiding this comment.
sparse support was added in 0.25 and the pandas version installed on 18.04 is 0.22.0.
|
Error in CI is for OSX which is a timeout issue that is not related. |
Co-authored-by: Thomas J. Fan <[email protected]>
The function being tested uses np.int, so match that in the test. The
test otherwise fails on 32-bit architectures.
Closes: #18084