Add files using upload-large-folder tool
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +9 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_agglomerative.py +1333 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_bicluster.py +628 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_feature_agglomeration.py +75 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_hierarchical_fast.cpython-310-x86_64-linux-gnu.so +3 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_k_means_common.cpython-310-x86_64-linux-gnu.so +3 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_k_means_common.pxd +19 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_k_means_elkan.cpython-310-x86_64-linux-gnu.so +3 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_mean_shift.py +536 -0
- videollama2/lib/python3.10/site-packages/sklearn/cluster/_spectral.py +791 -0
- videollama2/lib/python3.10/site-packages/sklearn/ensemble/__pycache__/_base.cpython-310.pyc +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/ensemble/__pycache__/_iforest.cpython-310.pyc +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/feature_selection/__init__.py +53 -0
- videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_from_model.py +444 -0
- videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_mutual_info.py +468 -0
- videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_rfe.py +765 -0
- videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_sequential.py +325 -0
- videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_univariate_selection.py +1052 -0
- videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_variance_threshold.py +135 -0
- videollama2/lib/python3.10/site-packages/sklearn/feature_selection/tests/test_feature_select.py +946 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/_classes.py +1769 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/_splitter.pxd +100 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__init__.py +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__pycache__/__init__.cpython-310.pyc +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__pycache__/test_export.cpython-310.pyc +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__pycache__/test_reingold_tilford.cpython-310.pyc +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__pycache__/test_tree.cpython-310.pyc +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/tests/test_export.py +512 -0
- videollama2/lib/python3.10/site-packages/sklearn/tree/tests/test_tree.py +2430 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_arpack.py +30 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_array_api.py +232 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_joblib.py +31 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_logistic_sigmoid.cpython-310-x86_64-linux-gnu.so +3 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_mask.py +62 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_openmp_helpers.cpython-310-x86_64-linux-gnu.so +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_openmp_helpers.pxd +6 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_param_validation.py +943 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_random.cpython-310-x86_64-linux-gnu.so +3 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_readonly_array_wrapper.cpython-310-x86_64-linux-gnu.so +3 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_seq_dataset.cpython-310-x86_64-linux-gnu.so +3 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_show_versions.py +108 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_sorting.cpython-310-x86_64-linux-gnu.so +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_tags.py +67 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_testing.py +1057 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_typedefs.pxd +28 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/_vector_sentinel.pxd +12 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/discovery.py +217 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/estimator_checks.py +0 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/murmurhash.cpython-310-x86_64-linux-gnu.so +3 -0
- videollama2/lib/python3.10/site-packages/sklearn/utils/parallel.py +123 -0
.gitattributes
CHANGED
|
@@ -1915,3 +1915,12 @@ videollama2/lib/python3.10/site-packages/sklearn/tree/_tree.cpython-310-x86_64-l
|
|
| 1915 |
videollama2/lib/python3.10/site-packages/sklearn/ensemble/_hist_gradient_boosting/_binning.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1916 |
videollama2/lib/python3.10/site-packages/sklearn/utils/arrayfuncs.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1917 |
videollama2/lib/python3.10/site-packages/sklearn/utils/_cython_blas.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1915 |
videollama2/lib/python3.10/site-packages/sklearn/ensemble/_hist_gradient_boosting/_binning.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1916 |
videollama2/lib/python3.10/site-packages/sklearn/utils/arrayfuncs.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1917 |
videollama2/lib/python3.10/site-packages/sklearn/utils/_cython_blas.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1918 |
+
videollama2/lib/python3.10/site-packages/sklearn/utils/_logistic_sigmoid.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1919 |
+
videollama2/lib/python3.10/site-packages/sklearn/utils/murmurhash.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1920 |
+
videollama2/lib/python3.10/site-packages/sklearn/utils/_random.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1921 |
+
videollama2/lib/python3.10/site-packages/sklearn/utils/sparsefuncs_fast.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1922 |
+
videollama2/lib/python3.10/site-packages/sklearn/utils/_seq_dataset.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1923 |
+
videollama2/lib/python3.10/site-packages/sklearn/utils/_readonly_array_wrapper.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1924 |
+
videollama2/lib/python3.10/site-packages/sklearn/cluster/_k_means_elkan.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1925 |
+
videollama2/lib/python3.10/site-packages/sklearn/cluster/_k_means_common.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1926 |
+
videollama2/lib/python3.10/site-packages/sklearn/cluster/_hierarchical_fast.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_agglomerative.py
ADDED
|
@@ -0,0 +1,1333 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Hierarchical Agglomerative Clustering
|
| 2 |
+
|
| 3 |
+
These routines perform some hierarchical agglomerative clustering of some
|
| 4 |
+
input data.
|
| 5 |
+
|
| 6 |
+
Authors : Vincent Michel, Bertrand Thirion, Alexandre Gramfort,
|
| 7 |
+
Gael Varoquaux
|
| 8 |
+
License: BSD 3 clause
|
| 9 |
+
"""
|
| 10 |
+
import warnings
|
| 11 |
+
from heapq import heapify, heappop, heappush, heappushpop
|
| 12 |
+
from numbers import Integral, Real
|
| 13 |
+
|
| 14 |
+
import numpy as np
|
| 15 |
+
from scipy import sparse
|
| 16 |
+
from scipy.sparse.csgraph import connected_components
|
| 17 |
+
|
| 18 |
+
from ..base import BaseEstimator, ClusterMixin, ClassNamePrefixFeaturesOutMixin
|
| 19 |
+
from ..metrics.pairwise import paired_distances
|
| 20 |
+
from ..metrics.pairwise import _VALID_METRICS
|
| 21 |
+
from ..metrics import DistanceMetric
|
| 22 |
+
from ..metrics._dist_metrics import METRIC_MAPPING
|
| 23 |
+
from ..utils import check_array
|
| 24 |
+
from ..utils._fast_dict import IntFloatDict
|
| 25 |
+
from ..utils.graph import _fix_connected_components
|
| 26 |
+
from ..utils._param_validation import Hidden, Interval, StrOptions, HasMethods
|
| 27 |
+
from ..utils.validation import check_memory
|
| 28 |
+
|
| 29 |
+
# mypy error: Module 'sklearn.cluster' has no attribute '_hierarchical_fast'
|
| 30 |
+
from . import _hierarchical_fast as _hierarchical # type: ignore
|
| 31 |
+
from ._feature_agglomeration import AgglomerationTransform
|
| 32 |
+
|
| 33 |
+
###############################################################################
|
| 34 |
+
# For non fully-connected graphs
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def _fix_connectivity(X, connectivity, affinity):
|
| 38 |
+
"""
|
| 39 |
+
Fixes the connectivity matrix.
|
| 40 |
+
|
| 41 |
+
The different steps are:
|
| 42 |
+
|
| 43 |
+
- copies it
|
| 44 |
+
- makes it symmetric
|
| 45 |
+
- converts it to LIL if necessary
|
| 46 |
+
- completes it if necessary.
|
| 47 |
+
|
| 48 |
+
Parameters
|
| 49 |
+
----------
|
| 50 |
+
X : array-like of shape (n_samples, n_features)
|
| 51 |
+
Feature matrix representing `n_samples` samples to be clustered.
|
| 52 |
+
|
| 53 |
+
connectivity : sparse matrix, default=None
|
| 54 |
+
Connectivity matrix. Defines for each sample the neighboring samples
|
| 55 |
+
following a given structure of the data. The matrix is assumed to
|
| 56 |
+
be symmetric and only the upper triangular half is used.
|
| 57 |
+
Default is `None`, i.e, the Ward algorithm is unstructured.
|
| 58 |
+
|
| 59 |
+
affinity : {"euclidean", "precomputed"}, default="euclidean"
|
| 60 |
+
Which affinity to use. At the moment `precomputed` and
|
| 61 |
+
``euclidean`` are supported. `euclidean` uses the
|
| 62 |
+
negative squared Euclidean distance between points.
|
| 63 |
+
|
| 64 |
+
Returns
|
| 65 |
+
-------
|
| 66 |
+
connectivity : sparse matrix
|
| 67 |
+
The fixed connectivity matrix.
|
| 68 |
+
|
| 69 |
+
n_connected_components : int
|
| 70 |
+
The number of connected components in the graph.
|
| 71 |
+
"""
|
| 72 |
+
n_samples = X.shape[0]
|
| 73 |
+
if connectivity.shape[0] != n_samples or connectivity.shape[1] != n_samples:
|
| 74 |
+
raise ValueError(
|
| 75 |
+
"Wrong shape for connectivity matrix: %s when X is %s"
|
| 76 |
+
% (connectivity.shape, X.shape)
|
| 77 |
+
)
|
| 78 |
+
|
| 79 |
+
# Make the connectivity matrix symmetric:
|
| 80 |
+
connectivity = connectivity + connectivity.T
|
| 81 |
+
|
| 82 |
+
# Convert connectivity matrix to LIL
|
| 83 |
+
if not sparse.isspmatrix_lil(connectivity):
|
| 84 |
+
if not sparse.isspmatrix(connectivity):
|
| 85 |
+
connectivity = sparse.lil_matrix(connectivity)
|
| 86 |
+
else:
|
| 87 |
+
connectivity = connectivity.tolil()
|
| 88 |
+
|
| 89 |
+
# Compute the number of nodes
|
| 90 |
+
n_connected_components, labels = connected_components(connectivity)
|
| 91 |
+
|
| 92 |
+
if n_connected_components > 1:
|
| 93 |
+
warnings.warn(
|
| 94 |
+
"the number of connected components of the "
|
| 95 |
+
"connectivity matrix is %d > 1. Completing it to avoid "
|
| 96 |
+
"stopping the tree early." % n_connected_components,
|
| 97 |
+
stacklevel=2,
|
| 98 |
+
)
|
| 99 |
+
# XXX: Can we do without completing the matrix?
|
| 100 |
+
connectivity = _fix_connected_components(
|
| 101 |
+
X=X,
|
| 102 |
+
graph=connectivity,
|
| 103 |
+
n_connected_components=n_connected_components,
|
| 104 |
+
component_labels=labels,
|
| 105 |
+
metric=affinity,
|
| 106 |
+
mode="connectivity",
|
| 107 |
+
)
|
| 108 |
+
|
| 109 |
+
return connectivity, n_connected_components
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
def _single_linkage_tree(
|
| 113 |
+
connectivity,
|
| 114 |
+
n_samples,
|
| 115 |
+
n_nodes,
|
| 116 |
+
n_clusters,
|
| 117 |
+
n_connected_components,
|
| 118 |
+
return_distance,
|
| 119 |
+
):
|
| 120 |
+
"""
|
| 121 |
+
Perform single linkage clustering on sparse data via the minimum
|
| 122 |
+
spanning tree from scipy.sparse.csgraph, then using union-find to label.
|
| 123 |
+
The parent array is then generated by walking through the tree.
|
| 124 |
+
"""
|
| 125 |
+
from scipy.sparse.csgraph import minimum_spanning_tree
|
| 126 |
+
|
| 127 |
+
# explicitly cast connectivity to ensure safety
|
| 128 |
+
connectivity = connectivity.astype(np.float64, copy=False)
|
| 129 |
+
|
| 130 |
+
# Ensure zero distances aren't ignored by setting them to "epsilon"
|
| 131 |
+
epsilon_value = np.finfo(dtype=connectivity.data.dtype).eps
|
| 132 |
+
connectivity.data[connectivity.data == 0] = epsilon_value
|
| 133 |
+
|
| 134 |
+
# Use scipy.sparse.csgraph to generate a minimum spanning tree
|
| 135 |
+
mst = minimum_spanning_tree(connectivity.tocsr())
|
| 136 |
+
|
| 137 |
+
# Convert the graph to scipy.cluster.hierarchy array format
|
| 138 |
+
mst = mst.tocoo()
|
| 139 |
+
|
| 140 |
+
# Undo the epsilon values
|
| 141 |
+
mst.data[mst.data == epsilon_value] = 0
|
| 142 |
+
|
| 143 |
+
mst_array = np.vstack([mst.row, mst.col, mst.data]).T
|
| 144 |
+
|
| 145 |
+
# Sort edges of the min_spanning_tree by weight
|
| 146 |
+
mst_array = mst_array[np.argsort(mst_array.T[2], kind="mergesort"), :]
|
| 147 |
+
|
| 148 |
+
# Convert edge list into standard hierarchical clustering format
|
| 149 |
+
single_linkage_tree = _hierarchical._single_linkage_label(mst_array)
|
| 150 |
+
children_ = single_linkage_tree[:, :2].astype(int)
|
| 151 |
+
|
| 152 |
+
# Compute parents
|
| 153 |
+
parent = np.arange(n_nodes, dtype=np.intp)
|
| 154 |
+
for i, (left, right) in enumerate(children_, n_samples):
|
| 155 |
+
if n_clusters is not None and i >= n_nodes:
|
| 156 |
+
break
|
| 157 |
+
if left < n_nodes:
|
| 158 |
+
parent[left] = i
|
| 159 |
+
if right < n_nodes:
|
| 160 |
+
parent[right] = i
|
| 161 |
+
|
| 162 |
+
if return_distance:
|
| 163 |
+
distances = single_linkage_tree[:, 2]
|
| 164 |
+
return children_, n_connected_components, n_samples, parent, distances
|
| 165 |
+
return children_, n_connected_components, n_samples, parent
|
| 166 |
+
|
| 167 |
+
|
| 168 |
+
###############################################################################
|
| 169 |
+
# Hierarchical tree building functions
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
def ward_tree(X, *, connectivity=None, n_clusters=None, return_distance=False):
|
| 173 |
+
"""Ward clustering based on a Feature matrix.
|
| 174 |
+
|
| 175 |
+
Recursively merges the pair of clusters that minimally increases
|
| 176 |
+
within-cluster variance.
|
| 177 |
+
|
| 178 |
+
The inertia matrix uses a Heapq-based representation.
|
| 179 |
+
|
| 180 |
+
This is the structured version, that takes into account some topological
|
| 181 |
+
structure between samples.
|
| 182 |
+
|
| 183 |
+
Read more in the :ref:`User Guide <hierarchical_clustering>`.
|
| 184 |
+
|
| 185 |
+
Parameters
|
| 186 |
+
----------
|
| 187 |
+
X : array-like of shape (n_samples, n_features)
|
| 188 |
+
Feature matrix representing `n_samples` samples to be clustered.
|
| 189 |
+
|
| 190 |
+
connectivity : sparse matrix, default=None
|
| 191 |
+
Connectivity matrix. Defines for each sample the neighboring samples
|
| 192 |
+
following a given structure of the data. The matrix is assumed to
|
| 193 |
+
be symmetric and only the upper triangular half is used.
|
| 194 |
+
Default is None, i.e, the Ward algorithm is unstructured.
|
| 195 |
+
|
| 196 |
+
n_clusters : int, default=None
|
| 197 |
+
`n_clusters` should be less than `n_samples`. Stop early the
|
| 198 |
+
construction of the tree at `n_clusters.` This is useful to decrease
|
| 199 |
+
computation time if the number of clusters is not small compared to the
|
| 200 |
+
number of samples. In this case, the complete tree is not computed, thus
|
| 201 |
+
the 'children' output is of limited use, and the 'parents' output should
|
| 202 |
+
rather be used. This option is valid only when specifying a connectivity
|
| 203 |
+
matrix.
|
| 204 |
+
|
| 205 |
+
return_distance : bool, default=False
|
| 206 |
+
If `True`, return the distance between the clusters.
|
| 207 |
+
|
| 208 |
+
Returns
|
| 209 |
+
-------
|
| 210 |
+
children : ndarray of shape (n_nodes-1, 2)
|
| 211 |
+
The children of each non-leaf node. Values less than `n_samples`
|
| 212 |
+
correspond to leaves of the tree which are the original samples.
|
| 213 |
+
A node `i` greater than or equal to `n_samples` is a non-leaf
|
| 214 |
+
node and has children `children_[i - n_samples]`. Alternatively
|
| 215 |
+
at the i-th iteration, children[i][0] and children[i][1]
|
| 216 |
+
are merged to form node `n_samples + i`.
|
| 217 |
+
|
| 218 |
+
n_connected_components : int
|
| 219 |
+
The number of connected components in the graph.
|
| 220 |
+
|
| 221 |
+
n_leaves : int
|
| 222 |
+
The number of leaves in the tree.
|
| 223 |
+
|
| 224 |
+
parents : ndarray of shape (n_nodes,) or None
|
| 225 |
+
The parent of each node. Only returned when a connectivity matrix
|
| 226 |
+
is specified, elsewhere 'None' is returned.
|
| 227 |
+
|
| 228 |
+
distances : ndarray of shape (n_nodes-1,)
|
| 229 |
+
Only returned if `return_distance` is set to `True` (for compatibility).
|
| 230 |
+
The distances between the centers of the nodes. `distances[i]`
|
| 231 |
+
corresponds to a weighted Euclidean distance between
|
| 232 |
+
the nodes `children[i, 1]` and `children[i, 2]`. If the nodes refer to
|
| 233 |
+
leaves of the tree, then `distances[i]` is their unweighted Euclidean
|
| 234 |
+
distance. Distances are updated in the following way
|
| 235 |
+
(from scipy.hierarchy.linkage):
|
| 236 |
+
|
| 237 |
+
The new entry :math:`d(u,v)` is computed as follows,
|
| 238 |
+
|
| 239 |
+
.. math::
|
| 240 |
+
|
| 241 |
+
d(u,v) = \\sqrt{\\frac{|v|+|s|}
|
| 242 |
+
{T}d(v,s)^2
|
| 243 |
+
+ \\frac{|v|+|t|}
|
| 244 |
+
{T}d(v,t)^2
|
| 245 |
+
- \\frac{|v|}
|
| 246 |
+
{T}d(s,t)^2}
|
| 247 |
+
|
| 248 |
+
where :math:`u` is the newly joined cluster consisting of
|
| 249 |
+
clusters :math:`s` and :math:`t`, :math:`v` is an unused
|
| 250 |
+
cluster in the forest, :math:`T=|v|+|s|+|t|`, and
|
| 251 |
+
:math:`|*|` is the cardinality of its argument. This is also
|
| 252 |
+
known as the incremental algorithm.
|
| 253 |
+
"""
|
| 254 |
+
X = np.asarray(X)
|
| 255 |
+
if X.ndim == 1:
|
| 256 |
+
X = np.reshape(X, (-1, 1))
|
| 257 |
+
n_samples, n_features = X.shape
|
| 258 |
+
|
| 259 |
+
if connectivity is None:
|
| 260 |
+
from scipy.cluster import hierarchy # imports PIL
|
| 261 |
+
|
| 262 |
+
if n_clusters is not None:
|
| 263 |
+
warnings.warn(
|
| 264 |
+
"Partial build of the tree is implemented "
|
| 265 |
+
"only for structured clustering (i.e. with "
|
| 266 |
+
"explicit connectivity). The algorithm "
|
| 267 |
+
"will build the full tree and only "
|
| 268 |
+
"retain the lower branches required "
|
| 269 |
+
"for the specified number of clusters",
|
| 270 |
+
stacklevel=2,
|
| 271 |
+
)
|
| 272 |
+
X = np.require(X, requirements="W")
|
| 273 |
+
out = hierarchy.ward(X)
|
| 274 |
+
children_ = out[:, :2].astype(np.intp)
|
| 275 |
+
|
| 276 |
+
if return_distance:
|
| 277 |
+
distances = out[:, 2]
|
| 278 |
+
return children_, 1, n_samples, None, distances
|
| 279 |
+
else:
|
| 280 |
+
return children_, 1, n_samples, None
|
| 281 |
+
|
| 282 |
+
connectivity, n_connected_components = _fix_connectivity(
|
| 283 |
+
X, connectivity, affinity="euclidean"
|
| 284 |
+
)
|
| 285 |
+
if n_clusters is None:
|
| 286 |
+
n_nodes = 2 * n_samples - 1
|
| 287 |
+
else:
|
| 288 |
+
if n_clusters > n_samples:
|
| 289 |
+
raise ValueError(
|
| 290 |
+
"Cannot provide more clusters than samples. "
|
| 291 |
+
"%i n_clusters was asked, and there are %i "
|
| 292 |
+
"samples." % (n_clusters, n_samples)
|
| 293 |
+
)
|
| 294 |
+
n_nodes = 2 * n_samples - n_clusters
|
| 295 |
+
|
| 296 |
+
# create inertia matrix
|
| 297 |
+
coord_row = []
|
| 298 |
+
coord_col = []
|
| 299 |
+
A = []
|
| 300 |
+
for ind, row in enumerate(connectivity.rows):
|
| 301 |
+
A.append(row)
|
| 302 |
+
# We keep only the upper triangular for the moments
|
| 303 |
+
# Generator expressions are faster than arrays on the following
|
| 304 |
+
row = [i for i in row if i < ind]
|
| 305 |
+
coord_row.extend(
|
| 306 |
+
len(row)
|
| 307 |
+
* [
|
| 308 |
+
ind,
|
| 309 |
+
]
|
| 310 |
+
)
|
| 311 |
+
coord_col.extend(row)
|
| 312 |
+
|
| 313 |
+
coord_row = np.array(coord_row, dtype=np.intp, order="C")
|
| 314 |
+
coord_col = np.array(coord_col, dtype=np.intp, order="C")
|
| 315 |
+
|
| 316 |
+
# build moments as a list
|
| 317 |
+
moments_1 = np.zeros(n_nodes, order="C")
|
| 318 |
+
moments_1[:n_samples] = 1
|
| 319 |
+
moments_2 = np.zeros((n_nodes, n_features), order="C")
|
| 320 |
+
moments_2[:n_samples] = X
|
| 321 |
+
inertia = np.empty(len(coord_row), dtype=np.float64, order="C")
|
| 322 |
+
_hierarchical.compute_ward_dist(moments_1, moments_2, coord_row, coord_col, inertia)
|
| 323 |
+
inertia = list(zip(inertia, coord_row, coord_col))
|
| 324 |
+
heapify(inertia)
|
| 325 |
+
|
| 326 |
+
# prepare the main fields
|
| 327 |
+
parent = np.arange(n_nodes, dtype=np.intp)
|
| 328 |
+
used_node = np.ones(n_nodes, dtype=bool)
|
| 329 |
+
children = []
|
| 330 |
+
if return_distance:
|
| 331 |
+
distances = np.empty(n_nodes - n_samples)
|
| 332 |
+
|
| 333 |
+
not_visited = np.empty(n_nodes, dtype=np.int8, order="C")
|
| 334 |
+
|
| 335 |
+
# recursive merge loop
|
| 336 |
+
for k in range(n_samples, n_nodes):
|
| 337 |
+
# identify the merge
|
| 338 |
+
while True:
|
| 339 |
+
inert, i, j = heappop(inertia)
|
| 340 |
+
if used_node[i] and used_node[j]:
|
| 341 |
+
break
|
| 342 |
+
parent[i], parent[j] = k, k
|
| 343 |
+
children.append((i, j))
|
| 344 |
+
used_node[i] = used_node[j] = False
|
| 345 |
+
if return_distance: # store inertia value
|
| 346 |
+
distances[k - n_samples] = inert
|
| 347 |
+
|
| 348 |
+
# update the moments
|
| 349 |
+
moments_1[k] = moments_1[i] + moments_1[j]
|
| 350 |
+
moments_2[k] = moments_2[i] + moments_2[j]
|
| 351 |
+
|
| 352 |
+
# update the structure matrix A and the inertia matrix
|
| 353 |
+
coord_col = []
|
| 354 |
+
not_visited.fill(1)
|
| 355 |
+
not_visited[k] = 0
|
| 356 |
+
_hierarchical._get_parents(A[i], coord_col, parent, not_visited)
|
| 357 |
+
_hierarchical._get_parents(A[j], coord_col, parent, not_visited)
|
| 358 |
+
# List comprehension is faster than a for loop
|
| 359 |
+
[A[col].append(k) for col in coord_col]
|
| 360 |
+
A.append(coord_col)
|
| 361 |
+
coord_col = np.array(coord_col, dtype=np.intp, order="C")
|
| 362 |
+
coord_row = np.empty(coord_col.shape, dtype=np.intp, order="C")
|
| 363 |
+
coord_row.fill(k)
|
| 364 |
+
n_additions = len(coord_row)
|
| 365 |
+
ini = np.empty(n_additions, dtype=np.float64, order="C")
|
| 366 |
+
|
| 367 |
+
_hierarchical.compute_ward_dist(moments_1, moments_2, coord_row, coord_col, ini)
|
| 368 |
+
|
| 369 |
+
# List comprehension is faster than a for loop
|
| 370 |
+
[heappush(inertia, (ini[idx], k, coord_col[idx])) for idx in range(n_additions)]
|
| 371 |
+
|
| 372 |
+
# Separate leaves in children (empty lists up to now)
|
| 373 |
+
n_leaves = n_samples
|
| 374 |
+
# sort children to get consistent output with unstructured version
|
| 375 |
+
children = [c[::-1] for c in children]
|
| 376 |
+
children = np.array(children) # return numpy array for efficient caching
|
| 377 |
+
|
| 378 |
+
if return_distance:
|
| 379 |
+
# 2 is scaling factor to compare w/ unstructured version
|
| 380 |
+
distances = np.sqrt(2.0 * distances)
|
| 381 |
+
return children, n_connected_components, n_leaves, parent, distances
|
| 382 |
+
else:
|
| 383 |
+
return children, n_connected_components, n_leaves, parent
|
| 384 |
+
|
| 385 |
+
|
| 386 |
+
# single average and complete linkage
|
| 387 |
+
def linkage_tree(
|
| 388 |
+
X,
|
| 389 |
+
connectivity=None,
|
| 390 |
+
n_clusters=None,
|
| 391 |
+
linkage="complete",
|
| 392 |
+
affinity="euclidean",
|
| 393 |
+
return_distance=False,
|
| 394 |
+
):
|
| 395 |
+
"""Linkage agglomerative clustering based on a Feature matrix.
|
| 396 |
+
|
| 397 |
+
The inertia matrix uses a Heapq-based representation.
|
| 398 |
+
|
| 399 |
+
This is the structured version, that takes into account some topological
|
| 400 |
+
structure between samples.
|
| 401 |
+
|
| 402 |
+
Read more in the :ref:`User Guide <hierarchical_clustering>`.
|
| 403 |
+
|
| 404 |
+
Parameters
|
| 405 |
+
----------
|
| 406 |
+
X : array-like of shape (n_samples, n_features)
|
| 407 |
+
Feature matrix representing `n_samples` samples to be clustered.
|
| 408 |
+
|
| 409 |
+
connectivity : sparse matrix, default=None
|
| 410 |
+
Connectivity matrix. Defines for each sample the neighboring samples
|
| 411 |
+
following a given structure of the data. The matrix is assumed to
|
| 412 |
+
be symmetric and only the upper triangular half is used.
|
| 413 |
+
Default is `None`, i.e, the Ward algorithm is unstructured.
|
| 414 |
+
|
| 415 |
+
n_clusters : int, default=None
|
| 416 |
+
Stop early the construction of the tree at `n_clusters`. This is
|
| 417 |
+
useful to decrease computation time if the number of clusters is
|
| 418 |
+
not small compared to the number of samples. In this case, the
|
| 419 |
+
complete tree is not computed, thus the 'children' output is of
|
| 420 |
+
limited use, and the 'parents' output should rather be used.
|
| 421 |
+
This option is valid only when specifying a connectivity matrix.
|
| 422 |
+
|
| 423 |
+
linkage : {"average", "complete", "single"}, default="complete"
|
| 424 |
+
Which linkage criteria to use. The linkage criterion determines which
|
| 425 |
+
distance to use between sets of observation.
|
| 426 |
+
- "average" uses the average of the distances of each observation of
|
| 427 |
+
the two sets.
|
| 428 |
+
- "complete" or maximum linkage uses the maximum distances between
|
| 429 |
+
all observations of the two sets.
|
| 430 |
+
- "single" uses the minimum of the distances between all
|
| 431 |
+
observations of the two sets.
|
| 432 |
+
|
| 433 |
+
affinity : str or callable, default='euclidean'
|
| 434 |
+
Which metric to use. Can be 'euclidean', 'manhattan', or any
|
| 435 |
+
distance known to paired distance (see metric.pairwise).
|
| 436 |
+
|
| 437 |
+
return_distance : bool, default=False
|
| 438 |
+
Whether or not to return the distances between the clusters.
|
| 439 |
+
|
| 440 |
+
Returns
|
| 441 |
+
-------
|
| 442 |
+
children : ndarray of shape (n_nodes-1, 2)
|
| 443 |
+
The children of each non-leaf node. Values less than `n_samples`
|
| 444 |
+
correspond to leaves of the tree which are the original samples.
|
| 445 |
+
A node `i` greater than or equal to `n_samples` is a non-leaf
|
| 446 |
+
node and has children `children_[i - n_samples]`. Alternatively
|
| 447 |
+
at the i-th iteration, children[i][0] and children[i][1]
|
| 448 |
+
are merged to form node `n_samples + i`.
|
| 449 |
+
|
| 450 |
+
n_connected_components : int
|
| 451 |
+
The number of connected components in the graph.
|
| 452 |
+
|
| 453 |
+
n_leaves : int
|
| 454 |
+
The number of leaves in the tree.
|
| 455 |
+
|
| 456 |
+
parents : ndarray of shape (n_nodes, ) or None
|
| 457 |
+
The parent of each node. Only returned when a connectivity matrix
|
| 458 |
+
is specified, elsewhere 'None' is returned.
|
| 459 |
+
|
| 460 |
+
distances : ndarray of shape (n_nodes-1,)
|
| 461 |
+
Returned when `return_distance` is set to `True`.
|
| 462 |
+
|
| 463 |
+
distances[i] refers to the distance between children[i][0] and
|
| 464 |
+
children[i][1] when they are merged.
|
| 465 |
+
|
| 466 |
+
See Also
|
| 467 |
+
--------
|
| 468 |
+
ward_tree : Hierarchical clustering with ward linkage.
|
| 469 |
+
"""
|
| 470 |
+
X = np.asarray(X)
|
| 471 |
+
if X.ndim == 1:
|
| 472 |
+
X = np.reshape(X, (-1, 1))
|
| 473 |
+
n_samples, n_features = X.shape
|
| 474 |
+
|
| 475 |
+
linkage_choices = {
|
| 476 |
+
"complete": _hierarchical.max_merge,
|
| 477 |
+
"average": _hierarchical.average_merge,
|
| 478 |
+
"single": None,
|
| 479 |
+
} # Single linkage is handled differently
|
| 480 |
+
try:
|
| 481 |
+
join_func = linkage_choices[linkage]
|
| 482 |
+
except KeyError as e:
|
| 483 |
+
raise ValueError(
|
| 484 |
+
"Unknown linkage option, linkage should be one of %s, but %s was given"
|
| 485 |
+
% (linkage_choices.keys(), linkage)
|
| 486 |
+
) from e
|
| 487 |
+
|
| 488 |
+
if affinity == "cosine" and np.any(~np.any(X, axis=1)):
|
| 489 |
+
raise ValueError("Cosine affinity cannot be used when X contains zero vectors")
|
| 490 |
+
|
| 491 |
+
if connectivity is None:
|
| 492 |
+
from scipy.cluster import hierarchy # imports PIL
|
| 493 |
+
|
| 494 |
+
if n_clusters is not None:
|
| 495 |
+
warnings.warn(
|
| 496 |
+
"Partial build of the tree is implemented "
|
| 497 |
+
"only for structured clustering (i.e. with "
|
| 498 |
+
"explicit connectivity). The algorithm "
|
| 499 |
+
"will build the full tree and only "
|
| 500 |
+
"retain the lower branches required "
|
| 501 |
+
"for the specified number of clusters",
|
| 502 |
+
stacklevel=2,
|
| 503 |
+
)
|
| 504 |
+
|
| 505 |
+
if affinity == "precomputed":
|
| 506 |
+
# for the linkage function of hierarchy to work on precomputed
|
| 507 |
+
# data, provide as first argument an ndarray of the shape returned
|
| 508 |
+
# by sklearn.metrics.pairwise_distances.
|
| 509 |
+
if X.shape[0] != X.shape[1]:
|
| 510 |
+
raise ValueError(
|
| 511 |
+
f"Distance matrix should be square, got matrix of shape {X.shape}"
|
| 512 |
+
)
|
| 513 |
+
i, j = np.triu_indices(X.shape[0], k=1)
|
| 514 |
+
X = X[i, j]
|
| 515 |
+
elif affinity == "l2":
|
| 516 |
+
# Translate to something understood by scipy
|
| 517 |
+
affinity = "euclidean"
|
| 518 |
+
elif affinity in ("l1", "manhattan"):
|
| 519 |
+
affinity = "cityblock"
|
| 520 |
+
elif callable(affinity):
|
| 521 |
+
X = affinity(X)
|
| 522 |
+
i, j = np.triu_indices(X.shape[0], k=1)
|
| 523 |
+
X = X[i, j]
|
| 524 |
+
if (
|
| 525 |
+
linkage == "single"
|
| 526 |
+
and affinity != "precomputed"
|
| 527 |
+
and not callable(affinity)
|
| 528 |
+
and affinity in METRIC_MAPPING
|
| 529 |
+
):
|
| 530 |
+
|
| 531 |
+
# We need the fast cythonized metric from neighbors
|
| 532 |
+
dist_metric = DistanceMetric.get_metric(affinity)
|
| 533 |
+
|
| 534 |
+
# The Cython routines used require contiguous arrays
|
| 535 |
+
X = np.ascontiguousarray(X, dtype=np.double)
|
| 536 |
+
|
| 537 |
+
mst = _hierarchical.mst_linkage_core(X, dist_metric)
|
| 538 |
+
# Sort edges of the min_spanning_tree by weight
|
| 539 |
+
mst = mst[np.argsort(mst.T[2], kind="mergesort"), :]
|
| 540 |
+
|
| 541 |
+
# Convert edge list into standard hierarchical clustering format
|
| 542 |
+
out = _hierarchical.single_linkage_label(mst)
|
| 543 |
+
else:
|
| 544 |
+
out = hierarchy.linkage(X, method=linkage, metric=affinity)
|
| 545 |
+
children_ = out[:, :2].astype(int, copy=False)
|
| 546 |
+
|
| 547 |
+
if return_distance:
|
| 548 |
+
distances = out[:, 2]
|
| 549 |
+
return children_, 1, n_samples, None, distances
|
| 550 |
+
return children_, 1, n_samples, None
|
| 551 |
+
|
| 552 |
+
connectivity, n_connected_components = _fix_connectivity(
|
| 553 |
+
X, connectivity, affinity=affinity
|
| 554 |
+
)
|
| 555 |
+
connectivity = connectivity.tocoo()
|
| 556 |
+
# Put the diagonal to zero
|
| 557 |
+
diag_mask = connectivity.row != connectivity.col
|
| 558 |
+
connectivity.row = connectivity.row[diag_mask]
|
| 559 |
+
connectivity.col = connectivity.col[diag_mask]
|
| 560 |
+
connectivity.data = connectivity.data[diag_mask]
|
| 561 |
+
del diag_mask
|
| 562 |
+
|
| 563 |
+
if affinity == "precomputed":
|
| 564 |
+
distances = X[connectivity.row, connectivity.col].astype(np.float64, copy=False)
|
| 565 |
+
else:
|
| 566 |
+
# FIXME We compute all the distances, while we could have only computed
|
| 567 |
+
# the "interesting" distances
|
| 568 |
+
distances = paired_distances(
|
| 569 |
+
X[connectivity.row], X[connectivity.col], metric=affinity
|
| 570 |
+
)
|
| 571 |
+
connectivity.data = distances
|
| 572 |
+
|
| 573 |
+
if n_clusters is None:
|
| 574 |
+
n_nodes = 2 * n_samples - 1
|
| 575 |
+
else:
|
| 576 |
+
assert n_clusters <= n_samples
|
| 577 |
+
n_nodes = 2 * n_samples - n_clusters
|
| 578 |
+
|
| 579 |
+
if linkage == "single":
|
| 580 |
+
return _single_linkage_tree(
|
| 581 |
+
connectivity,
|
| 582 |
+
n_samples,
|
| 583 |
+
n_nodes,
|
| 584 |
+
n_clusters,
|
| 585 |
+
n_connected_components,
|
| 586 |
+
return_distance,
|
| 587 |
+
)
|
| 588 |
+
|
| 589 |
+
if return_distance:
|
| 590 |
+
distances = np.empty(n_nodes - n_samples)
|
| 591 |
+
# create inertia heap and connection matrix
|
| 592 |
+
A = np.empty(n_nodes, dtype=object)
|
| 593 |
+
inertia = list()
|
| 594 |
+
|
| 595 |
+
# LIL seems to the best format to access the rows quickly,
|
| 596 |
+
# without the numpy overhead of slicing CSR indices and data.
|
| 597 |
+
connectivity = connectivity.tolil()
|
| 598 |
+
# We are storing the graph in a list of IntFloatDict
|
| 599 |
+
for ind, (data, row) in enumerate(zip(connectivity.data, connectivity.rows)):
|
| 600 |
+
A[ind] = IntFloatDict(
|
| 601 |
+
np.asarray(row, dtype=np.intp), np.asarray(data, dtype=np.float64)
|
| 602 |
+
)
|
| 603 |
+
# We keep only the upper triangular for the heap
|
| 604 |
+
# Generator expressions are faster than arrays on the following
|
| 605 |
+
inertia.extend(
|
| 606 |
+
_hierarchical.WeightedEdge(d, ind, r) for r, d in zip(row, data) if r < ind
|
| 607 |
+
)
|
| 608 |
+
del connectivity
|
| 609 |
+
|
| 610 |
+
heapify(inertia)
|
| 611 |
+
|
| 612 |
+
# prepare the main fields
|
| 613 |
+
parent = np.arange(n_nodes, dtype=np.intp)
|
| 614 |
+
used_node = np.ones(n_nodes, dtype=np.intp)
|
| 615 |
+
children = []
|
| 616 |
+
|
| 617 |
+
# recursive merge loop
|
| 618 |
+
for k in range(n_samples, n_nodes):
|
| 619 |
+
# identify the merge
|
| 620 |
+
while True:
|
| 621 |
+
edge = heappop(inertia)
|
| 622 |
+
if used_node[edge.a] and used_node[edge.b]:
|
| 623 |
+
break
|
| 624 |
+
i = edge.a
|
| 625 |
+
j = edge.b
|
| 626 |
+
|
| 627 |
+
if return_distance:
|
| 628 |
+
# store distances
|
| 629 |
+
distances[k - n_samples] = edge.weight
|
| 630 |
+
|
| 631 |
+
parent[i] = parent[j] = k
|
| 632 |
+
children.append((i, j))
|
| 633 |
+
# Keep track of the number of elements per cluster
|
| 634 |
+
n_i = used_node[i]
|
| 635 |
+
n_j = used_node[j]
|
| 636 |
+
used_node[k] = n_i + n_j
|
| 637 |
+
used_node[i] = used_node[j] = False
|
| 638 |
+
|
| 639 |
+
# update the structure matrix A and the inertia matrix
|
| 640 |
+
# a clever 'min', or 'max' operation between A[i] and A[j]
|
| 641 |
+
coord_col = join_func(A[i], A[j], used_node, n_i, n_j)
|
| 642 |
+
for col, d in coord_col:
|
| 643 |
+
A[col].append(k, d)
|
| 644 |
+
# Here we use the information from coord_col (containing the
|
| 645 |
+
# distances) to update the heap
|
| 646 |
+
heappush(inertia, _hierarchical.WeightedEdge(d, k, col))
|
| 647 |
+
A[k] = coord_col
|
| 648 |
+
# Clear A[i] and A[j] to save memory
|
| 649 |
+
A[i] = A[j] = 0
|
| 650 |
+
|
| 651 |
+
# Separate leaves in children (empty lists up to now)
|
| 652 |
+
n_leaves = n_samples
|
| 653 |
+
|
| 654 |
+
# # return numpy array for efficient caching
|
| 655 |
+
children = np.array(children)[:, ::-1]
|
| 656 |
+
|
| 657 |
+
if return_distance:
|
| 658 |
+
return children, n_connected_components, n_leaves, parent, distances
|
| 659 |
+
return children, n_connected_components, n_leaves, parent
|
| 660 |
+
|
| 661 |
+
|
| 662 |
+
# Matching names to tree-building strategies
|
| 663 |
+
def _complete_linkage(*args, **kwargs):
|
| 664 |
+
kwargs["linkage"] = "complete"
|
| 665 |
+
return linkage_tree(*args, **kwargs)
|
| 666 |
+
|
| 667 |
+
|
| 668 |
+
def _average_linkage(*args, **kwargs):
|
| 669 |
+
kwargs["linkage"] = "average"
|
| 670 |
+
return linkage_tree(*args, **kwargs)
|
| 671 |
+
|
| 672 |
+
|
| 673 |
+
def _single_linkage(*args, **kwargs):
|
| 674 |
+
kwargs["linkage"] = "single"
|
| 675 |
+
return linkage_tree(*args, **kwargs)
|
| 676 |
+
|
| 677 |
+
|
| 678 |
+
_TREE_BUILDERS = dict(
|
| 679 |
+
ward=ward_tree,
|
| 680 |
+
complete=_complete_linkage,
|
| 681 |
+
average=_average_linkage,
|
| 682 |
+
single=_single_linkage,
|
| 683 |
+
)
|
| 684 |
+
|
| 685 |
+
###############################################################################
|
| 686 |
+
# Functions for cutting hierarchical clustering tree
|
| 687 |
+
|
| 688 |
+
|
| 689 |
+
def _hc_cut(n_clusters, children, n_leaves):
|
| 690 |
+
"""Function cutting the ward tree for a given number of clusters.
|
| 691 |
+
|
| 692 |
+
Parameters
|
| 693 |
+
----------
|
| 694 |
+
n_clusters : int or ndarray
|
| 695 |
+
The number of clusters to form.
|
| 696 |
+
|
| 697 |
+
children : ndarray of shape (n_nodes-1, 2)
|
| 698 |
+
The children of each non-leaf node. Values less than `n_samples`
|
| 699 |
+
correspond to leaves of the tree which are the original samples.
|
| 700 |
+
A node `i` greater than or equal to `n_samples` is a non-leaf
|
| 701 |
+
node and has children `children_[i - n_samples]`. Alternatively
|
| 702 |
+
at the i-th iteration, children[i][0] and children[i][1]
|
| 703 |
+
are merged to form node `n_samples + i`.
|
| 704 |
+
|
| 705 |
+
n_leaves : int
|
| 706 |
+
Number of leaves of the tree.
|
| 707 |
+
|
| 708 |
+
Returns
|
| 709 |
+
-------
|
| 710 |
+
labels : array [n_samples]
|
| 711 |
+
Cluster labels for each point.
|
| 712 |
+
"""
|
| 713 |
+
if n_clusters > n_leaves:
|
| 714 |
+
raise ValueError(
|
| 715 |
+
"Cannot extract more clusters than samples: "
|
| 716 |
+
"%s clusters where given for a tree with %s leaves."
|
| 717 |
+
% (n_clusters, n_leaves)
|
| 718 |
+
)
|
| 719 |
+
# In this function, we store nodes as a heap to avoid recomputing
|
| 720 |
+
# the max of the nodes: the first element is always the smallest
|
| 721 |
+
# We use negated indices as heaps work on smallest elements, and we
|
| 722 |
+
# are interested in largest elements
|
| 723 |
+
# children[-1] is the root of the tree
|
| 724 |
+
nodes = [-(max(children[-1]) + 1)]
|
| 725 |
+
for _ in range(n_clusters - 1):
|
| 726 |
+
# As we have a heap, nodes[0] is the smallest element
|
| 727 |
+
these_children = children[-nodes[0] - n_leaves]
|
| 728 |
+
# Insert the 2 children and remove the largest node
|
| 729 |
+
heappush(nodes, -these_children[0])
|
| 730 |
+
heappushpop(nodes, -these_children[1])
|
| 731 |
+
label = np.zeros(n_leaves, dtype=np.intp)
|
| 732 |
+
for i, node in enumerate(nodes):
|
| 733 |
+
label[_hierarchical._hc_get_descendent(-node, children, n_leaves)] = i
|
| 734 |
+
return label
|
| 735 |
+
|
| 736 |
+
|
| 737 |
+
###############################################################################
|
| 738 |
+
|
| 739 |
+
|
| 740 |
+
class AgglomerativeClustering(ClusterMixin, BaseEstimator):
|
| 741 |
+
"""
|
| 742 |
+
Agglomerative Clustering.
|
| 743 |
+
|
| 744 |
+
Recursively merges pair of clusters of sample data; uses linkage distance.
|
| 745 |
+
|
| 746 |
+
Read more in the :ref:`User Guide <hierarchical_clustering>`.
|
| 747 |
+
|
| 748 |
+
Parameters
|
| 749 |
+
----------
|
| 750 |
+
n_clusters : int or None, default=2
|
| 751 |
+
The number of clusters to find. It must be ``None`` if
|
| 752 |
+
``distance_threshold`` is not ``None``.
|
| 753 |
+
|
| 754 |
+
affinity : str or callable, default='euclidean'
|
| 755 |
+
The metric to use when calculating distance between instances in a
|
| 756 |
+
feature array. If metric is a string or callable, it must be one of
|
| 757 |
+
the options allowed by :func:`sklearn.metrics.pairwise_distances` for
|
| 758 |
+
its metric parameter.
|
| 759 |
+
If linkage is "ward", only "euclidean" is accepted.
|
| 760 |
+
If "precomputed", a distance matrix (instead of a similarity matrix)
|
| 761 |
+
is needed as input for the fit method.
|
| 762 |
+
|
| 763 |
+
.. deprecated:: 1.2
|
| 764 |
+
`affinity` was deprecated in version 1.2 and will be renamed to
|
| 765 |
+
`metric` in 1.4.
|
| 766 |
+
|
| 767 |
+
metric : str or callable, default=None
|
| 768 |
+
Metric used to compute the linkage. Can be "euclidean", "l1", "l2",
|
| 769 |
+
"manhattan", "cosine", or "precomputed". If set to `None` then
|
| 770 |
+
"euclidean" is used. If linkage is "ward", only "euclidean" is
|
| 771 |
+
accepted. If "precomputed", a distance matrix is needed as input for
|
| 772 |
+
the fit method.
|
| 773 |
+
|
| 774 |
+
.. versionadded:: 1.2
|
| 775 |
+
|
| 776 |
+
memory : str or object with the joblib.Memory interface, default=None
|
| 777 |
+
Used to cache the output of the computation of the tree.
|
| 778 |
+
By default, no caching is done. If a string is given, it is the
|
| 779 |
+
path to the caching directory.
|
| 780 |
+
|
| 781 |
+
connectivity : array-like or callable, default=None
|
| 782 |
+
Connectivity matrix. Defines for each sample the neighboring
|
| 783 |
+
samples following a given structure of the data.
|
| 784 |
+
This can be a connectivity matrix itself or a callable that transforms
|
| 785 |
+
the data into a connectivity matrix, such as derived from
|
| 786 |
+
`kneighbors_graph`. Default is ``None``, i.e, the
|
| 787 |
+
hierarchical clustering algorithm is unstructured.
|
| 788 |
+
|
| 789 |
+
compute_full_tree : 'auto' or bool, default='auto'
|
| 790 |
+
Stop early the construction of the tree at ``n_clusters``. This is
|
| 791 |
+
useful to decrease computation time if the number of clusters is not
|
| 792 |
+
small compared to the number of samples. This option is useful only
|
| 793 |
+
when specifying a connectivity matrix. Note also that when varying the
|
| 794 |
+
number of clusters and using caching, it may be advantageous to compute
|
| 795 |
+
the full tree. It must be ``True`` if ``distance_threshold`` is not
|
| 796 |
+
``None``. By default `compute_full_tree` is "auto", which is equivalent
|
| 797 |
+
to `True` when `distance_threshold` is not `None` or that `n_clusters`
|
| 798 |
+
is inferior to the maximum between 100 or `0.02 * n_samples`.
|
| 799 |
+
Otherwise, "auto" is equivalent to `False`.
|
| 800 |
+
|
| 801 |
+
linkage : {'ward', 'complete', 'average', 'single'}, default='ward'
|
| 802 |
+
Which linkage criterion to use. The linkage criterion determines which
|
| 803 |
+
distance to use between sets of observation. The algorithm will merge
|
| 804 |
+
the pairs of cluster that minimize this criterion.
|
| 805 |
+
|
| 806 |
+
- 'ward' minimizes the variance of the clusters being merged.
|
| 807 |
+
- 'average' uses the average of the distances of each observation of
|
| 808 |
+
the two sets.
|
| 809 |
+
- 'complete' or 'maximum' linkage uses the maximum distances between
|
| 810 |
+
all observations of the two sets.
|
| 811 |
+
- 'single' uses the minimum of the distances between all observations
|
| 812 |
+
of the two sets.
|
| 813 |
+
|
| 814 |
+
.. versionadded:: 0.20
|
| 815 |
+
Added the 'single' option
|
| 816 |
+
|
| 817 |
+
distance_threshold : float, default=None
|
| 818 |
+
The linkage distance threshold at or above which clusters will not be
|
| 819 |
+
merged. If not ``None``, ``n_clusters`` must be ``None`` and
|
| 820 |
+
``compute_full_tree`` must be ``True``.
|
| 821 |
+
|
| 822 |
+
.. versionadded:: 0.21
|
| 823 |
+
|
| 824 |
+
compute_distances : bool, default=False
|
| 825 |
+
Computes distances between clusters even if `distance_threshold` is not
|
| 826 |
+
used. This can be used to make dendrogram visualization, but introduces
|
| 827 |
+
a computational and memory overhead.
|
| 828 |
+
|
| 829 |
+
.. versionadded:: 0.24
|
| 830 |
+
|
| 831 |
+
Attributes
|
| 832 |
+
----------
|
| 833 |
+
n_clusters_ : int
|
| 834 |
+
The number of clusters found by the algorithm. If
|
| 835 |
+
``distance_threshold=None``, it will be equal to the given
|
| 836 |
+
``n_clusters``.
|
| 837 |
+
|
| 838 |
+
labels_ : ndarray of shape (n_samples)
|
| 839 |
+
Cluster labels for each point.
|
| 840 |
+
|
| 841 |
+
n_leaves_ : int
|
| 842 |
+
Number of leaves in the hierarchical tree.
|
| 843 |
+
|
| 844 |
+
n_connected_components_ : int
|
| 845 |
+
The estimated number of connected components in the graph.
|
| 846 |
+
|
| 847 |
+
.. versionadded:: 0.21
|
| 848 |
+
``n_connected_components_`` was added to replace ``n_components_``.
|
| 849 |
+
|
| 850 |
+
n_features_in_ : int
|
| 851 |
+
Number of features seen during :term:`fit`.
|
| 852 |
+
|
| 853 |
+
.. versionadded:: 0.24
|
| 854 |
+
|
| 855 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 856 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 857 |
+
has feature names that are all strings.
|
| 858 |
+
|
| 859 |
+
.. versionadded:: 1.0
|
| 860 |
+
|
| 861 |
+
children_ : array-like of shape (n_samples-1, 2)
|
| 862 |
+
The children of each non-leaf node. Values less than `n_samples`
|
| 863 |
+
correspond to leaves of the tree which are the original samples.
|
| 864 |
+
A node `i` greater than or equal to `n_samples` is a non-leaf
|
| 865 |
+
node and has children `children_[i - n_samples]`. Alternatively
|
| 866 |
+
at the i-th iteration, children[i][0] and children[i][1]
|
| 867 |
+
are merged to form node `n_samples + i`.
|
| 868 |
+
|
| 869 |
+
distances_ : array-like of shape (n_nodes-1,)
|
| 870 |
+
Distances between nodes in the corresponding place in `children_`.
|
| 871 |
+
Only computed if `distance_threshold` is used or `compute_distances`
|
| 872 |
+
is set to `True`.
|
| 873 |
+
|
| 874 |
+
See Also
|
| 875 |
+
--------
|
| 876 |
+
FeatureAgglomeration : Agglomerative clustering but for features instead of
|
| 877 |
+
samples.
|
| 878 |
+
ward_tree : Hierarchical clustering with ward linkage.
|
| 879 |
+
|
| 880 |
+
Examples
|
| 881 |
+
--------
|
| 882 |
+
>>> from sklearn.cluster import AgglomerativeClustering
|
| 883 |
+
>>> import numpy as np
|
| 884 |
+
>>> X = np.array([[1, 2], [1, 4], [1, 0],
|
| 885 |
+
... [4, 2], [4, 4], [4, 0]])
|
| 886 |
+
>>> clustering = AgglomerativeClustering().fit(X)
|
| 887 |
+
>>> clustering
|
| 888 |
+
AgglomerativeClustering()
|
| 889 |
+
>>> clustering.labels_
|
| 890 |
+
array([1, 1, 1, 0, 0, 0])
|
| 891 |
+
"""
|
| 892 |
+
|
| 893 |
+
_parameter_constraints: dict = {
|
| 894 |
+
"n_clusters": [Interval(Integral, 1, None, closed="left"), None],
|
| 895 |
+
"affinity": [
|
| 896 |
+
Hidden(StrOptions({"deprecated"})),
|
| 897 |
+
StrOptions(set(_VALID_METRICS) | {"precomputed"}),
|
| 898 |
+
callable,
|
| 899 |
+
],
|
| 900 |
+
"metric": [
|
| 901 |
+
StrOptions(set(_VALID_METRICS) | {"precomputed"}),
|
| 902 |
+
callable,
|
| 903 |
+
None,
|
| 904 |
+
],
|
| 905 |
+
"memory": [str, HasMethods("cache"), None],
|
| 906 |
+
"connectivity": ["array-like", callable, None],
|
| 907 |
+
"compute_full_tree": [StrOptions({"auto"}), "boolean"],
|
| 908 |
+
"linkage": [StrOptions(set(_TREE_BUILDERS.keys()))],
|
| 909 |
+
"distance_threshold": [Interval(Real, 0, None, closed="left"), None],
|
| 910 |
+
"compute_distances": ["boolean"],
|
| 911 |
+
}
|
| 912 |
+
|
| 913 |
+
def __init__(
|
| 914 |
+
self,
|
| 915 |
+
n_clusters=2,
|
| 916 |
+
*,
|
| 917 |
+
affinity="deprecated", # TODO(1.4): Remove
|
| 918 |
+
metric=None, # TODO(1.4): Set to "euclidean"
|
| 919 |
+
memory=None,
|
| 920 |
+
connectivity=None,
|
| 921 |
+
compute_full_tree="auto",
|
| 922 |
+
linkage="ward",
|
| 923 |
+
distance_threshold=None,
|
| 924 |
+
compute_distances=False,
|
| 925 |
+
):
|
| 926 |
+
self.n_clusters = n_clusters
|
| 927 |
+
self.distance_threshold = distance_threshold
|
| 928 |
+
self.memory = memory
|
| 929 |
+
self.connectivity = connectivity
|
| 930 |
+
self.compute_full_tree = compute_full_tree
|
| 931 |
+
self.linkage = linkage
|
| 932 |
+
self.affinity = affinity
|
| 933 |
+
self.metric = metric
|
| 934 |
+
self.compute_distances = compute_distances
|
| 935 |
+
|
| 936 |
+
def fit(self, X, y=None):
|
| 937 |
+
"""Fit the hierarchical clustering from features, or distance matrix.
|
| 938 |
+
|
| 939 |
+
Parameters
|
| 940 |
+
----------
|
| 941 |
+
X : array-like, shape (n_samples, n_features) or \
|
| 942 |
+
(n_samples, n_samples)
|
| 943 |
+
Training instances to cluster, or distances between instances if
|
| 944 |
+
``metric='precomputed'``.
|
| 945 |
+
|
| 946 |
+
y : Ignored
|
| 947 |
+
Not used, present here for API consistency by convention.
|
| 948 |
+
|
| 949 |
+
Returns
|
| 950 |
+
-------
|
| 951 |
+
self : object
|
| 952 |
+
Returns the fitted instance.
|
| 953 |
+
"""
|
| 954 |
+
self._validate_params()
|
| 955 |
+
X = self._validate_data(X, ensure_min_samples=2)
|
| 956 |
+
return self._fit(X)
|
| 957 |
+
|
| 958 |
+
def _fit(self, X):
|
| 959 |
+
"""Fit without validation
|
| 960 |
+
|
| 961 |
+
Parameters
|
| 962 |
+
----------
|
| 963 |
+
X : ndarray of shape (n_samples, n_features) or (n_samples, n_samples)
|
| 964 |
+
Training instances to cluster, or distances between instances if
|
| 965 |
+
``affinity='precomputed'``.
|
| 966 |
+
|
| 967 |
+
Returns
|
| 968 |
+
-------
|
| 969 |
+
self : object
|
| 970 |
+
Returns the fitted instance.
|
| 971 |
+
"""
|
| 972 |
+
memory = check_memory(self.memory)
|
| 973 |
+
|
| 974 |
+
self._metric = self.metric
|
| 975 |
+
# TODO(1.4): Remove
|
| 976 |
+
if self.affinity != "deprecated":
|
| 977 |
+
if self.metric is not None:
|
| 978 |
+
raise ValueError(
|
| 979 |
+
"Both `affinity` and `metric` attributes were set. Attribute"
|
| 980 |
+
" `affinity` was deprecated in version 1.2 and will be removed in"
|
| 981 |
+
" 1.4. To avoid this error, only set the `metric` attribute."
|
| 982 |
+
)
|
| 983 |
+
warnings.warn(
|
| 984 |
+
"Attribute `affinity` was deprecated in version 1.2 and will be removed"
|
| 985 |
+
" in 1.4. Use `metric` instead",
|
| 986 |
+
FutureWarning,
|
| 987 |
+
)
|
| 988 |
+
self._metric = self.affinity
|
| 989 |
+
elif self.metric is None:
|
| 990 |
+
self._metric = "euclidean"
|
| 991 |
+
|
| 992 |
+
if not ((self.n_clusters is None) ^ (self.distance_threshold is None)):
|
| 993 |
+
raise ValueError(
|
| 994 |
+
"Exactly one of n_clusters and "
|
| 995 |
+
"distance_threshold has to be set, and the other "
|
| 996 |
+
"needs to be None."
|
| 997 |
+
)
|
| 998 |
+
|
| 999 |
+
if self.distance_threshold is not None and not self.compute_full_tree:
|
| 1000 |
+
raise ValueError(
|
| 1001 |
+
"compute_full_tree must be True if distance_threshold is set."
|
| 1002 |
+
)
|
| 1003 |
+
|
| 1004 |
+
if self.linkage == "ward" and self._metric != "euclidean":
|
| 1005 |
+
raise ValueError(
|
| 1006 |
+
f"{self._metric} was provided as metric. Ward can only "
|
| 1007 |
+
"work with euclidean distances."
|
| 1008 |
+
)
|
| 1009 |
+
|
| 1010 |
+
tree_builder = _TREE_BUILDERS[self.linkage]
|
| 1011 |
+
|
| 1012 |
+
connectivity = self.connectivity
|
| 1013 |
+
if self.connectivity is not None:
|
| 1014 |
+
if callable(self.connectivity):
|
| 1015 |
+
connectivity = self.connectivity(X)
|
| 1016 |
+
connectivity = check_array(
|
| 1017 |
+
connectivity, accept_sparse=["csr", "coo", "lil"]
|
| 1018 |
+
)
|
| 1019 |
+
|
| 1020 |
+
n_samples = len(X)
|
| 1021 |
+
compute_full_tree = self.compute_full_tree
|
| 1022 |
+
if self.connectivity is None:
|
| 1023 |
+
compute_full_tree = True
|
| 1024 |
+
if compute_full_tree == "auto":
|
| 1025 |
+
if self.distance_threshold is not None:
|
| 1026 |
+
compute_full_tree = True
|
| 1027 |
+
else:
|
| 1028 |
+
# Early stopping is likely to give a speed up only for
|
| 1029 |
+
# a large number of clusters. The actual threshold
|
| 1030 |
+
# implemented here is heuristic
|
| 1031 |
+
compute_full_tree = self.n_clusters < max(100, 0.02 * n_samples)
|
| 1032 |
+
n_clusters = self.n_clusters
|
| 1033 |
+
if compute_full_tree:
|
| 1034 |
+
n_clusters = None
|
| 1035 |
+
|
| 1036 |
+
# Construct the tree
|
| 1037 |
+
kwargs = {}
|
| 1038 |
+
if self.linkage != "ward":
|
| 1039 |
+
kwargs["linkage"] = self.linkage
|
| 1040 |
+
kwargs["affinity"] = self._metric
|
| 1041 |
+
|
| 1042 |
+
distance_threshold = self.distance_threshold
|
| 1043 |
+
|
| 1044 |
+
return_distance = (distance_threshold is not None) or self.compute_distances
|
| 1045 |
+
|
| 1046 |
+
out = memory.cache(tree_builder)(
|
| 1047 |
+
X,
|
| 1048 |
+
connectivity=connectivity,
|
| 1049 |
+
n_clusters=n_clusters,
|
| 1050 |
+
return_distance=return_distance,
|
| 1051 |
+
**kwargs,
|
| 1052 |
+
)
|
| 1053 |
+
(self.children_, self.n_connected_components_, self.n_leaves_, parents) = out[
|
| 1054 |
+
:4
|
| 1055 |
+
]
|
| 1056 |
+
|
| 1057 |
+
if return_distance:
|
| 1058 |
+
self.distances_ = out[-1]
|
| 1059 |
+
|
| 1060 |
+
if self.distance_threshold is not None: # distance_threshold is used
|
| 1061 |
+
self.n_clusters_ = (
|
| 1062 |
+
np.count_nonzero(self.distances_ >= distance_threshold) + 1
|
| 1063 |
+
)
|
| 1064 |
+
else: # n_clusters is used
|
| 1065 |
+
self.n_clusters_ = self.n_clusters
|
| 1066 |
+
|
| 1067 |
+
# Cut the tree
|
| 1068 |
+
if compute_full_tree:
|
| 1069 |
+
self.labels_ = _hc_cut(self.n_clusters_, self.children_, self.n_leaves_)
|
| 1070 |
+
else:
|
| 1071 |
+
labels = _hierarchical.hc_get_heads(parents, copy=False)
|
| 1072 |
+
# copy to avoid holding a reference on the original array
|
| 1073 |
+
labels = np.copy(labels[:n_samples])
|
| 1074 |
+
# Reassign cluster numbers
|
| 1075 |
+
self.labels_ = np.searchsorted(np.unique(labels), labels)
|
| 1076 |
+
return self
|
| 1077 |
+
|
| 1078 |
+
def fit_predict(self, X, y=None):
|
| 1079 |
+
"""Fit and return the result of each sample's clustering assignment.
|
| 1080 |
+
|
| 1081 |
+
In addition to fitting, this method also return the result of the
|
| 1082 |
+
clustering assignment for each sample in the training set.
|
| 1083 |
+
|
| 1084 |
+
Parameters
|
| 1085 |
+
----------
|
| 1086 |
+
X : array-like of shape (n_samples, n_features) or \
|
| 1087 |
+
(n_samples, n_samples)
|
| 1088 |
+
Training instances to cluster, or distances between instances if
|
| 1089 |
+
``affinity='precomputed'``.
|
| 1090 |
+
|
| 1091 |
+
y : Ignored
|
| 1092 |
+
Not used, present here for API consistency by convention.
|
| 1093 |
+
|
| 1094 |
+
Returns
|
| 1095 |
+
-------
|
| 1096 |
+
labels : ndarray of shape (n_samples,)
|
| 1097 |
+
Cluster labels.
|
| 1098 |
+
"""
|
| 1099 |
+
return super().fit_predict(X, y)
|
| 1100 |
+
|
| 1101 |
+
|
| 1102 |
+
class FeatureAgglomeration(
|
| 1103 |
+
ClassNamePrefixFeaturesOutMixin, AgglomerativeClustering, AgglomerationTransform
|
| 1104 |
+
):
|
| 1105 |
+
"""Agglomerate features.
|
| 1106 |
+
|
| 1107 |
+
Recursively merges pair of clusters of features.
|
| 1108 |
+
|
| 1109 |
+
Read more in the :ref:`User Guide <hierarchical_clustering>`.
|
| 1110 |
+
|
| 1111 |
+
Parameters
|
| 1112 |
+
----------
|
| 1113 |
+
n_clusters : int or None, default=2
|
| 1114 |
+
The number of clusters to find. It must be ``None`` if
|
| 1115 |
+
``distance_threshold`` is not ``None``.
|
| 1116 |
+
|
| 1117 |
+
affinity : str or callable, default='euclidean'
|
| 1118 |
+
The metric to use when calculating distance between instances in a
|
| 1119 |
+
feature array. If metric is a string or callable, it must be one of
|
| 1120 |
+
the options allowed by :func:`sklearn.metrics.pairwise_distances` for
|
| 1121 |
+
its metric parameter.
|
| 1122 |
+
If linkage is "ward", only "euclidean" is accepted.
|
| 1123 |
+
If "precomputed", a distance matrix (instead of a similarity matrix)
|
| 1124 |
+
is needed as input for the fit method.
|
| 1125 |
+
|
| 1126 |
+
.. deprecated:: 1.2
|
| 1127 |
+
`affinity` was deprecated in version 1.2 and will be renamed to
|
| 1128 |
+
`metric` in 1.4.
|
| 1129 |
+
|
| 1130 |
+
metric : str or callable, default=None
|
| 1131 |
+
Metric used to compute the linkage. Can be "euclidean", "l1", "l2",
|
| 1132 |
+
"manhattan", "cosine", or "precomputed". If set to `None` then
|
| 1133 |
+
"euclidean" is used. If linkage is "ward", only "euclidean" is
|
| 1134 |
+
accepted. If "precomputed", a distance matrix is needed as input for
|
| 1135 |
+
the fit method.
|
| 1136 |
+
|
| 1137 |
+
.. versionadded:: 1.2
|
| 1138 |
+
|
| 1139 |
+
memory : str or object with the joblib.Memory interface, default=None
|
| 1140 |
+
Used to cache the output of the computation of the tree.
|
| 1141 |
+
By default, no caching is done. If a string is given, it is the
|
| 1142 |
+
path to the caching directory.
|
| 1143 |
+
|
| 1144 |
+
connectivity : array-like or callable, default=None
|
| 1145 |
+
Connectivity matrix. Defines for each feature the neighboring
|
| 1146 |
+
features following a given structure of the data.
|
| 1147 |
+
This can be a connectivity matrix itself or a callable that transforms
|
| 1148 |
+
the data into a connectivity matrix, such as derived from
|
| 1149 |
+
`kneighbors_graph`. Default is `None`, i.e, the
|
| 1150 |
+
hierarchical clustering algorithm is unstructured.
|
| 1151 |
+
|
| 1152 |
+
compute_full_tree : 'auto' or bool, default='auto'
|
| 1153 |
+
Stop early the construction of the tree at `n_clusters`. This is useful
|
| 1154 |
+
to decrease computation time if the number of clusters is not small
|
| 1155 |
+
compared to the number of features. This option is useful only when
|
| 1156 |
+
specifying a connectivity matrix. Note also that when varying the
|
| 1157 |
+
number of clusters and using caching, it may be advantageous to compute
|
| 1158 |
+
the full tree. It must be ``True`` if ``distance_threshold`` is not
|
| 1159 |
+
``None``. By default `compute_full_tree` is "auto", which is equivalent
|
| 1160 |
+
to `True` when `distance_threshold` is not `None` or that `n_clusters`
|
| 1161 |
+
is inferior to the maximum between 100 or `0.02 * n_samples`.
|
| 1162 |
+
Otherwise, "auto" is equivalent to `False`.
|
| 1163 |
+
|
| 1164 |
+
linkage : {"ward", "complete", "average", "single"}, default="ward"
|
| 1165 |
+
Which linkage criterion to use. The linkage criterion determines which
|
| 1166 |
+
distance to use between sets of features. The algorithm will merge
|
| 1167 |
+
the pairs of cluster that minimize this criterion.
|
| 1168 |
+
|
| 1169 |
+
- "ward" minimizes the variance of the clusters being merged.
|
| 1170 |
+
- "complete" or maximum linkage uses the maximum distances between
|
| 1171 |
+
all features of the two sets.
|
| 1172 |
+
- "average" uses the average of the distances of each feature of
|
| 1173 |
+
the two sets.
|
| 1174 |
+
- "single" uses the minimum of the distances between all features
|
| 1175 |
+
of the two sets.
|
| 1176 |
+
|
| 1177 |
+
pooling_func : callable, default=np.mean
|
| 1178 |
+
This combines the values of agglomerated features into a single
|
| 1179 |
+
value, and should accept an array of shape [M, N] and the keyword
|
| 1180 |
+
argument `axis=1`, and reduce it to an array of size [M].
|
| 1181 |
+
|
| 1182 |
+
distance_threshold : float, default=None
|
| 1183 |
+
The linkage distance threshold at or above which clusters will not be
|
| 1184 |
+
merged. If not ``None``, ``n_clusters`` must be ``None`` and
|
| 1185 |
+
``compute_full_tree`` must be ``True``.
|
| 1186 |
+
|
| 1187 |
+
.. versionadded:: 0.21
|
| 1188 |
+
|
| 1189 |
+
compute_distances : bool, default=False
|
| 1190 |
+
Computes distances between clusters even if `distance_threshold` is not
|
| 1191 |
+
used. This can be used to make dendrogram visualization, but introduces
|
| 1192 |
+
a computational and memory overhead.
|
| 1193 |
+
|
| 1194 |
+
.. versionadded:: 0.24
|
| 1195 |
+
|
| 1196 |
+
Attributes
|
| 1197 |
+
----------
|
| 1198 |
+
n_clusters_ : int
|
| 1199 |
+
The number of clusters found by the algorithm. If
|
| 1200 |
+
``distance_threshold=None``, it will be equal to the given
|
| 1201 |
+
``n_clusters``.
|
| 1202 |
+
|
| 1203 |
+
labels_ : array-like of (n_features,)
|
| 1204 |
+
Cluster labels for each feature.
|
| 1205 |
+
|
| 1206 |
+
n_leaves_ : int
|
| 1207 |
+
Number of leaves in the hierarchical tree.
|
| 1208 |
+
|
| 1209 |
+
n_connected_components_ : int
|
| 1210 |
+
The estimated number of connected components in the graph.
|
| 1211 |
+
|
| 1212 |
+
.. versionadded:: 0.21
|
| 1213 |
+
``n_connected_components_`` was added to replace ``n_components_``.
|
| 1214 |
+
|
| 1215 |
+
n_features_in_ : int
|
| 1216 |
+
Number of features seen during :term:`fit`.
|
| 1217 |
+
|
| 1218 |
+
.. versionadded:: 0.24
|
| 1219 |
+
|
| 1220 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 1221 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 1222 |
+
has feature names that are all strings.
|
| 1223 |
+
|
| 1224 |
+
.. versionadded:: 1.0
|
| 1225 |
+
|
| 1226 |
+
children_ : array-like of shape (n_nodes-1, 2)
|
| 1227 |
+
The children of each non-leaf node. Values less than `n_features`
|
| 1228 |
+
correspond to leaves of the tree which are the original samples.
|
| 1229 |
+
A node `i` greater than or equal to `n_features` is a non-leaf
|
| 1230 |
+
node and has children `children_[i - n_features]`. Alternatively
|
| 1231 |
+
at the i-th iteration, children[i][0] and children[i][1]
|
| 1232 |
+
are merged to form node `n_features + i`.
|
| 1233 |
+
|
| 1234 |
+
distances_ : array-like of shape (n_nodes-1,)
|
| 1235 |
+
Distances between nodes in the corresponding place in `children_`.
|
| 1236 |
+
Only computed if `distance_threshold` is used or `compute_distances`
|
| 1237 |
+
is set to `True`.
|
| 1238 |
+
|
| 1239 |
+
See Also
|
| 1240 |
+
--------
|
| 1241 |
+
AgglomerativeClustering : Agglomerative clustering samples instead of
|
| 1242 |
+
features.
|
| 1243 |
+
ward_tree : Hierarchical clustering with ward linkage.
|
| 1244 |
+
|
| 1245 |
+
Examples
|
| 1246 |
+
--------
|
| 1247 |
+
>>> import numpy as np
|
| 1248 |
+
>>> from sklearn import datasets, cluster
|
| 1249 |
+
>>> digits = datasets.load_digits()
|
| 1250 |
+
>>> images = digits.images
|
| 1251 |
+
>>> X = np.reshape(images, (len(images), -1))
|
| 1252 |
+
>>> agglo = cluster.FeatureAgglomeration(n_clusters=32)
|
| 1253 |
+
>>> agglo.fit(X)
|
| 1254 |
+
FeatureAgglomeration(n_clusters=32)
|
| 1255 |
+
>>> X_reduced = agglo.transform(X)
|
| 1256 |
+
>>> X_reduced.shape
|
| 1257 |
+
(1797, 32)
|
| 1258 |
+
"""
|
| 1259 |
+
|
| 1260 |
+
_parameter_constraints: dict = {
|
| 1261 |
+
"n_clusters": [Interval(Integral, 1, None, closed="left"), None],
|
| 1262 |
+
"affinity": [
|
| 1263 |
+
Hidden(StrOptions({"deprecated"})),
|
| 1264 |
+
StrOptions(set(_VALID_METRICS) | {"precomputed"}),
|
| 1265 |
+
callable,
|
| 1266 |
+
],
|
| 1267 |
+
"metric": [
|
| 1268 |
+
StrOptions(set(_VALID_METRICS) | {"precomputed"}),
|
| 1269 |
+
callable,
|
| 1270 |
+
None,
|
| 1271 |
+
],
|
| 1272 |
+
"memory": [str, HasMethods("cache"), None],
|
| 1273 |
+
"connectivity": ["array-like", callable, None],
|
| 1274 |
+
"compute_full_tree": [StrOptions({"auto"}), "boolean"],
|
| 1275 |
+
"linkage": [StrOptions(set(_TREE_BUILDERS.keys()))],
|
| 1276 |
+
"pooling_func": [callable],
|
| 1277 |
+
"distance_threshold": [Interval(Real, 0, None, closed="left"), None],
|
| 1278 |
+
"compute_distances": ["boolean"],
|
| 1279 |
+
}
|
| 1280 |
+
|
| 1281 |
+
def __init__(
|
| 1282 |
+
self,
|
| 1283 |
+
n_clusters=2,
|
| 1284 |
+
*,
|
| 1285 |
+
affinity="deprecated", # TODO(1.4): Remove
|
| 1286 |
+
metric=None, # TODO(1.4): Set to "euclidean"
|
| 1287 |
+
memory=None,
|
| 1288 |
+
connectivity=None,
|
| 1289 |
+
compute_full_tree="auto",
|
| 1290 |
+
linkage="ward",
|
| 1291 |
+
pooling_func=np.mean,
|
| 1292 |
+
distance_threshold=None,
|
| 1293 |
+
compute_distances=False,
|
| 1294 |
+
):
|
| 1295 |
+
super().__init__(
|
| 1296 |
+
n_clusters=n_clusters,
|
| 1297 |
+
memory=memory,
|
| 1298 |
+
connectivity=connectivity,
|
| 1299 |
+
compute_full_tree=compute_full_tree,
|
| 1300 |
+
linkage=linkage,
|
| 1301 |
+
affinity=affinity,
|
| 1302 |
+
metric=metric,
|
| 1303 |
+
distance_threshold=distance_threshold,
|
| 1304 |
+
compute_distances=compute_distances,
|
| 1305 |
+
)
|
| 1306 |
+
self.pooling_func = pooling_func
|
| 1307 |
+
|
| 1308 |
+
def fit(self, X, y=None):
|
| 1309 |
+
"""Fit the hierarchical clustering on the data.
|
| 1310 |
+
|
| 1311 |
+
Parameters
|
| 1312 |
+
----------
|
| 1313 |
+
X : array-like of shape (n_samples, n_features)
|
| 1314 |
+
The data.
|
| 1315 |
+
|
| 1316 |
+
y : Ignored
|
| 1317 |
+
Not used, present here for API consistency by convention.
|
| 1318 |
+
|
| 1319 |
+
Returns
|
| 1320 |
+
-------
|
| 1321 |
+
self : object
|
| 1322 |
+
Returns the transformer.
|
| 1323 |
+
"""
|
| 1324 |
+
self._validate_params()
|
| 1325 |
+
X = self._validate_data(X, ensure_min_features=2)
|
| 1326 |
+
super()._fit(X.T)
|
| 1327 |
+
self._n_features_out = self.n_clusters_
|
| 1328 |
+
return self
|
| 1329 |
+
|
| 1330 |
+
@property
|
| 1331 |
+
def fit_predict(self):
|
| 1332 |
+
"""Fit and return the result of each sample's clustering assignment."""
|
| 1333 |
+
raise AttributeError
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_bicluster.py
ADDED
|
@@ -0,0 +1,628 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Spectral biclustering algorithms."""
|
| 2 |
+
# Authors : Kemal Eren
|
| 3 |
+
# License: BSD 3 clause
|
| 4 |
+
|
| 5 |
+
from abc import ABCMeta, abstractmethod
|
| 6 |
+
|
| 7 |
+
import numpy as np
|
| 8 |
+
from numbers import Integral
|
| 9 |
+
|
| 10 |
+
from scipy.linalg import norm
|
| 11 |
+
from scipy.sparse import dia_matrix, issparse
|
| 12 |
+
from scipy.sparse.linalg import eigsh, svds
|
| 13 |
+
|
| 14 |
+
from . import KMeans, MiniBatchKMeans
|
| 15 |
+
from ..base import BaseEstimator, BiclusterMixin
|
| 16 |
+
from ..utils import check_random_state
|
| 17 |
+
from ..utils import check_scalar
|
| 18 |
+
|
| 19 |
+
from ..utils.extmath import make_nonnegative, randomized_svd, safe_sparse_dot
|
| 20 |
+
|
| 21 |
+
from ..utils.validation import assert_all_finite
|
| 22 |
+
from ..utils._param_validation import Interval, StrOptions
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
__all__ = ["SpectralCoclustering", "SpectralBiclustering"]
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def _scale_normalize(X):
|
| 29 |
+
"""Normalize ``X`` by scaling rows and columns independently.
|
| 30 |
+
|
| 31 |
+
Returns the normalized matrix and the row and column scaling
|
| 32 |
+
factors.
|
| 33 |
+
"""
|
| 34 |
+
X = make_nonnegative(X)
|
| 35 |
+
row_diag = np.asarray(1.0 / np.sqrt(X.sum(axis=1))).squeeze()
|
| 36 |
+
col_diag = np.asarray(1.0 / np.sqrt(X.sum(axis=0))).squeeze()
|
| 37 |
+
row_diag = np.where(np.isnan(row_diag), 0, row_diag)
|
| 38 |
+
col_diag = np.where(np.isnan(col_diag), 0, col_diag)
|
| 39 |
+
if issparse(X):
|
| 40 |
+
n_rows, n_cols = X.shape
|
| 41 |
+
r = dia_matrix((row_diag, [0]), shape=(n_rows, n_rows))
|
| 42 |
+
c = dia_matrix((col_diag, [0]), shape=(n_cols, n_cols))
|
| 43 |
+
an = r * X * c
|
| 44 |
+
else:
|
| 45 |
+
an = row_diag[:, np.newaxis] * X * col_diag
|
| 46 |
+
return an, row_diag, col_diag
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def _bistochastic_normalize(X, max_iter=1000, tol=1e-5):
|
| 50 |
+
"""Normalize rows and columns of ``X`` simultaneously so that all
|
| 51 |
+
rows sum to one constant and all columns sum to a different
|
| 52 |
+
constant.
|
| 53 |
+
"""
|
| 54 |
+
# According to paper, this can also be done more efficiently with
|
| 55 |
+
# deviation reduction and balancing algorithms.
|
| 56 |
+
X = make_nonnegative(X)
|
| 57 |
+
X_scaled = X
|
| 58 |
+
for _ in range(max_iter):
|
| 59 |
+
X_new, _, _ = _scale_normalize(X_scaled)
|
| 60 |
+
if issparse(X):
|
| 61 |
+
dist = norm(X_scaled.data - X.data)
|
| 62 |
+
else:
|
| 63 |
+
dist = norm(X_scaled - X_new)
|
| 64 |
+
X_scaled = X_new
|
| 65 |
+
if dist is not None and dist < tol:
|
| 66 |
+
break
|
| 67 |
+
return X_scaled
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def _log_normalize(X):
|
| 71 |
+
"""Normalize ``X`` according to Kluger's log-interactions scheme."""
|
| 72 |
+
X = make_nonnegative(X, min_value=1)
|
| 73 |
+
if issparse(X):
|
| 74 |
+
raise ValueError(
|
| 75 |
+
"Cannot compute log of a sparse matrix,"
|
| 76 |
+
" because log(x) diverges to -infinity as x"
|
| 77 |
+
" goes to 0."
|
| 78 |
+
)
|
| 79 |
+
L = np.log(X)
|
| 80 |
+
row_avg = L.mean(axis=1)[:, np.newaxis]
|
| 81 |
+
col_avg = L.mean(axis=0)
|
| 82 |
+
avg = L.mean()
|
| 83 |
+
return L - row_avg - col_avg + avg
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
class BaseSpectral(BiclusterMixin, BaseEstimator, metaclass=ABCMeta):
|
| 87 |
+
"""Base class for spectral biclustering."""
|
| 88 |
+
|
| 89 |
+
_parameter_constraints: dict = {
|
| 90 |
+
"svd_method": [StrOptions({"randomized", "arpack"})],
|
| 91 |
+
"n_svd_vecs": [Interval(Integral, 0, None, closed="left"), None],
|
| 92 |
+
"mini_batch": ["boolean"],
|
| 93 |
+
"init": [StrOptions({"k-means++", "random"}), np.ndarray],
|
| 94 |
+
"n_init": [Interval(Integral, 1, None, closed="left")],
|
| 95 |
+
"random_state": ["random_state"],
|
| 96 |
+
}
|
| 97 |
+
|
| 98 |
+
@abstractmethod
|
| 99 |
+
def __init__(
|
| 100 |
+
self,
|
| 101 |
+
n_clusters=3,
|
| 102 |
+
svd_method="randomized",
|
| 103 |
+
n_svd_vecs=None,
|
| 104 |
+
mini_batch=False,
|
| 105 |
+
init="k-means++",
|
| 106 |
+
n_init=10,
|
| 107 |
+
random_state=None,
|
| 108 |
+
):
|
| 109 |
+
self.n_clusters = n_clusters
|
| 110 |
+
self.svd_method = svd_method
|
| 111 |
+
self.n_svd_vecs = n_svd_vecs
|
| 112 |
+
self.mini_batch = mini_batch
|
| 113 |
+
self.init = init
|
| 114 |
+
self.n_init = n_init
|
| 115 |
+
self.random_state = random_state
|
| 116 |
+
|
| 117 |
+
@abstractmethod
|
| 118 |
+
def _check_parameters(self, n_samples):
|
| 119 |
+
"""Validate parameters depending on the input data."""
|
| 120 |
+
|
| 121 |
+
def fit(self, X, y=None):
|
| 122 |
+
"""Create a biclustering for X.
|
| 123 |
+
|
| 124 |
+
Parameters
|
| 125 |
+
----------
|
| 126 |
+
X : array-like of shape (n_samples, n_features)
|
| 127 |
+
Training data.
|
| 128 |
+
|
| 129 |
+
y : Ignored
|
| 130 |
+
Not used, present for API consistency by convention.
|
| 131 |
+
|
| 132 |
+
Returns
|
| 133 |
+
-------
|
| 134 |
+
self : object
|
| 135 |
+
SpectralBiclustering instance.
|
| 136 |
+
"""
|
| 137 |
+
self._validate_params()
|
| 138 |
+
|
| 139 |
+
X = self._validate_data(X, accept_sparse="csr", dtype=np.float64)
|
| 140 |
+
self._check_parameters(X.shape[0])
|
| 141 |
+
self._fit(X)
|
| 142 |
+
return self
|
| 143 |
+
|
| 144 |
+
def _svd(self, array, n_components, n_discard):
|
| 145 |
+
"""Returns first `n_components` left and right singular
|
| 146 |
+
vectors u and v, discarding the first `n_discard`.
|
| 147 |
+
"""
|
| 148 |
+
if self.svd_method == "randomized":
|
| 149 |
+
kwargs = {}
|
| 150 |
+
if self.n_svd_vecs is not None:
|
| 151 |
+
kwargs["n_oversamples"] = self.n_svd_vecs
|
| 152 |
+
u, _, vt = randomized_svd(
|
| 153 |
+
array, n_components, random_state=self.random_state, **kwargs
|
| 154 |
+
)
|
| 155 |
+
|
| 156 |
+
elif self.svd_method == "arpack":
|
| 157 |
+
u, _, vt = svds(array, k=n_components, ncv=self.n_svd_vecs)
|
| 158 |
+
if np.any(np.isnan(vt)):
|
| 159 |
+
# some eigenvalues of A * A.T are negative, causing
|
| 160 |
+
# sqrt() to be np.nan. This causes some vectors in vt
|
| 161 |
+
# to be np.nan.
|
| 162 |
+
A = safe_sparse_dot(array.T, array)
|
| 163 |
+
random_state = check_random_state(self.random_state)
|
| 164 |
+
# initialize with [-1,1] as in ARPACK
|
| 165 |
+
v0 = random_state.uniform(-1, 1, A.shape[0])
|
| 166 |
+
_, v = eigsh(A, ncv=self.n_svd_vecs, v0=v0)
|
| 167 |
+
vt = v.T
|
| 168 |
+
if np.any(np.isnan(u)):
|
| 169 |
+
A = safe_sparse_dot(array, array.T)
|
| 170 |
+
random_state = check_random_state(self.random_state)
|
| 171 |
+
# initialize with [-1,1] as in ARPACK
|
| 172 |
+
v0 = random_state.uniform(-1, 1, A.shape[0])
|
| 173 |
+
_, u = eigsh(A, ncv=self.n_svd_vecs, v0=v0)
|
| 174 |
+
|
| 175 |
+
assert_all_finite(u)
|
| 176 |
+
assert_all_finite(vt)
|
| 177 |
+
u = u[:, n_discard:]
|
| 178 |
+
vt = vt[n_discard:]
|
| 179 |
+
return u, vt.T
|
| 180 |
+
|
| 181 |
+
def _k_means(self, data, n_clusters):
|
| 182 |
+
if self.mini_batch:
|
| 183 |
+
model = MiniBatchKMeans(
|
| 184 |
+
n_clusters,
|
| 185 |
+
init=self.init,
|
| 186 |
+
n_init=self.n_init,
|
| 187 |
+
random_state=self.random_state,
|
| 188 |
+
)
|
| 189 |
+
else:
|
| 190 |
+
model = KMeans(
|
| 191 |
+
n_clusters,
|
| 192 |
+
init=self.init,
|
| 193 |
+
n_init=self.n_init,
|
| 194 |
+
random_state=self.random_state,
|
| 195 |
+
)
|
| 196 |
+
model.fit(data)
|
| 197 |
+
centroid = model.cluster_centers_
|
| 198 |
+
labels = model.labels_
|
| 199 |
+
return centroid, labels
|
| 200 |
+
|
| 201 |
+
def _more_tags(self):
|
| 202 |
+
return {
|
| 203 |
+
"_xfail_checks": {
|
| 204 |
+
"check_estimators_dtypes": "raises nan error",
|
| 205 |
+
"check_fit2d_1sample": "_scale_normalize fails",
|
| 206 |
+
"check_fit2d_1feature": "raises apply_along_axis error",
|
| 207 |
+
"check_estimator_sparse_data": "does not fail gracefully",
|
| 208 |
+
"check_methods_subset_invariance": "empty array passed inside",
|
| 209 |
+
"check_dont_overwrite_parameters": "empty array passed inside",
|
| 210 |
+
"check_fit2d_predict1d": "empty array passed inside",
|
| 211 |
+
}
|
| 212 |
+
}
|
| 213 |
+
|
| 214 |
+
|
| 215 |
+
class SpectralCoclustering(BaseSpectral):
|
| 216 |
+
"""Spectral Co-Clustering algorithm (Dhillon, 2001).
|
| 217 |
+
|
| 218 |
+
Clusters rows and columns of an array `X` to solve the relaxed
|
| 219 |
+
normalized cut of the bipartite graph created from `X` as follows:
|
| 220 |
+
the edge between row vertex `i` and column vertex `j` has weight
|
| 221 |
+
`X[i, j]`.
|
| 222 |
+
|
| 223 |
+
The resulting bicluster structure is block-diagonal, since each
|
| 224 |
+
row and each column belongs to exactly one bicluster.
|
| 225 |
+
|
| 226 |
+
Supports sparse matrices, as long as they are nonnegative.
|
| 227 |
+
|
| 228 |
+
Read more in the :ref:`User Guide <spectral_coclustering>`.
|
| 229 |
+
|
| 230 |
+
Parameters
|
| 231 |
+
----------
|
| 232 |
+
n_clusters : int, default=3
|
| 233 |
+
The number of biclusters to find.
|
| 234 |
+
|
| 235 |
+
svd_method : {'randomized', 'arpack'}, default='randomized'
|
| 236 |
+
Selects the algorithm for finding singular vectors. May be
|
| 237 |
+
'randomized' or 'arpack'. If 'randomized', use
|
| 238 |
+
:func:`sklearn.utils.extmath.randomized_svd`, which may be faster
|
| 239 |
+
for large matrices. If 'arpack', use
|
| 240 |
+
:func:`scipy.sparse.linalg.svds`, which is more accurate, but
|
| 241 |
+
possibly slower in some cases.
|
| 242 |
+
|
| 243 |
+
n_svd_vecs : int, default=None
|
| 244 |
+
Number of vectors to use in calculating the SVD. Corresponds
|
| 245 |
+
to `ncv` when `svd_method=arpack` and `n_oversamples` when
|
| 246 |
+
`svd_method` is 'randomized`.
|
| 247 |
+
|
| 248 |
+
mini_batch : bool, default=False
|
| 249 |
+
Whether to use mini-batch k-means, which is faster but may get
|
| 250 |
+
different results.
|
| 251 |
+
|
| 252 |
+
init : {'k-means++', 'random'}, or ndarray of shape \
|
| 253 |
+
(n_clusters, n_features), default='k-means++'
|
| 254 |
+
Method for initialization of k-means algorithm; defaults to
|
| 255 |
+
'k-means++'.
|
| 256 |
+
|
| 257 |
+
n_init : int, default=10
|
| 258 |
+
Number of random initializations that are tried with the
|
| 259 |
+
k-means algorithm.
|
| 260 |
+
|
| 261 |
+
If mini-batch k-means is used, the best initialization is
|
| 262 |
+
chosen and the algorithm runs once. Otherwise, the algorithm
|
| 263 |
+
is run for each initialization and the best solution chosen.
|
| 264 |
+
|
| 265 |
+
random_state : int, RandomState instance, default=None
|
| 266 |
+
Used for randomizing the singular value decomposition and the k-means
|
| 267 |
+
initialization. Use an int to make the randomness deterministic.
|
| 268 |
+
See :term:`Glossary <random_state>`.
|
| 269 |
+
|
| 270 |
+
Attributes
|
| 271 |
+
----------
|
| 272 |
+
rows_ : array-like of shape (n_row_clusters, n_rows)
|
| 273 |
+
Results of the clustering. `rows[i, r]` is True if
|
| 274 |
+
cluster `i` contains row `r`. Available only after calling ``fit``.
|
| 275 |
+
|
| 276 |
+
columns_ : array-like of shape (n_column_clusters, n_columns)
|
| 277 |
+
Results of the clustering, like `rows`.
|
| 278 |
+
|
| 279 |
+
row_labels_ : array-like of shape (n_rows,)
|
| 280 |
+
The bicluster label of each row.
|
| 281 |
+
|
| 282 |
+
column_labels_ : array-like of shape (n_cols,)
|
| 283 |
+
The bicluster label of each column.
|
| 284 |
+
|
| 285 |
+
biclusters_ : tuple of two ndarrays
|
| 286 |
+
The tuple contains the `rows_` and `columns_` arrays.
|
| 287 |
+
|
| 288 |
+
n_features_in_ : int
|
| 289 |
+
Number of features seen during :term:`fit`.
|
| 290 |
+
|
| 291 |
+
.. versionadded:: 0.24
|
| 292 |
+
|
| 293 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 294 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 295 |
+
has feature names that are all strings.
|
| 296 |
+
|
| 297 |
+
.. versionadded:: 1.0
|
| 298 |
+
|
| 299 |
+
See Also
|
| 300 |
+
--------
|
| 301 |
+
SpectralBiclustering : Partitions rows and columns under the assumption
|
| 302 |
+
that the data has an underlying checkerboard structure.
|
| 303 |
+
|
| 304 |
+
References
|
| 305 |
+
----------
|
| 306 |
+
* :doi:`Dhillon, Inderjit S, 2001. Co-clustering documents and words using
|
| 307 |
+
bipartite spectral graph partitioning.
|
| 308 |
+
<10.1145/502512.502550>`
|
| 309 |
+
|
| 310 |
+
Examples
|
| 311 |
+
--------
|
| 312 |
+
>>> from sklearn.cluster import SpectralCoclustering
|
| 313 |
+
>>> import numpy as np
|
| 314 |
+
>>> X = np.array([[1, 1], [2, 1], [1, 0],
|
| 315 |
+
... [4, 7], [3, 5], [3, 6]])
|
| 316 |
+
>>> clustering = SpectralCoclustering(n_clusters=2, random_state=0).fit(X)
|
| 317 |
+
>>> clustering.row_labels_ #doctest: +SKIP
|
| 318 |
+
array([0, 1, 1, 0, 0, 0], dtype=int32)
|
| 319 |
+
>>> clustering.column_labels_ #doctest: +SKIP
|
| 320 |
+
array([0, 0], dtype=int32)
|
| 321 |
+
>>> clustering
|
| 322 |
+
SpectralCoclustering(n_clusters=2, random_state=0)
|
| 323 |
+
"""
|
| 324 |
+
|
| 325 |
+
_parameter_constraints: dict = {
|
| 326 |
+
**BaseSpectral._parameter_constraints,
|
| 327 |
+
"n_clusters": [Interval(Integral, 1, None, closed="left")],
|
| 328 |
+
}
|
| 329 |
+
|
| 330 |
+
def __init__(
|
| 331 |
+
self,
|
| 332 |
+
n_clusters=3,
|
| 333 |
+
*,
|
| 334 |
+
svd_method="randomized",
|
| 335 |
+
n_svd_vecs=None,
|
| 336 |
+
mini_batch=False,
|
| 337 |
+
init="k-means++",
|
| 338 |
+
n_init=10,
|
| 339 |
+
random_state=None,
|
| 340 |
+
):
|
| 341 |
+
super().__init__(
|
| 342 |
+
n_clusters, svd_method, n_svd_vecs, mini_batch, init, n_init, random_state
|
| 343 |
+
)
|
| 344 |
+
|
| 345 |
+
def _check_parameters(self, n_samples):
|
| 346 |
+
if self.n_clusters > n_samples:
|
| 347 |
+
raise ValueError(
|
| 348 |
+
f"n_clusters should be <= n_samples={n_samples}. Got"
|
| 349 |
+
f" {self.n_clusters} instead."
|
| 350 |
+
)
|
| 351 |
+
|
| 352 |
+
def _fit(self, X):
|
| 353 |
+
normalized_data, row_diag, col_diag = _scale_normalize(X)
|
| 354 |
+
n_sv = 1 + int(np.ceil(np.log2(self.n_clusters)))
|
| 355 |
+
u, v = self._svd(normalized_data, n_sv, n_discard=1)
|
| 356 |
+
z = np.vstack((row_diag[:, np.newaxis] * u, col_diag[:, np.newaxis] * v))
|
| 357 |
+
|
| 358 |
+
_, labels = self._k_means(z, self.n_clusters)
|
| 359 |
+
|
| 360 |
+
n_rows = X.shape[0]
|
| 361 |
+
self.row_labels_ = labels[:n_rows]
|
| 362 |
+
self.column_labels_ = labels[n_rows:]
|
| 363 |
+
|
| 364 |
+
self.rows_ = np.vstack([self.row_labels_ == c for c in range(self.n_clusters)])
|
| 365 |
+
self.columns_ = np.vstack(
|
| 366 |
+
[self.column_labels_ == c for c in range(self.n_clusters)]
|
| 367 |
+
)
|
| 368 |
+
|
| 369 |
+
|
| 370 |
+
class SpectralBiclustering(BaseSpectral):
|
| 371 |
+
"""Spectral biclustering (Kluger, 2003).
|
| 372 |
+
|
| 373 |
+
Partitions rows and columns under the assumption that the data has
|
| 374 |
+
an underlying checkerboard structure. For instance, if there are
|
| 375 |
+
two row partitions and three column partitions, each row will
|
| 376 |
+
belong to three biclusters, and each column will belong to two
|
| 377 |
+
biclusters. The outer product of the corresponding row and column
|
| 378 |
+
label vectors gives this checkerboard structure.
|
| 379 |
+
|
| 380 |
+
Read more in the :ref:`User Guide <spectral_biclustering>`.
|
| 381 |
+
|
| 382 |
+
Parameters
|
| 383 |
+
----------
|
| 384 |
+
n_clusters : int or tuple (n_row_clusters, n_column_clusters), default=3
|
| 385 |
+
The number of row and column clusters in the checkerboard
|
| 386 |
+
structure.
|
| 387 |
+
|
| 388 |
+
method : {'bistochastic', 'scale', 'log'}, default='bistochastic'
|
| 389 |
+
Method of normalizing and converting singular vectors into
|
| 390 |
+
biclusters. May be one of 'scale', 'bistochastic', or 'log'.
|
| 391 |
+
The authors recommend using 'log'. If the data is sparse,
|
| 392 |
+
however, log normalization will not work, which is why the
|
| 393 |
+
default is 'bistochastic'.
|
| 394 |
+
|
| 395 |
+
.. warning::
|
| 396 |
+
if `method='log'`, the data must not be sparse.
|
| 397 |
+
|
| 398 |
+
n_components : int, default=6
|
| 399 |
+
Number of singular vectors to check.
|
| 400 |
+
|
| 401 |
+
n_best : int, default=3
|
| 402 |
+
Number of best singular vectors to which to project the data
|
| 403 |
+
for clustering.
|
| 404 |
+
|
| 405 |
+
svd_method : {'randomized', 'arpack'}, default='randomized'
|
| 406 |
+
Selects the algorithm for finding singular vectors. May be
|
| 407 |
+
'randomized' or 'arpack'. If 'randomized', uses
|
| 408 |
+
:func:`~sklearn.utils.extmath.randomized_svd`, which may be faster
|
| 409 |
+
for large matrices. If 'arpack', uses
|
| 410 |
+
`scipy.sparse.linalg.svds`, which is more accurate, but
|
| 411 |
+
possibly slower in some cases.
|
| 412 |
+
|
| 413 |
+
n_svd_vecs : int, default=None
|
| 414 |
+
Number of vectors to use in calculating the SVD. Corresponds
|
| 415 |
+
to `ncv` when `svd_method=arpack` and `n_oversamples` when
|
| 416 |
+
`svd_method` is 'randomized`.
|
| 417 |
+
|
| 418 |
+
mini_batch : bool, default=False
|
| 419 |
+
Whether to use mini-batch k-means, which is faster but may get
|
| 420 |
+
different results.
|
| 421 |
+
|
| 422 |
+
init : {'k-means++', 'random'} or ndarray of shape (n_clusters, n_features), \
|
| 423 |
+
default='k-means++'
|
| 424 |
+
Method for initialization of k-means algorithm; defaults to
|
| 425 |
+
'k-means++'.
|
| 426 |
+
|
| 427 |
+
n_init : int, default=10
|
| 428 |
+
Number of random initializations that are tried with the
|
| 429 |
+
k-means algorithm.
|
| 430 |
+
|
| 431 |
+
If mini-batch k-means is used, the best initialization is
|
| 432 |
+
chosen and the algorithm runs once. Otherwise, the algorithm
|
| 433 |
+
is run for each initialization and the best solution chosen.
|
| 434 |
+
|
| 435 |
+
random_state : int, RandomState instance, default=None
|
| 436 |
+
Used for randomizing the singular value decomposition and the k-means
|
| 437 |
+
initialization. Use an int to make the randomness deterministic.
|
| 438 |
+
See :term:`Glossary <random_state>`.
|
| 439 |
+
|
| 440 |
+
Attributes
|
| 441 |
+
----------
|
| 442 |
+
rows_ : array-like of shape (n_row_clusters, n_rows)
|
| 443 |
+
Results of the clustering. `rows[i, r]` is True if
|
| 444 |
+
cluster `i` contains row `r`. Available only after calling ``fit``.
|
| 445 |
+
|
| 446 |
+
columns_ : array-like of shape (n_column_clusters, n_columns)
|
| 447 |
+
Results of the clustering, like `rows`.
|
| 448 |
+
|
| 449 |
+
row_labels_ : array-like of shape (n_rows,)
|
| 450 |
+
Row partition labels.
|
| 451 |
+
|
| 452 |
+
column_labels_ : array-like of shape (n_cols,)
|
| 453 |
+
Column partition labels.
|
| 454 |
+
|
| 455 |
+
biclusters_ : tuple of two ndarrays
|
| 456 |
+
The tuple contains the `rows_` and `columns_` arrays.
|
| 457 |
+
|
| 458 |
+
n_features_in_ : int
|
| 459 |
+
Number of features seen during :term:`fit`.
|
| 460 |
+
|
| 461 |
+
.. versionadded:: 0.24
|
| 462 |
+
|
| 463 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 464 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 465 |
+
has feature names that are all strings.
|
| 466 |
+
|
| 467 |
+
.. versionadded:: 1.0
|
| 468 |
+
|
| 469 |
+
See Also
|
| 470 |
+
--------
|
| 471 |
+
SpectralCoclustering : Spectral Co-Clustering algorithm (Dhillon, 2001).
|
| 472 |
+
|
| 473 |
+
References
|
| 474 |
+
----------
|
| 475 |
+
|
| 476 |
+
* :doi:`Kluger, Yuval, et. al., 2003. Spectral biclustering of microarray
|
| 477 |
+
data: coclustering genes and conditions.
|
| 478 |
+
<10.1101/gr.648603>`
|
| 479 |
+
|
| 480 |
+
Examples
|
| 481 |
+
--------
|
| 482 |
+
>>> from sklearn.cluster import SpectralBiclustering
|
| 483 |
+
>>> import numpy as np
|
| 484 |
+
>>> X = np.array([[1, 1], [2, 1], [1, 0],
|
| 485 |
+
... [4, 7], [3, 5], [3, 6]])
|
| 486 |
+
>>> clustering = SpectralBiclustering(n_clusters=2, random_state=0).fit(X)
|
| 487 |
+
>>> clustering.row_labels_
|
| 488 |
+
array([1, 1, 1, 0, 0, 0], dtype=int32)
|
| 489 |
+
>>> clustering.column_labels_
|
| 490 |
+
array([0, 1], dtype=int32)
|
| 491 |
+
>>> clustering
|
| 492 |
+
SpectralBiclustering(n_clusters=2, random_state=0)
|
| 493 |
+
"""
|
| 494 |
+
|
| 495 |
+
_parameter_constraints: dict = {
|
| 496 |
+
**BaseSpectral._parameter_constraints,
|
| 497 |
+
"n_clusters": [Interval(Integral, 1, None, closed="left"), tuple],
|
| 498 |
+
"method": [StrOptions({"bistochastic", "scale", "log"})],
|
| 499 |
+
"n_components": [Interval(Integral, 1, None, closed="left")],
|
| 500 |
+
"n_best": [Interval(Integral, 1, None, closed="left")],
|
| 501 |
+
}
|
| 502 |
+
|
| 503 |
+
def __init__(
|
| 504 |
+
self,
|
| 505 |
+
n_clusters=3,
|
| 506 |
+
*,
|
| 507 |
+
method="bistochastic",
|
| 508 |
+
n_components=6,
|
| 509 |
+
n_best=3,
|
| 510 |
+
svd_method="randomized",
|
| 511 |
+
n_svd_vecs=None,
|
| 512 |
+
mini_batch=False,
|
| 513 |
+
init="k-means++",
|
| 514 |
+
n_init=10,
|
| 515 |
+
random_state=None,
|
| 516 |
+
):
|
| 517 |
+
super().__init__(
|
| 518 |
+
n_clusters, svd_method, n_svd_vecs, mini_batch, init, n_init, random_state
|
| 519 |
+
)
|
| 520 |
+
self.method = method
|
| 521 |
+
self.n_components = n_components
|
| 522 |
+
self.n_best = n_best
|
| 523 |
+
|
| 524 |
+
def _check_parameters(self, n_samples):
|
| 525 |
+
if isinstance(self.n_clusters, Integral):
|
| 526 |
+
if self.n_clusters > n_samples:
|
| 527 |
+
raise ValueError(
|
| 528 |
+
f"n_clusters should be <= n_samples={n_samples}. Got"
|
| 529 |
+
f" {self.n_clusters} instead."
|
| 530 |
+
)
|
| 531 |
+
else: # tuple
|
| 532 |
+
try:
|
| 533 |
+
n_row_clusters, n_column_clusters = self.n_clusters
|
| 534 |
+
check_scalar(
|
| 535 |
+
n_row_clusters,
|
| 536 |
+
"n_row_clusters",
|
| 537 |
+
target_type=Integral,
|
| 538 |
+
min_val=1,
|
| 539 |
+
max_val=n_samples,
|
| 540 |
+
)
|
| 541 |
+
check_scalar(
|
| 542 |
+
n_column_clusters,
|
| 543 |
+
"n_column_clusters",
|
| 544 |
+
target_type=Integral,
|
| 545 |
+
min_val=1,
|
| 546 |
+
max_val=n_samples,
|
| 547 |
+
)
|
| 548 |
+
except (ValueError, TypeError) as e:
|
| 549 |
+
raise ValueError(
|
| 550 |
+
"Incorrect parameter n_clusters has value:"
|
| 551 |
+
f" {self.n_clusters}. It should either be a single integer"
|
| 552 |
+
" or an iterable with two integers:"
|
| 553 |
+
" (n_row_clusters, n_column_clusters)"
|
| 554 |
+
" And the values are should be in the"
|
| 555 |
+
" range: (1, n_samples)"
|
| 556 |
+
) from e
|
| 557 |
+
|
| 558 |
+
if self.n_best > self.n_components:
|
| 559 |
+
raise ValueError(
|
| 560 |
+
f"n_best={self.n_best} must be <= n_components={self.n_components}."
|
| 561 |
+
)
|
| 562 |
+
|
| 563 |
+
def _fit(self, X):
|
| 564 |
+
n_sv = self.n_components
|
| 565 |
+
if self.method == "bistochastic":
|
| 566 |
+
normalized_data = _bistochastic_normalize(X)
|
| 567 |
+
n_sv += 1
|
| 568 |
+
elif self.method == "scale":
|
| 569 |
+
normalized_data, _, _ = _scale_normalize(X)
|
| 570 |
+
n_sv += 1
|
| 571 |
+
elif self.method == "log":
|
| 572 |
+
normalized_data = _log_normalize(X)
|
| 573 |
+
n_discard = 0 if self.method == "log" else 1
|
| 574 |
+
u, v = self._svd(normalized_data, n_sv, n_discard)
|
| 575 |
+
ut = u.T
|
| 576 |
+
vt = v.T
|
| 577 |
+
|
| 578 |
+
try:
|
| 579 |
+
n_row_clusters, n_col_clusters = self.n_clusters
|
| 580 |
+
except TypeError:
|
| 581 |
+
n_row_clusters = n_col_clusters = self.n_clusters
|
| 582 |
+
|
| 583 |
+
best_ut = self._fit_best_piecewise(ut, self.n_best, n_row_clusters)
|
| 584 |
+
|
| 585 |
+
best_vt = self._fit_best_piecewise(vt, self.n_best, n_col_clusters)
|
| 586 |
+
|
| 587 |
+
self.row_labels_ = self._project_and_cluster(X, best_vt.T, n_row_clusters)
|
| 588 |
+
|
| 589 |
+
self.column_labels_ = self._project_and_cluster(X.T, best_ut.T, n_col_clusters)
|
| 590 |
+
|
| 591 |
+
self.rows_ = np.vstack(
|
| 592 |
+
[
|
| 593 |
+
self.row_labels_ == label
|
| 594 |
+
for label in range(n_row_clusters)
|
| 595 |
+
for _ in range(n_col_clusters)
|
| 596 |
+
]
|
| 597 |
+
)
|
| 598 |
+
self.columns_ = np.vstack(
|
| 599 |
+
[
|
| 600 |
+
self.column_labels_ == label
|
| 601 |
+
for _ in range(n_row_clusters)
|
| 602 |
+
for label in range(n_col_clusters)
|
| 603 |
+
]
|
| 604 |
+
)
|
| 605 |
+
|
| 606 |
+
def _fit_best_piecewise(self, vectors, n_best, n_clusters):
|
| 607 |
+
"""Find the ``n_best`` vectors that are best approximated by piecewise
|
| 608 |
+
constant vectors.
|
| 609 |
+
|
| 610 |
+
The piecewise vectors are found by k-means; the best is chosen
|
| 611 |
+
according to Euclidean distance.
|
| 612 |
+
|
| 613 |
+
"""
|
| 614 |
+
|
| 615 |
+
def make_piecewise(v):
|
| 616 |
+
centroid, labels = self._k_means(v.reshape(-1, 1), n_clusters)
|
| 617 |
+
return centroid[labels].ravel()
|
| 618 |
+
|
| 619 |
+
piecewise_vectors = np.apply_along_axis(make_piecewise, axis=1, arr=vectors)
|
| 620 |
+
dists = np.apply_along_axis(norm, axis=1, arr=(vectors - piecewise_vectors))
|
| 621 |
+
result = vectors[np.argsort(dists)[:n_best]]
|
| 622 |
+
return result
|
| 623 |
+
|
| 624 |
+
def _project_and_cluster(self, data, vectors, n_clusters):
|
| 625 |
+
"""Project ``data`` to ``vectors`` and cluster the result."""
|
| 626 |
+
projected = safe_sparse_dot(data, vectors)
|
| 627 |
+
_, labels = self._k_means(projected, n_clusters)
|
| 628 |
+
return labels
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_feature_agglomeration.py
ADDED
|
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Feature agglomeration. Base classes and functions for performing feature
|
| 3 |
+
agglomeration.
|
| 4 |
+
"""
|
| 5 |
+
# Author: V. Michel, A. Gramfort
|
| 6 |
+
# License: BSD 3 clause
|
| 7 |
+
|
| 8 |
+
import numpy as np
|
| 9 |
+
|
| 10 |
+
from ..base import TransformerMixin
|
| 11 |
+
from ..utils.validation import check_is_fitted
|
| 12 |
+
from scipy.sparse import issparse
|
| 13 |
+
|
| 14 |
+
###############################################################################
|
| 15 |
+
# Mixin class for feature agglomeration.
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
class AgglomerationTransform(TransformerMixin):
|
| 19 |
+
"""
|
| 20 |
+
A class for feature agglomeration via the transform interface.
|
| 21 |
+
"""
|
| 22 |
+
|
| 23 |
+
def transform(self, X):
|
| 24 |
+
"""
|
| 25 |
+
Transform a new matrix using the built clustering.
|
| 26 |
+
|
| 27 |
+
Parameters
|
| 28 |
+
----------
|
| 29 |
+
X : array-like of shape (n_samples, n_features) or \
|
| 30 |
+
(n_samples, n_samples)
|
| 31 |
+
A M by N array of M observations in N dimensions or a length
|
| 32 |
+
M array of M one-dimensional observations.
|
| 33 |
+
|
| 34 |
+
Returns
|
| 35 |
+
-------
|
| 36 |
+
Y : ndarray of shape (n_samples, n_clusters) or (n_clusters,)
|
| 37 |
+
The pooled values for each feature cluster.
|
| 38 |
+
"""
|
| 39 |
+
check_is_fitted(self)
|
| 40 |
+
|
| 41 |
+
X = self._validate_data(X, reset=False)
|
| 42 |
+
if self.pooling_func == np.mean and not issparse(X):
|
| 43 |
+
size = np.bincount(self.labels_)
|
| 44 |
+
n_samples = X.shape[0]
|
| 45 |
+
# a fast way to compute the mean of grouped features
|
| 46 |
+
nX = np.array(
|
| 47 |
+
[np.bincount(self.labels_, X[i, :]) / size for i in range(n_samples)]
|
| 48 |
+
)
|
| 49 |
+
else:
|
| 50 |
+
nX = [
|
| 51 |
+
self.pooling_func(X[:, self.labels_ == l], axis=1)
|
| 52 |
+
for l in np.unique(self.labels_)
|
| 53 |
+
]
|
| 54 |
+
nX = np.array(nX).T
|
| 55 |
+
return nX
|
| 56 |
+
|
| 57 |
+
def inverse_transform(self, Xred):
|
| 58 |
+
"""
|
| 59 |
+
Inverse the transformation and return a vector of size `n_features`.
|
| 60 |
+
|
| 61 |
+
Parameters
|
| 62 |
+
----------
|
| 63 |
+
Xred : array-like of shape (n_samples, n_clusters) or (n_clusters,)
|
| 64 |
+
The values to be assigned to each cluster of samples.
|
| 65 |
+
|
| 66 |
+
Returns
|
| 67 |
+
-------
|
| 68 |
+
X : ndarray of shape (n_samples, n_features) or (n_features,)
|
| 69 |
+
A vector of size `n_samples` with the values of `Xred` assigned to
|
| 70 |
+
each of the cluster of samples.
|
| 71 |
+
"""
|
| 72 |
+
check_is_fitted(self)
|
| 73 |
+
|
| 74 |
+
unil, inverse = np.unique(self.labels_, return_inverse=True)
|
| 75 |
+
return Xred[..., inverse]
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_hierarchical_fast.cpython-310-x86_64-linux-gnu.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fbab2dfb377c76cc63b49b08c32f7f7c3848f951b42019bc08cd41b58f02d397
|
| 3 |
+
size 327769
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_k_means_common.cpython-310-x86_64-linux-gnu.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f2d3838bfb6b911e555204f2e11c0bc2097fe95370c871103e0b4f622b5dd059
|
| 3 |
+
size 417737
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_k_means_common.pxd
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from cython cimport floating
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
cdef floating _euclidean_dense_dense(floating*, floating*, int, bint) nogil
|
| 5 |
+
|
| 6 |
+
cdef floating _euclidean_sparse_dense(floating[::1], int[::1], floating[::1],
|
| 7 |
+
floating, bint) nogil
|
| 8 |
+
|
| 9 |
+
cpdef void _relocate_empty_clusters_dense(
|
| 10 |
+
floating[:, ::1], floating[::1], floating[:, ::1],
|
| 11 |
+
floating[:, ::1], floating[::1], int[::1])
|
| 12 |
+
|
| 13 |
+
cpdef void _relocate_empty_clusters_sparse(
|
| 14 |
+
floating[::1], int[::1], int[::1], floating[::1], floating[:, ::1],
|
| 15 |
+
floating[:, ::1], floating[::1], int[::1])
|
| 16 |
+
|
| 17 |
+
cdef void _average_centers(floating[:, ::1], floating[::1])
|
| 18 |
+
|
| 19 |
+
cdef void _center_shift(floating[:, ::1], floating[:, ::1], floating[::1])
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_k_means_elkan.cpython-310-x86_64-linux-gnu.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:edfd181e6be2fe2da302115e3d5729c53f06d0a7d66d42a19b5e21f87d37ea6d
|
| 3 |
+
size 431033
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_mean_shift.py
ADDED
|
@@ -0,0 +1,536 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Mean shift clustering algorithm.
|
| 2 |
+
|
| 3 |
+
Mean shift clustering aims to discover *blobs* in a smooth density of
|
| 4 |
+
samples. It is a centroid based algorithm, which works by updating candidates
|
| 5 |
+
for centroids to be the mean of the points within a given region. These
|
| 6 |
+
candidates are then filtered in a post-processing stage to eliminate
|
| 7 |
+
near-duplicates to form the final set of centroids.
|
| 8 |
+
|
| 9 |
+
Seeding is performed using a binning technique for scalability.
|
| 10 |
+
"""
|
| 11 |
+
|
| 12 |
+
# Authors: Conrad Lee <conradlee@gmail.com>
|
| 13 |
+
# Alexandre Gramfort <alexandre.gramfort@inria.fr>
|
| 14 |
+
# Gael Varoquaux <gael.varoquaux@normalesup.org>
|
| 15 |
+
# Martino Sorbaro <martino.sorbaro@ed.ac.uk>
|
| 16 |
+
|
| 17 |
+
import numpy as np
|
| 18 |
+
import warnings
|
| 19 |
+
from numbers import Integral, Real
|
| 20 |
+
|
| 21 |
+
from collections import defaultdict
|
| 22 |
+
from ..utils._param_validation import Interval
|
| 23 |
+
from ..utils.validation import check_is_fitted
|
| 24 |
+
from ..utils.parallel import delayed, Parallel
|
| 25 |
+
from ..utils import check_random_state, gen_batches, check_array
|
| 26 |
+
from ..base import BaseEstimator, ClusterMixin
|
| 27 |
+
from ..neighbors import NearestNeighbors
|
| 28 |
+
from ..metrics.pairwise import pairwise_distances_argmin
|
| 29 |
+
from .._config import config_context
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def estimate_bandwidth(X, *, quantile=0.3, n_samples=None, random_state=0, n_jobs=None):
|
| 33 |
+
"""Estimate the bandwidth to use with the mean-shift algorithm.
|
| 34 |
+
|
| 35 |
+
That this function takes time at least quadratic in n_samples. For large
|
| 36 |
+
datasets, it's wise to set that parameter to a small value.
|
| 37 |
+
|
| 38 |
+
Parameters
|
| 39 |
+
----------
|
| 40 |
+
X : array-like of shape (n_samples, n_features)
|
| 41 |
+
Input points.
|
| 42 |
+
|
| 43 |
+
quantile : float, default=0.3
|
| 44 |
+
Should be between [0, 1]
|
| 45 |
+
0.5 means that the median of all pairwise distances is used.
|
| 46 |
+
|
| 47 |
+
n_samples : int, default=None
|
| 48 |
+
The number of samples to use. If not given, all samples are used.
|
| 49 |
+
|
| 50 |
+
random_state : int, RandomState instance, default=None
|
| 51 |
+
The generator used to randomly select the samples from input points
|
| 52 |
+
for bandwidth estimation. Use an int to make the randomness
|
| 53 |
+
deterministic.
|
| 54 |
+
See :term:`Glossary <random_state>`.
|
| 55 |
+
|
| 56 |
+
n_jobs : int, default=None
|
| 57 |
+
The number of parallel jobs to run for neighbors search.
|
| 58 |
+
``None`` means 1 unless in a :obj:`joblib.parallel_backend` context.
|
| 59 |
+
``-1`` means using all processors. See :term:`Glossary <n_jobs>`
|
| 60 |
+
for more details.
|
| 61 |
+
|
| 62 |
+
Returns
|
| 63 |
+
-------
|
| 64 |
+
bandwidth : float
|
| 65 |
+
The bandwidth parameter.
|
| 66 |
+
"""
|
| 67 |
+
X = check_array(X)
|
| 68 |
+
|
| 69 |
+
random_state = check_random_state(random_state)
|
| 70 |
+
if n_samples is not None:
|
| 71 |
+
idx = random_state.permutation(X.shape[0])[:n_samples]
|
| 72 |
+
X = X[idx]
|
| 73 |
+
n_neighbors = int(X.shape[0] * quantile)
|
| 74 |
+
if n_neighbors < 1: # cannot fit NearestNeighbors with n_neighbors = 0
|
| 75 |
+
n_neighbors = 1
|
| 76 |
+
nbrs = NearestNeighbors(n_neighbors=n_neighbors, n_jobs=n_jobs)
|
| 77 |
+
nbrs.fit(X)
|
| 78 |
+
|
| 79 |
+
bandwidth = 0.0
|
| 80 |
+
for batch in gen_batches(len(X), 500):
|
| 81 |
+
d, _ = nbrs.kneighbors(X[batch, :], return_distance=True)
|
| 82 |
+
bandwidth += np.max(d, axis=1).sum()
|
| 83 |
+
|
| 84 |
+
return bandwidth / X.shape[0]
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
# separate function for each seed's iterative loop
|
| 88 |
+
def _mean_shift_single_seed(my_mean, X, nbrs, max_iter):
|
| 89 |
+
# For each seed, climb gradient until convergence or max_iter
|
| 90 |
+
bandwidth = nbrs.get_params()["radius"]
|
| 91 |
+
stop_thresh = 1e-3 * bandwidth # when mean has converged
|
| 92 |
+
completed_iterations = 0
|
| 93 |
+
while True:
|
| 94 |
+
# Find mean of points within bandwidth
|
| 95 |
+
i_nbrs = nbrs.radius_neighbors([my_mean], bandwidth, return_distance=False)[0]
|
| 96 |
+
points_within = X[i_nbrs]
|
| 97 |
+
if len(points_within) == 0:
|
| 98 |
+
break # Depending on seeding strategy this condition may occur
|
| 99 |
+
my_old_mean = my_mean # save the old mean
|
| 100 |
+
my_mean = np.mean(points_within, axis=0)
|
| 101 |
+
# If converged or at max_iter, adds the cluster
|
| 102 |
+
if (
|
| 103 |
+
np.linalg.norm(my_mean - my_old_mean) < stop_thresh
|
| 104 |
+
or completed_iterations == max_iter
|
| 105 |
+
):
|
| 106 |
+
break
|
| 107 |
+
completed_iterations += 1
|
| 108 |
+
return tuple(my_mean), len(points_within), completed_iterations
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
def mean_shift(
|
| 112 |
+
X,
|
| 113 |
+
*,
|
| 114 |
+
bandwidth=None,
|
| 115 |
+
seeds=None,
|
| 116 |
+
bin_seeding=False,
|
| 117 |
+
min_bin_freq=1,
|
| 118 |
+
cluster_all=True,
|
| 119 |
+
max_iter=300,
|
| 120 |
+
n_jobs=None,
|
| 121 |
+
):
|
| 122 |
+
"""Perform mean shift clustering of data using a flat kernel.
|
| 123 |
+
|
| 124 |
+
Read more in the :ref:`User Guide <mean_shift>`.
|
| 125 |
+
|
| 126 |
+
Parameters
|
| 127 |
+
----------
|
| 128 |
+
|
| 129 |
+
X : array-like of shape (n_samples, n_features)
|
| 130 |
+
Input data.
|
| 131 |
+
|
| 132 |
+
bandwidth : float, default=None
|
| 133 |
+
Kernel bandwidth.
|
| 134 |
+
|
| 135 |
+
If bandwidth is not given, it is determined using a heuristic based on
|
| 136 |
+
the median of all pairwise distances. This will take quadratic time in
|
| 137 |
+
the number of samples. The sklearn.cluster.estimate_bandwidth function
|
| 138 |
+
can be used to do this more efficiently.
|
| 139 |
+
|
| 140 |
+
seeds : array-like of shape (n_seeds, n_features) or None
|
| 141 |
+
Point used as initial kernel locations. If None and bin_seeding=False,
|
| 142 |
+
each data point is used as a seed. If None and bin_seeding=True,
|
| 143 |
+
see bin_seeding.
|
| 144 |
+
|
| 145 |
+
bin_seeding : bool, default=False
|
| 146 |
+
If true, initial kernel locations are not locations of all
|
| 147 |
+
points, but rather the location of the discretized version of
|
| 148 |
+
points, where points are binned onto a grid whose coarseness
|
| 149 |
+
corresponds to the bandwidth. Setting this option to True will speed
|
| 150 |
+
up the algorithm because fewer seeds will be initialized.
|
| 151 |
+
Ignored if seeds argument is not None.
|
| 152 |
+
|
| 153 |
+
min_bin_freq : int, default=1
|
| 154 |
+
To speed up the algorithm, accept only those bins with at least
|
| 155 |
+
min_bin_freq points as seeds.
|
| 156 |
+
|
| 157 |
+
cluster_all : bool, default=True
|
| 158 |
+
If true, then all points are clustered, even those orphans that are
|
| 159 |
+
not within any kernel. Orphans are assigned to the nearest kernel.
|
| 160 |
+
If false, then orphans are given cluster label -1.
|
| 161 |
+
|
| 162 |
+
max_iter : int, default=300
|
| 163 |
+
Maximum number of iterations, per seed point before the clustering
|
| 164 |
+
operation terminates (for that seed point), if has not converged yet.
|
| 165 |
+
|
| 166 |
+
n_jobs : int, default=None
|
| 167 |
+
The number of jobs to use for the computation. The following tasks benefit
|
| 168 |
+
from the parallelization:
|
| 169 |
+
|
| 170 |
+
- The search of nearest neighbors for bandwidth estimation and label
|
| 171 |
+
assignments. See the details in the docstring of the
|
| 172 |
+
``NearestNeighbors`` class.
|
| 173 |
+
- Hill-climbing optimization for all seeds.
|
| 174 |
+
|
| 175 |
+
See :term:`Glossary <n_jobs>` for more details.
|
| 176 |
+
|
| 177 |
+
``None`` means 1 unless in a :obj:`joblib.parallel_backend` context.
|
| 178 |
+
``-1`` means using all processors. See :term:`Glossary <n_jobs>`
|
| 179 |
+
for more details.
|
| 180 |
+
|
| 181 |
+
.. versionadded:: 0.17
|
| 182 |
+
Parallel Execution using *n_jobs*.
|
| 183 |
+
|
| 184 |
+
Returns
|
| 185 |
+
-------
|
| 186 |
+
|
| 187 |
+
cluster_centers : ndarray of shape (n_clusters, n_features)
|
| 188 |
+
Coordinates of cluster centers.
|
| 189 |
+
|
| 190 |
+
labels : ndarray of shape (n_samples,)
|
| 191 |
+
Cluster labels for each point.
|
| 192 |
+
|
| 193 |
+
Notes
|
| 194 |
+
-----
|
| 195 |
+
For an example, see :ref:`examples/cluster/plot_mean_shift.py
|
| 196 |
+
<sphx_glr_auto_examples_cluster_plot_mean_shift.py>`.
|
| 197 |
+
"""
|
| 198 |
+
model = MeanShift(
|
| 199 |
+
bandwidth=bandwidth,
|
| 200 |
+
seeds=seeds,
|
| 201 |
+
min_bin_freq=min_bin_freq,
|
| 202 |
+
bin_seeding=bin_seeding,
|
| 203 |
+
cluster_all=cluster_all,
|
| 204 |
+
n_jobs=n_jobs,
|
| 205 |
+
max_iter=max_iter,
|
| 206 |
+
).fit(X)
|
| 207 |
+
return model.cluster_centers_, model.labels_
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
def get_bin_seeds(X, bin_size, min_bin_freq=1):
|
| 211 |
+
"""Find seeds for mean_shift.
|
| 212 |
+
|
| 213 |
+
Finds seeds by first binning data onto a grid whose lines are
|
| 214 |
+
spaced bin_size apart, and then choosing those bins with at least
|
| 215 |
+
min_bin_freq points.
|
| 216 |
+
|
| 217 |
+
Parameters
|
| 218 |
+
----------
|
| 219 |
+
|
| 220 |
+
X : array-like of shape (n_samples, n_features)
|
| 221 |
+
Input points, the same points that will be used in mean_shift.
|
| 222 |
+
|
| 223 |
+
bin_size : float
|
| 224 |
+
Controls the coarseness of the binning. Smaller values lead
|
| 225 |
+
to more seeding (which is computationally more expensive). If you're
|
| 226 |
+
not sure how to set this, set it to the value of the bandwidth used
|
| 227 |
+
in clustering.mean_shift.
|
| 228 |
+
|
| 229 |
+
min_bin_freq : int, default=1
|
| 230 |
+
Only bins with at least min_bin_freq will be selected as seeds.
|
| 231 |
+
Raising this value decreases the number of seeds found, which
|
| 232 |
+
makes mean_shift computationally cheaper.
|
| 233 |
+
|
| 234 |
+
Returns
|
| 235 |
+
-------
|
| 236 |
+
bin_seeds : array-like of shape (n_samples, n_features)
|
| 237 |
+
Points used as initial kernel positions in clustering.mean_shift.
|
| 238 |
+
"""
|
| 239 |
+
if bin_size == 0:
|
| 240 |
+
return X
|
| 241 |
+
|
| 242 |
+
# Bin points
|
| 243 |
+
bin_sizes = defaultdict(int)
|
| 244 |
+
for point in X:
|
| 245 |
+
binned_point = np.round(point / bin_size)
|
| 246 |
+
bin_sizes[tuple(binned_point)] += 1
|
| 247 |
+
|
| 248 |
+
# Select only those bins as seeds which have enough members
|
| 249 |
+
bin_seeds = np.array(
|
| 250 |
+
[point for point, freq in bin_sizes.items() if freq >= min_bin_freq],
|
| 251 |
+
dtype=np.float32,
|
| 252 |
+
)
|
| 253 |
+
if len(bin_seeds) == len(X):
|
| 254 |
+
warnings.warn(
|
| 255 |
+
"Binning data failed with provided bin_size=%f, using data points as seeds."
|
| 256 |
+
% bin_size
|
| 257 |
+
)
|
| 258 |
+
return X
|
| 259 |
+
bin_seeds = bin_seeds * bin_size
|
| 260 |
+
return bin_seeds
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
class MeanShift(ClusterMixin, BaseEstimator):
|
| 264 |
+
"""Mean shift clustering using a flat kernel.
|
| 265 |
+
|
| 266 |
+
Mean shift clustering aims to discover "blobs" in a smooth density of
|
| 267 |
+
samples. It is a centroid-based algorithm, which works by updating
|
| 268 |
+
candidates for centroids to be the mean of the points within a given
|
| 269 |
+
region. These candidates are then filtered in a post-processing stage to
|
| 270 |
+
eliminate near-duplicates to form the final set of centroids.
|
| 271 |
+
|
| 272 |
+
Seeding is performed using a binning technique for scalability.
|
| 273 |
+
|
| 274 |
+
Read more in the :ref:`User Guide <mean_shift>`.
|
| 275 |
+
|
| 276 |
+
Parameters
|
| 277 |
+
----------
|
| 278 |
+
bandwidth : float, default=None
|
| 279 |
+
Bandwidth used in the flat kernel.
|
| 280 |
+
|
| 281 |
+
If not given, the bandwidth is estimated using
|
| 282 |
+
sklearn.cluster.estimate_bandwidth; see the documentation for that
|
| 283 |
+
function for hints on scalability (see also the Notes, below).
|
| 284 |
+
|
| 285 |
+
seeds : array-like of shape (n_samples, n_features), default=None
|
| 286 |
+
Seeds used to initialize kernels. If not set,
|
| 287 |
+
the seeds are calculated by clustering.get_bin_seeds
|
| 288 |
+
with bandwidth as the grid size and default values for
|
| 289 |
+
other parameters.
|
| 290 |
+
|
| 291 |
+
bin_seeding : bool, default=False
|
| 292 |
+
If true, initial kernel locations are not locations of all
|
| 293 |
+
points, but rather the location of the discretized version of
|
| 294 |
+
points, where points are binned onto a grid whose coarseness
|
| 295 |
+
corresponds to the bandwidth. Setting this option to True will speed
|
| 296 |
+
up the algorithm because fewer seeds will be initialized.
|
| 297 |
+
The default value is False.
|
| 298 |
+
Ignored if seeds argument is not None.
|
| 299 |
+
|
| 300 |
+
min_bin_freq : int, default=1
|
| 301 |
+
To speed up the algorithm, accept only those bins with at least
|
| 302 |
+
min_bin_freq points as seeds.
|
| 303 |
+
|
| 304 |
+
cluster_all : bool, default=True
|
| 305 |
+
If true, then all points are clustered, even those orphans that are
|
| 306 |
+
not within any kernel. Orphans are assigned to the nearest kernel.
|
| 307 |
+
If false, then orphans are given cluster label -1.
|
| 308 |
+
|
| 309 |
+
n_jobs : int, default=None
|
| 310 |
+
The number of jobs to use for the computation. The following tasks benefit
|
| 311 |
+
from the parallelization:
|
| 312 |
+
|
| 313 |
+
- The search of nearest neighbors for bandwidth estimation and label
|
| 314 |
+
assignments. See the details in the docstring of the
|
| 315 |
+
``NearestNeighbors`` class.
|
| 316 |
+
- Hill-climbing optimization for all seeds.
|
| 317 |
+
|
| 318 |
+
See :term:`Glossary <n_jobs>` for more details.
|
| 319 |
+
|
| 320 |
+
``None`` means 1 unless in a :obj:`joblib.parallel_backend` context.
|
| 321 |
+
``-1`` means using all processors. See :term:`Glossary <n_jobs>`
|
| 322 |
+
for more details.
|
| 323 |
+
|
| 324 |
+
max_iter : int, default=300
|
| 325 |
+
Maximum number of iterations, per seed point before the clustering
|
| 326 |
+
operation terminates (for that seed point), if has not converged yet.
|
| 327 |
+
|
| 328 |
+
.. versionadded:: 0.22
|
| 329 |
+
|
| 330 |
+
Attributes
|
| 331 |
+
----------
|
| 332 |
+
cluster_centers_ : ndarray of shape (n_clusters, n_features)
|
| 333 |
+
Coordinates of cluster centers.
|
| 334 |
+
|
| 335 |
+
labels_ : ndarray of shape (n_samples,)
|
| 336 |
+
Labels of each point.
|
| 337 |
+
|
| 338 |
+
n_iter_ : int
|
| 339 |
+
Maximum number of iterations performed on each seed.
|
| 340 |
+
|
| 341 |
+
.. versionadded:: 0.22
|
| 342 |
+
|
| 343 |
+
n_features_in_ : int
|
| 344 |
+
Number of features seen during :term:`fit`.
|
| 345 |
+
|
| 346 |
+
.. versionadded:: 0.24
|
| 347 |
+
|
| 348 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 349 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 350 |
+
has feature names that are all strings.
|
| 351 |
+
|
| 352 |
+
.. versionadded:: 1.0
|
| 353 |
+
|
| 354 |
+
See Also
|
| 355 |
+
--------
|
| 356 |
+
KMeans : K-Means clustering.
|
| 357 |
+
|
| 358 |
+
Notes
|
| 359 |
+
-----
|
| 360 |
+
|
| 361 |
+
Scalability:
|
| 362 |
+
|
| 363 |
+
Because this implementation uses a flat kernel and
|
| 364 |
+
a Ball Tree to look up members of each kernel, the complexity will tend
|
| 365 |
+
towards O(T*n*log(n)) in lower dimensions, with n the number of samples
|
| 366 |
+
and T the number of points. In higher dimensions the complexity will
|
| 367 |
+
tend towards O(T*n^2).
|
| 368 |
+
|
| 369 |
+
Scalability can be boosted by using fewer seeds, for example by using
|
| 370 |
+
a higher value of min_bin_freq in the get_bin_seeds function.
|
| 371 |
+
|
| 372 |
+
Note that the estimate_bandwidth function is much less scalable than the
|
| 373 |
+
mean shift algorithm and will be the bottleneck if it is used.
|
| 374 |
+
|
| 375 |
+
References
|
| 376 |
+
----------
|
| 377 |
+
|
| 378 |
+
Dorin Comaniciu and Peter Meer, "Mean Shift: A robust approach toward
|
| 379 |
+
feature space analysis". IEEE Transactions on Pattern Analysis and
|
| 380 |
+
Machine Intelligence. 2002. pp. 603-619.
|
| 381 |
+
|
| 382 |
+
Examples
|
| 383 |
+
--------
|
| 384 |
+
>>> from sklearn.cluster import MeanShift
|
| 385 |
+
>>> import numpy as np
|
| 386 |
+
>>> X = np.array([[1, 1], [2, 1], [1, 0],
|
| 387 |
+
... [4, 7], [3, 5], [3, 6]])
|
| 388 |
+
>>> clustering = MeanShift(bandwidth=2).fit(X)
|
| 389 |
+
>>> clustering.labels_
|
| 390 |
+
array([1, 1, 1, 0, 0, 0])
|
| 391 |
+
>>> clustering.predict([[0, 0], [5, 5]])
|
| 392 |
+
array([1, 0])
|
| 393 |
+
>>> clustering
|
| 394 |
+
MeanShift(bandwidth=2)
|
| 395 |
+
"""
|
| 396 |
+
|
| 397 |
+
_parameter_constraints: dict = {
|
| 398 |
+
"bandwidth": [Interval(Real, 0, None, closed="neither"), None],
|
| 399 |
+
"seeds": ["array-like", None],
|
| 400 |
+
"bin_seeding": ["boolean"],
|
| 401 |
+
"min_bin_freq": [Interval(Integral, 1, None, closed="left")],
|
| 402 |
+
"cluster_all": ["boolean"],
|
| 403 |
+
"n_jobs": [Integral, None],
|
| 404 |
+
"max_iter": [Interval(Integral, 0, None, closed="left")],
|
| 405 |
+
}
|
| 406 |
+
|
| 407 |
+
def __init__(
|
| 408 |
+
self,
|
| 409 |
+
*,
|
| 410 |
+
bandwidth=None,
|
| 411 |
+
seeds=None,
|
| 412 |
+
bin_seeding=False,
|
| 413 |
+
min_bin_freq=1,
|
| 414 |
+
cluster_all=True,
|
| 415 |
+
n_jobs=None,
|
| 416 |
+
max_iter=300,
|
| 417 |
+
):
|
| 418 |
+
self.bandwidth = bandwidth
|
| 419 |
+
self.seeds = seeds
|
| 420 |
+
self.bin_seeding = bin_seeding
|
| 421 |
+
self.cluster_all = cluster_all
|
| 422 |
+
self.min_bin_freq = min_bin_freq
|
| 423 |
+
self.n_jobs = n_jobs
|
| 424 |
+
self.max_iter = max_iter
|
| 425 |
+
|
| 426 |
+
def fit(self, X, y=None):
|
| 427 |
+
"""Perform clustering.
|
| 428 |
+
|
| 429 |
+
Parameters
|
| 430 |
+
----------
|
| 431 |
+
X : array-like of shape (n_samples, n_features)
|
| 432 |
+
Samples to cluster.
|
| 433 |
+
|
| 434 |
+
y : Ignored
|
| 435 |
+
Not used, present for API consistency by convention.
|
| 436 |
+
|
| 437 |
+
Returns
|
| 438 |
+
-------
|
| 439 |
+
self : object
|
| 440 |
+
Fitted instance.
|
| 441 |
+
"""
|
| 442 |
+
self._validate_params()
|
| 443 |
+
X = self._validate_data(X)
|
| 444 |
+
bandwidth = self.bandwidth
|
| 445 |
+
if bandwidth is None:
|
| 446 |
+
bandwidth = estimate_bandwidth(X, n_jobs=self.n_jobs)
|
| 447 |
+
|
| 448 |
+
seeds = self.seeds
|
| 449 |
+
if seeds is None:
|
| 450 |
+
if self.bin_seeding:
|
| 451 |
+
seeds = get_bin_seeds(X, bandwidth, self.min_bin_freq)
|
| 452 |
+
else:
|
| 453 |
+
seeds = X
|
| 454 |
+
n_samples, n_features = X.shape
|
| 455 |
+
center_intensity_dict = {}
|
| 456 |
+
|
| 457 |
+
# We use n_jobs=1 because this will be used in nested calls under
|
| 458 |
+
# parallel calls to _mean_shift_single_seed so there is no need for
|
| 459 |
+
# for further parallelism.
|
| 460 |
+
nbrs = NearestNeighbors(radius=bandwidth, n_jobs=1).fit(X)
|
| 461 |
+
|
| 462 |
+
# execute iterations on all seeds in parallel
|
| 463 |
+
all_res = Parallel(n_jobs=self.n_jobs)(
|
| 464 |
+
delayed(_mean_shift_single_seed)(seed, X, nbrs, self.max_iter)
|
| 465 |
+
for seed in seeds
|
| 466 |
+
)
|
| 467 |
+
# copy results in a dictionary
|
| 468 |
+
for i in range(len(seeds)):
|
| 469 |
+
if all_res[i][1]: # i.e. len(points_within) > 0
|
| 470 |
+
center_intensity_dict[all_res[i][0]] = all_res[i][1]
|
| 471 |
+
|
| 472 |
+
self.n_iter_ = max([x[2] for x in all_res])
|
| 473 |
+
|
| 474 |
+
if not center_intensity_dict:
|
| 475 |
+
# nothing near seeds
|
| 476 |
+
raise ValueError(
|
| 477 |
+
"No point was within bandwidth=%f of any seed. Try a different seeding"
|
| 478 |
+
" strategy or increase the bandwidth."
|
| 479 |
+
% bandwidth
|
| 480 |
+
)
|
| 481 |
+
|
| 482 |
+
# POST PROCESSING: remove near duplicate points
|
| 483 |
+
# If the distance between two kernels is less than the bandwidth,
|
| 484 |
+
# then we have to remove one because it is a duplicate. Remove the
|
| 485 |
+
# one with fewer points.
|
| 486 |
+
|
| 487 |
+
sorted_by_intensity = sorted(
|
| 488 |
+
center_intensity_dict.items(),
|
| 489 |
+
key=lambda tup: (tup[1], tup[0]),
|
| 490 |
+
reverse=True,
|
| 491 |
+
)
|
| 492 |
+
sorted_centers = np.array([tup[0] for tup in sorted_by_intensity])
|
| 493 |
+
unique = np.ones(len(sorted_centers), dtype=bool)
|
| 494 |
+
nbrs = NearestNeighbors(radius=bandwidth, n_jobs=self.n_jobs).fit(
|
| 495 |
+
sorted_centers
|
| 496 |
+
)
|
| 497 |
+
for i, center in enumerate(sorted_centers):
|
| 498 |
+
if unique[i]:
|
| 499 |
+
neighbor_idxs = nbrs.radius_neighbors([center], return_distance=False)[
|
| 500 |
+
0
|
| 501 |
+
]
|
| 502 |
+
unique[neighbor_idxs] = 0
|
| 503 |
+
unique[i] = 1 # leave the current point as unique
|
| 504 |
+
cluster_centers = sorted_centers[unique]
|
| 505 |
+
|
| 506 |
+
# ASSIGN LABELS: a point belongs to the cluster that it is closest to
|
| 507 |
+
nbrs = NearestNeighbors(n_neighbors=1, n_jobs=self.n_jobs).fit(cluster_centers)
|
| 508 |
+
labels = np.zeros(n_samples, dtype=int)
|
| 509 |
+
distances, idxs = nbrs.kneighbors(X)
|
| 510 |
+
if self.cluster_all:
|
| 511 |
+
labels = idxs.flatten()
|
| 512 |
+
else:
|
| 513 |
+
labels.fill(-1)
|
| 514 |
+
bool_selector = distances.flatten() <= bandwidth
|
| 515 |
+
labels[bool_selector] = idxs.flatten()[bool_selector]
|
| 516 |
+
|
| 517 |
+
self.cluster_centers_, self.labels_ = cluster_centers, labels
|
| 518 |
+
return self
|
| 519 |
+
|
| 520 |
+
def predict(self, X):
|
| 521 |
+
"""Predict the closest cluster each sample in X belongs to.
|
| 522 |
+
|
| 523 |
+
Parameters
|
| 524 |
+
----------
|
| 525 |
+
X : array-like of shape (n_samples, n_features)
|
| 526 |
+
New data to predict.
|
| 527 |
+
|
| 528 |
+
Returns
|
| 529 |
+
-------
|
| 530 |
+
labels : ndarray of shape (n_samples,)
|
| 531 |
+
Index of the cluster each sample belongs to.
|
| 532 |
+
"""
|
| 533 |
+
check_is_fitted(self)
|
| 534 |
+
X = self._validate_data(X, reset=False)
|
| 535 |
+
with config_context(assume_finite=True):
|
| 536 |
+
return pairwise_distances_argmin(X, self.cluster_centers_)
|
videollama2/lib/python3.10/site-packages/sklearn/cluster/_spectral.py
ADDED
|
@@ -0,0 +1,791 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Algorithms for spectral clustering"""
|
| 2 |
+
|
| 3 |
+
# Author: Gael Varoquaux <gael.varoquaux@normalesup.org>
|
| 4 |
+
# Brian Cheung
|
| 5 |
+
# Wei LI <kuantkid@gmail.com>
|
| 6 |
+
# Andrew Knyazev <Andrew.Knyazev@ucdenver.edu>
|
| 7 |
+
# License: BSD 3 clause
|
| 8 |
+
|
| 9 |
+
from numbers import Integral, Real
|
| 10 |
+
import warnings
|
| 11 |
+
|
| 12 |
+
import numpy as np
|
| 13 |
+
|
| 14 |
+
from scipy.linalg import LinAlgError, qr, svd
|
| 15 |
+
from scipy.sparse import csc_matrix
|
| 16 |
+
|
| 17 |
+
from ..base import BaseEstimator, ClusterMixin
|
| 18 |
+
from ..utils._param_validation import Interval, StrOptions
|
| 19 |
+
from ..utils import check_random_state, as_float_array
|
| 20 |
+
from ..metrics.pairwise import pairwise_kernels, KERNEL_PARAMS
|
| 21 |
+
from ..neighbors import kneighbors_graph, NearestNeighbors
|
| 22 |
+
from ..manifold import spectral_embedding
|
| 23 |
+
from ._kmeans import k_means
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
def cluster_qr(vectors):
|
| 27 |
+
"""Find the discrete partition closest to the eigenvector embedding.
|
| 28 |
+
|
| 29 |
+
This implementation was proposed in [1]_.
|
| 30 |
+
|
| 31 |
+
.. versionadded:: 1.1
|
| 32 |
+
|
| 33 |
+
Parameters
|
| 34 |
+
----------
|
| 35 |
+
vectors : array-like, shape: (n_samples, n_clusters)
|
| 36 |
+
The embedding space of the samples.
|
| 37 |
+
|
| 38 |
+
Returns
|
| 39 |
+
-------
|
| 40 |
+
labels : array of integers, shape: n_samples
|
| 41 |
+
The cluster labels of vectors.
|
| 42 |
+
|
| 43 |
+
References
|
| 44 |
+
----------
|
| 45 |
+
.. [1] :doi:`Simple, direct, and efficient multi-way spectral clustering, 2019
|
| 46 |
+
Anil Damle, Victor Minden, Lexing Ying
|
| 47 |
+
<10.1093/imaiai/iay008>`
|
| 48 |
+
|
| 49 |
+
"""
|
| 50 |
+
|
| 51 |
+
k = vectors.shape[1]
|
| 52 |
+
_, _, piv = qr(vectors.T, pivoting=True)
|
| 53 |
+
ut, _, v = svd(vectors[piv[:k], :].T)
|
| 54 |
+
vectors = abs(np.dot(vectors, np.dot(ut, v.conj())))
|
| 55 |
+
return vectors.argmax(axis=1)
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
def discretize(
|
| 59 |
+
vectors, *, copy=True, max_svd_restarts=30, n_iter_max=20, random_state=None
|
| 60 |
+
):
|
| 61 |
+
"""Search for a partition matrix which is closest to the eigenvector embedding.
|
| 62 |
+
|
| 63 |
+
This implementation was proposed in [1]_.
|
| 64 |
+
|
| 65 |
+
Parameters
|
| 66 |
+
----------
|
| 67 |
+
vectors : array-like of shape (n_samples, n_clusters)
|
| 68 |
+
The embedding space of the samples.
|
| 69 |
+
|
| 70 |
+
copy : bool, default=True
|
| 71 |
+
Whether to copy vectors, or perform in-place normalization.
|
| 72 |
+
|
| 73 |
+
max_svd_restarts : int, default=30
|
| 74 |
+
Maximum number of attempts to restart SVD if convergence fails
|
| 75 |
+
|
| 76 |
+
n_iter_max : int, default=30
|
| 77 |
+
Maximum number of iterations to attempt in rotation and partition
|
| 78 |
+
matrix search if machine precision convergence is not reached
|
| 79 |
+
|
| 80 |
+
random_state : int, RandomState instance, default=None
|
| 81 |
+
Determines random number generation for rotation matrix initialization.
|
| 82 |
+
Use an int to make the randomness deterministic.
|
| 83 |
+
See :term:`Glossary <random_state>`.
|
| 84 |
+
|
| 85 |
+
Returns
|
| 86 |
+
-------
|
| 87 |
+
labels : array of integers, shape: n_samples
|
| 88 |
+
The labels of the clusters.
|
| 89 |
+
|
| 90 |
+
References
|
| 91 |
+
----------
|
| 92 |
+
|
| 93 |
+
.. [1] `Multiclass spectral clustering, 2003
|
| 94 |
+
Stella X. Yu, Jianbo Shi
|
| 95 |
+
<https://people.eecs.berkeley.edu/~jordan/courses/281B-spring04/readings/yu-shi.pdf>`_
|
| 96 |
+
|
| 97 |
+
Notes
|
| 98 |
+
-----
|
| 99 |
+
|
| 100 |
+
The eigenvector embedding is used to iteratively search for the
|
| 101 |
+
closest discrete partition. First, the eigenvector embedding is
|
| 102 |
+
normalized to the space of partition matrices. An optimal discrete
|
| 103 |
+
partition matrix closest to this normalized embedding multiplied by
|
| 104 |
+
an initial rotation is calculated. Fixing this discrete partition
|
| 105 |
+
matrix, an optimal rotation matrix is calculated. These two
|
| 106 |
+
calculations are performed until convergence. The discrete partition
|
| 107 |
+
matrix is returned as the clustering solution. Used in spectral
|
| 108 |
+
clustering, this method tends to be faster and more robust to random
|
| 109 |
+
initialization than k-means.
|
| 110 |
+
|
| 111 |
+
"""
|
| 112 |
+
|
| 113 |
+
random_state = check_random_state(random_state)
|
| 114 |
+
|
| 115 |
+
vectors = as_float_array(vectors, copy=copy)
|
| 116 |
+
|
| 117 |
+
eps = np.finfo(float).eps
|
| 118 |
+
n_samples, n_components = vectors.shape
|
| 119 |
+
|
| 120 |
+
# Normalize the eigenvectors to an equal length of a vector of ones.
|
| 121 |
+
# Reorient the eigenvectors to point in the negative direction with respect
|
| 122 |
+
# to the first element. This may have to do with constraining the
|
| 123 |
+
# eigenvectors to lie in a specific quadrant to make the discretization
|
| 124 |
+
# search easier.
|
| 125 |
+
norm_ones = np.sqrt(n_samples)
|
| 126 |
+
for i in range(vectors.shape[1]):
|
| 127 |
+
vectors[:, i] = (vectors[:, i] / np.linalg.norm(vectors[:, i])) * norm_ones
|
| 128 |
+
if vectors[0, i] != 0:
|
| 129 |
+
vectors[:, i] = -1 * vectors[:, i] * np.sign(vectors[0, i])
|
| 130 |
+
|
| 131 |
+
# Normalize the rows of the eigenvectors. Samples should lie on the unit
|
| 132 |
+
# hypersphere centered at the origin. This transforms the samples in the
|
| 133 |
+
# embedding space to the space of partition matrices.
|
| 134 |
+
vectors = vectors / np.sqrt((vectors**2).sum(axis=1))[:, np.newaxis]
|
| 135 |
+
|
| 136 |
+
svd_restarts = 0
|
| 137 |
+
has_converged = False
|
| 138 |
+
|
| 139 |
+
# If there is an exception we try to randomize and rerun SVD again
|
| 140 |
+
# do this max_svd_restarts times.
|
| 141 |
+
while (svd_restarts < max_svd_restarts) and not has_converged:
|
| 142 |
+
|
| 143 |
+
# Initialize first column of rotation matrix with a row of the
|
| 144 |
+
# eigenvectors
|
| 145 |
+
rotation = np.zeros((n_components, n_components))
|
| 146 |
+
rotation[:, 0] = vectors[random_state.randint(n_samples), :].T
|
| 147 |
+
|
| 148 |
+
# To initialize the rest of the rotation matrix, find the rows
|
| 149 |
+
# of the eigenvectors that are as orthogonal to each other as
|
| 150 |
+
# possible
|
| 151 |
+
c = np.zeros(n_samples)
|
| 152 |
+
for j in range(1, n_components):
|
| 153 |
+
# Accumulate c to ensure row is as orthogonal as possible to
|
| 154 |
+
# previous picks as well as current one
|
| 155 |
+
c += np.abs(np.dot(vectors, rotation[:, j - 1]))
|
| 156 |
+
rotation[:, j] = vectors[c.argmin(), :].T
|
| 157 |
+
|
| 158 |
+
last_objective_value = 0.0
|
| 159 |
+
n_iter = 0
|
| 160 |
+
|
| 161 |
+
while not has_converged:
|
| 162 |
+
n_iter += 1
|
| 163 |
+
|
| 164 |
+
t_discrete = np.dot(vectors, rotation)
|
| 165 |
+
|
| 166 |
+
labels = t_discrete.argmax(axis=1)
|
| 167 |
+
vectors_discrete = csc_matrix(
|
| 168 |
+
(np.ones(len(labels)), (np.arange(0, n_samples), labels)),
|
| 169 |
+
shape=(n_samples, n_components),
|
| 170 |
+
)
|
| 171 |
+
|
| 172 |
+
t_svd = vectors_discrete.T * vectors
|
| 173 |
+
|
| 174 |
+
try:
|
| 175 |
+
U, S, Vh = np.linalg.svd(t_svd)
|
| 176 |
+
except LinAlgError:
|
| 177 |
+
svd_restarts += 1
|
| 178 |
+
print("SVD did not converge, randomizing and trying again")
|
| 179 |
+
break
|
| 180 |
+
|
| 181 |
+
ncut_value = 2.0 * (n_samples - S.sum())
|
| 182 |
+
if (abs(ncut_value - last_objective_value) < eps) or (n_iter > n_iter_max):
|
| 183 |
+
has_converged = True
|
| 184 |
+
else:
|
| 185 |
+
# otherwise calculate rotation and continue
|
| 186 |
+
last_objective_value = ncut_value
|
| 187 |
+
rotation = np.dot(Vh.T, U.T)
|
| 188 |
+
|
| 189 |
+
if not has_converged:
|
| 190 |
+
raise LinAlgError("SVD did not converge")
|
| 191 |
+
return labels
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
def spectral_clustering(
|
| 195 |
+
affinity,
|
| 196 |
+
*,
|
| 197 |
+
n_clusters=8,
|
| 198 |
+
n_components=None,
|
| 199 |
+
eigen_solver=None,
|
| 200 |
+
random_state=None,
|
| 201 |
+
n_init=10,
|
| 202 |
+
eigen_tol="auto",
|
| 203 |
+
assign_labels="kmeans",
|
| 204 |
+
verbose=False,
|
| 205 |
+
):
|
| 206 |
+
"""Apply clustering to a projection of the normalized Laplacian.
|
| 207 |
+
|
| 208 |
+
In practice Spectral Clustering is very useful when the structure of
|
| 209 |
+
the individual clusters is highly non-convex or more generally when
|
| 210 |
+
a measure of the center and spread of the cluster is not a suitable
|
| 211 |
+
description of the complete cluster. For instance, when clusters are
|
| 212 |
+
nested circles on the 2D plane.
|
| 213 |
+
|
| 214 |
+
If affinity is the adjacency matrix of a graph, this method can be
|
| 215 |
+
used to find normalized graph cuts [1]_, [2]_.
|
| 216 |
+
|
| 217 |
+
Read more in the :ref:`User Guide <spectral_clustering>`.
|
| 218 |
+
|
| 219 |
+
Parameters
|
| 220 |
+
----------
|
| 221 |
+
affinity : {array-like, sparse matrix} of shape (n_samples, n_samples)
|
| 222 |
+
The affinity matrix describing the relationship of the samples to
|
| 223 |
+
embed. **Must be symmetric**.
|
| 224 |
+
|
| 225 |
+
Possible examples:
|
| 226 |
+
- adjacency matrix of a graph,
|
| 227 |
+
- heat kernel of the pairwise distance matrix of the samples,
|
| 228 |
+
- symmetric k-nearest neighbours connectivity matrix of the samples.
|
| 229 |
+
|
| 230 |
+
n_clusters : int, default=None
|
| 231 |
+
Number of clusters to extract.
|
| 232 |
+
|
| 233 |
+
n_components : int, default=n_clusters
|
| 234 |
+
Number of eigenvectors to use for the spectral embedding.
|
| 235 |
+
|
| 236 |
+
eigen_solver : {None, 'arpack', 'lobpcg', or 'amg'}
|
| 237 |
+
The eigenvalue decomposition method. If None then ``'arpack'`` is used.
|
| 238 |
+
See [4]_ for more details regarding ``'lobpcg'``.
|
| 239 |
+
Eigensolver ``'amg'`` runs ``'lobpcg'`` with optional
|
| 240 |
+
Algebraic MultiGrid preconditioning and requires pyamg to be installed.
|
| 241 |
+
It can be faster on very large sparse problems [6]_ and [7]_.
|
| 242 |
+
|
| 243 |
+
random_state : int, RandomState instance, default=None
|
| 244 |
+
A pseudo random number generator used for the initialization
|
| 245 |
+
of the lobpcg eigenvectors decomposition when `eigen_solver ==
|
| 246 |
+
'amg'`, and for the K-Means initialization. Use an int to make
|
| 247 |
+
the results deterministic across calls (See
|
| 248 |
+
:term:`Glossary <random_state>`).
|
| 249 |
+
|
| 250 |
+
.. note::
|
| 251 |
+
When using `eigen_solver == 'amg'`,
|
| 252 |
+
it is necessary to also fix the global numpy seed with
|
| 253 |
+
`np.random.seed(int)` to get deterministic results. See
|
| 254 |
+
https://github.com/pyamg/pyamg/issues/139 for further
|
| 255 |
+
information.
|
| 256 |
+
|
| 257 |
+
n_init : int, default=10
|
| 258 |
+
Number of time the k-means algorithm will be run with different
|
| 259 |
+
centroid seeds. The final results will be the best output of n_init
|
| 260 |
+
consecutive runs in terms of inertia. Only used if
|
| 261 |
+
``assign_labels='kmeans'``.
|
| 262 |
+
|
| 263 |
+
eigen_tol : float, default="auto"
|
| 264 |
+
Stopping criterion for eigendecomposition of the Laplacian matrix.
|
| 265 |
+
If `eigen_tol="auto"` then the passed tolerance will depend on the
|
| 266 |
+
`eigen_solver`:
|
| 267 |
+
|
| 268 |
+
- If `eigen_solver="arpack"`, then `eigen_tol=0.0`;
|
| 269 |
+
- If `eigen_solver="lobpcg"` or `eigen_solver="amg"`, then
|
| 270 |
+
`eigen_tol=None` which configures the underlying `lobpcg` solver to
|
| 271 |
+
automatically resolve the value according to their heuristics. See,
|
| 272 |
+
:func:`scipy.sparse.linalg.lobpcg` for details.
|
| 273 |
+
|
| 274 |
+
Note that when using `eigen_solver="lobpcg"` or `eigen_solver="amg"`
|
| 275 |
+
values of `tol<1e-5` may lead to convergence issues and should be
|
| 276 |
+
avoided.
|
| 277 |
+
|
| 278 |
+
.. versionadded:: 1.2
|
| 279 |
+
Added 'auto' option.
|
| 280 |
+
|
| 281 |
+
assign_labels : {'kmeans', 'discretize', 'cluster_qr'}, default='kmeans'
|
| 282 |
+
The strategy to use to assign labels in the embedding
|
| 283 |
+
space. There are three ways to assign labels after the Laplacian
|
| 284 |
+
embedding. k-means can be applied and is a popular choice. But it can
|
| 285 |
+
also be sensitive to initialization. Discretization is another
|
| 286 |
+
approach which is less sensitive to random initialization [3]_.
|
| 287 |
+
The cluster_qr method [5]_ directly extracts clusters from eigenvectors
|
| 288 |
+
in spectral clustering. In contrast to k-means and discretization, cluster_qr
|
| 289 |
+
has no tuning parameters and is not an iterative method, yet may outperform
|
| 290 |
+
k-means and discretization in terms of both quality and speed.
|
| 291 |
+
|
| 292 |
+
.. versionchanged:: 1.1
|
| 293 |
+
Added new labeling method 'cluster_qr'.
|
| 294 |
+
|
| 295 |
+
verbose : bool, default=False
|
| 296 |
+
Verbosity mode.
|
| 297 |
+
|
| 298 |
+
.. versionadded:: 0.24
|
| 299 |
+
|
| 300 |
+
Returns
|
| 301 |
+
-------
|
| 302 |
+
labels : array of integers, shape: n_samples
|
| 303 |
+
The labels of the clusters.
|
| 304 |
+
|
| 305 |
+
Notes
|
| 306 |
+
-----
|
| 307 |
+
The graph should contain only one connected component, elsewhere
|
| 308 |
+
the results make little sense.
|
| 309 |
+
|
| 310 |
+
This algorithm solves the normalized cut for `k=2`: it is a
|
| 311 |
+
normalized spectral clustering.
|
| 312 |
+
|
| 313 |
+
References
|
| 314 |
+
----------
|
| 315 |
+
|
| 316 |
+
.. [1] :doi:`Normalized cuts and image segmentation, 2000
|
| 317 |
+
Jianbo Shi, Jitendra Malik
|
| 318 |
+
<10.1109/34.868688>`
|
| 319 |
+
|
| 320 |
+
.. [2] :doi:`A Tutorial on Spectral Clustering, 2007
|
| 321 |
+
Ulrike von Luxburg
|
| 322 |
+
<10.1007/s11222-007-9033-z>`
|
| 323 |
+
|
| 324 |
+
.. [3] `Multiclass spectral clustering, 2003
|
| 325 |
+
Stella X. Yu, Jianbo Shi
|
| 326 |
+
<https://people.eecs.berkeley.edu/~jordan/courses/281B-spring04/readings/yu-shi.pdf>`_
|
| 327 |
+
|
| 328 |
+
.. [4] :doi:`Toward the Optimal Preconditioned Eigensolver:
|
| 329 |
+
Locally Optimal Block Preconditioned Conjugate Gradient Method, 2001
|
| 330 |
+
A. V. Knyazev
|
| 331 |
+
SIAM Journal on Scientific Computing 23, no. 2, pp. 517-541.
|
| 332 |
+
<10.1137/S1064827500366124>`
|
| 333 |
+
|
| 334 |
+
.. [5] :doi:`Simple, direct, and efficient multi-way spectral clustering, 2019
|
| 335 |
+
Anil Damle, Victor Minden, Lexing Ying
|
| 336 |
+
<10.1093/imaiai/iay008>`
|
| 337 |
+
|
| 338 |
+
.. [6] :doi:`Multiscale Spectral Image Segmentation Multiscale preconditioning
|
| 339 |
+
for computing eigenvalues of graph Laplacians in image segmentation, 2006
|
| 340 |
+
Andrew Knyazev
|
| 341 |
+
<10.13140/RG.2.2.35280.02565>`
|
| 342 |
+
|
| 343 |
+
.. [7] :doi:`Preconditioned spectral clustering for stochastic block partition
|
| 344 |
+
streaming graph challenge (Preliminary version at arXiv.)
|
| 345 |
+
David Zhuzhunashvili, Andrew Knyazev
|
| 346 |
+
<10.1109/HPEC.2017.8091045>`
|
| 347 |
+
"""
|
| 348 |
+
if assign_labels not in ("kmeans", "discretize", "cluster_qr"):
|
| 349 |
+
raise ValueError(
|
| 350 |
+
"The 'assign_labels' parameter should be "
|
| 351 |
+
"'kmeans' or 'discretize', or 'cluster_qr', "
|
| 352 |
+
f"but {assign_labels!r} was given"
|
| 353 |
+
)
|
| 354 |
+
if isinstance(affinity, np.matrix):
|
| 355 |
+
raise TypeError(
|
| 356 |
+
"spectral_clustering does not support passing in affinity as an "
|
| 357 |
+
"np.matrix. Please convert to a numpy array with np.asarray. For "
|
| 358 |
+
"more information see: "
|
| 359 |
+
"https://numpy.org/doc/stable/reference/generated/numpy.matrix.html", # noqa
|
| 360 |
+
)
|
| 361 |
+
|
| 362 |
+
random_state = check_random_state(random_state)
|
| 363 |
+
n_components = n_clusters if n_components is None else n_components
|
| 364 |
+
|
| 365 |
+
# We now obtain the real valued solution matrix to the
|
| 366 |
+
# relaxed Ncut problem, solving the eigenvalue problem
|
| 367 |
+
# L_sym x = lambda x and recovering u = D^-1/2 x.
|
| 368 |
+
# The first eigenvector is constant only for fully connected graphs
|
| 369 |
+
# and should be kept for spectral clustering (drop_first = False)
|
| 370 |
+
# See spectral_embedding documentation.
|
| 371 |
+
maps = spectral_embedding(
|
| 372 |
+
affinity,
|
| 373 |
+
n_components=n_components,
|
| 374 |
+
eigen_solver=eigen_solver,
|
| 375 |
+
random_state=random_state,
|
| 376 |
+
eigen_tol=eigen_tol,
|
| 377 |
+
drop_first=False,
|
| 378 |
+
)
|
| 379 |
+
if verbose:
|
| 380 |
+
print(f"Computing label assignment using {assign_labels}")
|
| 381 |
+
|
| 382 |
+
if assign_labels == "kmeans":
|
| 383 |
+
_, labels, _ = k_means(
|
| 384 |
+
maps, n_clusters, random_state=random_state, n_init=n_init, verbose=verbose
|
| 385 |
+
)
|
| 386 |
+
elif assign_labels == "cluster_qr":
|
| 387 |
+
labels = cluster_qr(maps)
|
| 388 |
+
else:
|
| 389 |
+
labels = discretize(maps, random_state=random_state)
|
| 390 |
+
|
| 391 |
+
return labels
|
| 392 |
+
|
| 393 |
+
|
| 394 |
+
class SpectralClustering(ClusterMixin, BaseEstimator):
|
| 395 |
+
"""Apply clustering to a projection of the normalized Laplacian.
|
| 396 |
+
|
| 397 |
+
In practice Spectral Clustering is very useful when the structure of
|
| 398 |
+
the individual clusters is highly non-convex, or more generally when
|
| 399 |
+
a measure of the center and spread of the cluster is not a suitable
|
| 400 |
+
description of the complete cluster, such as when clusters are
|
| 401 |
+
nested circles on the 2D plane.
|
| 402 |
+
|
| 403 |
+
If the affinity matrix is the adjacency matrix of a graph, this method
|
| 404 |
+
can be used to find normalized graph cuts [1]_, [2]_.
|
| 405 |
+
|
| 406 |
+
When calling ``fit``, an affinity matrix is constructed using either
|
| 407 |
+
a kernel function such the Gaussian (aka RBF) kernel with Euclidean
|
| 408 |
+
distance ``d(X, X)``::
|
| 409 |
+
|
| 410 |
+
np.exp(-gamma * d(X,X) ** 2)
|
| 411 |
+
|
| 412 |
+
or a k-nearest neighbors connectivity matrix.
|
| 413 |
+
|
| 414 |
+
Alternatively, a user-provided affinity matrix can be specified by
|
| 415 |
+
setting ``affinity='precomputed'``.
|
| 416 |
+
|
| 417 |
+
Read more in the :ref:`User Guide <spectral_clustering>`.
|
| 418 |
+
|
| 419 |
+
Parameters
|
| 420 |
+
----------
|
| 421 |
+
n_clusters : int, default=8
|
| 422 |
+
The dimension of the projection subspace.
|
| 423 |
+
|
| 424 |
+
eigen_solver : {'arpack', 'lobpcg', 'amg'}, default=None
|
| 425 |
+
The eigenvalue decomposition strategy to use. AMG requires pyamg
|
| 426 |
+
to be installed. It can be faster on very large, sparse problems,
|
| 427 |
+
but may also lead to instabilities. If None, then ``'arpack'`` is
|
| 428 |
+
used. See [4]_ for more details regarding `'lobpcg'`.
|
| 429 |
+
|
| 430 |
+
n_components : int, default=None
|
| 431 |
+
Number of eigenvectors to use for the spectral embedding. If None,
|
| 432 |
+
defaults to `n_clusters`.
|
| 433 |
+
|
| 434 |
+
random_state : int, RandomState instance, default=None
|
| 435 |
+
A pseudo random number generator used for the initialization
|
| 436 |
+
of the lobpcg eigenvectors decomposition when `eigen_solver ==
|
| 437 |
+
'amg'`, and for the K-Means initialization. Use an int to make
|
| 438 |
+
the results deterministic across calls (See
|
| 439 |
+
:term:`Glossary <random_state>`).
|
| 440 |
+
|
| 441 |
+
.. note::
|
| 442 |
+
When using `eigen_solver == 'amg'`,
|
| 443 |
+
it is necessary to also fix the global numpy seed with
|
| 444 |
+
`np.random.seed(int)` to get deterministic results. See
|
| 445 |
+
https://github.com/pyamg/pyamg/issues/139 for further
|
| 446 |
+
information.
|
| 447 |
+
|
| 448 |
+
n_init : int, default=10
|
| 449 |
+
Number of time the k-means algorithm will be run with different
|
| 450 |
+
centroid seeds. The final results will be the best output of n_init
|
| 451 |
+
consecutive runs in terms of inertia. Only used if
|
| 452 |
+
``assign_labels='kmeans'``.
|
| 453 |
+
|
| 454 |
+
gamma : float, default=1.0
|
| 455 |
+
Kernel coefficient for rbf, poly, sigmoid, laplacian and chi2 kernels.
|
| 456 |
+
Ignored for ``affinity='nearest_neighbors'``.
|
| 457 |
+
|
| 458 |
+
affinity : str or callable, default='rbf'
|
| 459 |
+
How to construct the affinity matrix.
|
| 460 |
+
- 'nearest_neighbors': construct the affinity matrix by computing a
|
| 461 |
+
graph of nearest neighbors.
|
| 462 |
+
- 'rbf': construct the affinity matrix using a radial basis function
|
| 463 |
+
(RBF) kernel.
|
| 464 |
+
- 'precomputed': interpret ``X`` as a precomputed affinity matrix,
|
| 465 |
+
where larger values indicate greater similarity between instances.
|
| 466 |
+
- 'precomputed_nearest_neighbors': interpret ``X`` as a sparse graph
|
| 467 |
+
of precomputed distances, and construct a binary affinity matrix
|
| 468 |
+
from the ``n_neighbors`` nearest neighbors of each instance.
|
| 469 |
+
- one of the kernels supported by
|
| 470 |
+
:func:`~sklearn.metrics.pairwise_kernels`.
|
| 471 |
+
|
| 472 |
+
Only kernels that produce similarity scores (non-negative values that
|
| 473 |
+
increase with similarity) should be used. This property is not checked
|
| 474 |
+
by the clustering algorithm.
|
| 475 |
+
|
| 476 |
+
n_neighbors : int, default=10
|
| 477 |
+
Number of neighbors to use when constructing the affinity matrix using
|
| 478 |
+
the nearest neighbors method. Ignored for ``affinity='rbf'``.
|
| 479 |
+
|
| 480 |
+
eigen_tol : float, default="auto"
|
| 481 |
+
Stopping criterion for eigendecomposition of the Laplacian matrix.
|
| 482 |
+
If `eigen_tol="auto"` then the passed tolerance will depend on the
|
| 483 |
+
`eigen_solver`:
|
| 484 |
+
|
| 485 |
+
- If `eigen_solver="arpack"`, then `eigen_tol=0.0`;
|
| 486 |
+
- If `eigen_solver="lobpcg"` or `eigen_solver="amg"`, then
|
| 487 |
+
`eigen_tol=None` which configures the underlying `lobpcg` solver to
|
| 488 |
+
automatically resolve the value according to their heuristics. See,
|
| 489 |
+
:func:`scipy.sparse.linalg.lobpcg` for details.
|
| 490 |
+
|
| 491 |
+
Note that when using `eigen_solver="lobpcg"` or `eigen_solver="amg"`
|
| 492 |
+
values of `tol<1e-5` may lead to convergence issues and should be
|
| 493 |
+
avoided.
|
| 494 |
+
|
| 495 |
+
.. versionadded:: 1.2
|
| 496 |
+
Added 'auto' option.
|
| 497 |
+
|
| 498 |
+
assign_labels : {'kmeans', 'discretize', 'cluster_qr'}, default='kmeans'
|
| 499 |
+
The strategy for assigning labels in the embedding space. There are two
|
| 500 |
+
ways to assign labels after the Laplacian embedding. k-means is a
|
| 501 |
+
popular choice, but it can be sensitive to initialization.
|
| 502 |
+
Discretization is another approach which is less sensitive to random
|
| 503 |
+
initialization [3]_.
|
| 504 |
+
The cluster_qr method [5]_ directly extract clusters from eigenvectors
|
| 505 |
+
in spectral clustering. In contrast to k-means and discretization, cluster_qr
|
| 506 |
+
has no tuning parameters and runs no iterations, yet may outperform
|
| 507 |
+
k-means and discretization in terms of both quality and speed.
|
| 508 |
+
|
| 509 |
+
.. versionchanged:: 1.1
|
| 510 |
+
Added new labeling method 'cluster_qr'.
|
| 511 |
+
|
| 512 |
+
degree : float, default=3
|
| 513 |
+
Degree of the polynomial kernel. Ignored by other kernels.
|
| 514 |
+
|
| 515 |
+
coef0 : float, default=1
|
| 516 |
+
Zero coefficient for polynomial and sigmoid kernels.
|
| 517 |
+
Ignored by other kernels.
|
| 518 |
+
|
| 519 |
+
kernel_params : dict of str to any, default=None
|
| 520 |
+
Parameters (keyword arguments) and values for kernel passed as
|
| 521 |
+
callable object. Ignored by other kernels.
|
| 522 |
+
|
| 523 |
+
n_jobs : int, default=None
|
| 524 |
+
The number of parallel jobs to run when `affinity='nearest_neighbors'`
|
| 525 |
+
or `affinity='precomputed_nearest_neighbors'`. The neighbors search
|
| 526 |
+
will be done in parallel.
|
| 527 |
+
``None`` means 1 unless in a :obj:`joblib.parallel_backend` context.
|
| 528 |
+
``-1`` means using all processors. See :term:`Glossary <n_jobs>`
|
| 529 |
+
for more details.
|
| 530 |
+
|
| 531 |
+
verbose : bool, default=False
|
| 532 |
+
Verbosity mode.
|
| 533 |
+
|
| 534 |
+
.. versionadded:: 0.24
|
| 535 |
+
|
| 536 |
+
Attributes
|
| 537 |
+
----------
|
| 538 |
+
affinity_matrix_ : array-like of shape (n_samples, n_samples)
|
| 539 |
+
Affinity matrix used for clustering. Available only after calling
|
| 540 |
+
``fit``.
|
| 541 |
+
|
| 542 |
+
labels_ : ndarray of shape (n_samples,)
|
| 543 |
+
Labels of each point
|
| 544 |
+
|
| 545 |
+
n_features_in_ : int
|
| 546 |
+
Number of features seen during :term:`fit`.
|
| 547 |
+
|
| 548 |
+
.. versionadded:: 0.24
|
| 549 |
+
|
| 550 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 551 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 552 |
+
has feature names that are all strings.
|
| 553 |
+
|
| 554 |
+
.. versionadded:: 1.0
|
| 555 |
+
|
| 556 |
+
See Also
|
| 557 |
+
--------
|
| 558 |
+
sklearn.cluster.KMeans : K-Means clustering.
|
| 559 |
+
sklearn.cluster.DBSCAN : Density-Based Spatial Clustering of
|
| 560 |
+
Applications with Noise.
|
| 561 |
+
|
| 562 |
+
Notes
|
| 563 |
+
-----
|
| 564 |
+
A distance matrix for which 0 indicates identical elements and high values
|
| 565 |
+
indicate very dissimilar elements can be transformed into an affinity /
|
| 566 |
+
similarity matrix that is well-suited for the algorithm by
|
| 567 |
+
applying the Gaussian (aka RBF, heat) kernel::
|
| 568 |
+
|
| 569 |
+
np.exp(- dist_matrix ** 2 / (2. * delta ** 2))
|
| 570 |
+
|
| 571 |
+
where ``delta`` is a free parameter representing the width of the Gaussian
|
| 572 |
+
kernel.
|
| 573 |
+
|
| 574 |
+
An alternative is to take a symmetric version of the k-nearest neighbors
|
| 575 |
+
connectivity matrix of the points.
|
| 576 |
+
|
| 577 |
+
If the pyamg package is installed, it is used: this greatly
|
| 578 |
+
speeds up computation.
|
| 579 |
+
|
| 580 |
+
References
|
| 581 |
+
----------
|
| 582 |
+
.. [1] :doi:`Normalized cuts and image segmentation, 2000
|
| 583 |
+
Jianbo Shi, Jitendra Malik
|
| 584 |
+
<10.1109/34.868688>`
|
| 585 |
+
|
| 586 |
+
.. [2] :doi:`A Tutorial on Spectral Clustering, 2007
|
| 587 |
+
Ulrike von Luxburg
|
| 588 |
+
<10.1007/s11222-007-9033-z>`
|
| 589 |
+
|
| 590 |
+
.. [3] `Multiclass spectral clustering, 2003
|
| 591 |
+
Stella X. Yu, Jianbo Shi
|
| 592 |
+
<https://people.eecs.berkeley.edu/~jordan/courses/281B-spring04/readings/yu-shi.pdf>`_
|
| 593 |
+
|
| 594 |
+
.. [4] :doi:`Toward the Optimal Preconditioned Eigensolver:
|
| 595 |
+
Locally Optimal Block Preconditioned Conjugate Gradient Method, 2001
|
| 596 |
+
A. V. Knyazev
|
| 597 |
+
SIAM Journal on Scientific Computing 23, no. 2, pp. 517-541.
|
| 598 |
+
<10.1137/S1064827500366124>`
|
| 599 |
+
|
| 600 |
+
.. [5] :doi:`Simple, direct, and efficient multi-way spectral clustering, 2019
|
| 601 |
+
Anil Damle, Victor Minden, Lexing Ying
|
| 602 |
+
<10.1093/imaiai/iay008>`
|
| 603 |
+
|
| 604 |
+
Examples
|
| 605 |
+
--------
|
| 606 |
+
>>> from sklearn.cluster import SpectralClustering
|
| 607 |
+
>>> import numpy as np
|
| 608 |
+
>>> X = np.array([[1, 1], [2, 1], [1, 0],
|
| 609 |
+
... [4, 7], [3, 5], [3, 6]])
|
| 610 |
+
>>> clustering = SpectralClustering(n_clusters=2,
|
| 611 |
+
... assign_labels='discretize',
|
| 612 |
+
... random_state=0).fit(X)
|
| 613 |
+
>>> clustering.labels_
|
| 614 |
+
array([1, 1, 1, 0, 0, 0])
|
| 615 |
+
>>> clustering
|
| 616 |
+
SpectralClustering(assign_labels='discretize', n_clusters=2,
|
| 617 |
+
random_state=0)
|
| 618 |
+
"""
|
| 619 |
+
|
| 620 |
+
_parameter_constraints: dict = {
|
| 621 |
+
"n_clusters": [Interval(Integral, 1, None, closed="left")],
|
| 622 |
+
"eigen_solver": [StrOptions({"arpack", "lobpcg", "amg"}), None],
|
| 623 |
+
"n_components": [Interval(Integral, 1, None, closed="left"), None],
|
| 624 |
+
"random_state": ["random_state"],
|
| 625 |
+
"n_init": [Interval(Integral, 1, None, closed="left")],
|
| 626 |
+
"gamma": [Interval(Real, 0, None, closed="left")],
|
| 627 |
+
"affinity": [
|
| 628 |
+
callable,
|
| 629 |
+
StrOptions(
|
| 630 |
+
set(KERNEL_PARAMS)
|
| 631 |
+
| {"nearest_neighbors", "precomputed", "precomputed_nearest_neighbors"}
|
| 632 |
+
),
|
| 633 |
+
],
|
| 634 |
+
"n_neighbors": [Interval(Integral, 1, None, closed="left")],
|
| 635 |
+
"eigen_tol": [
|
| 636 |
+
Interval(Real, 0.0, None, closed="left"),
|
| 637 |
+
StrOptions({"auto"}),
|
| 638 |
+
],
|
| 639 |
+
"assign_labels": [StrOptions({"kmeans", "discretize", "cluster_qr"})],
|
| 640 |
+
"degree": [Interval(Integral, 0, None, closed="left")],
|
| 641 |
+
"coef0": [Interval(Real, None, None, closed="neither")],
|
| 642 |
+
"kernel_params": [dict, None],
|
| 643 |
+
"n_jobs": [Integral, None],
|
| 644 |
+
"verbose": ["verbose"],
|
| 645 |
+
}
|
| 646 |
+
|
| 647 |
+
def __init__(
|
| 648 |
+
self,
|
| 649 |
+
n_clusters=8,
|
| 650 |
+
*,
|
| 651 |
+
eigen_solver=None,
|
| 652 |
+
n_components=None,
|
| 653 |
+
random_state=None,
|
| 654 |
+
n_init=10,
|
| 655 |
+
gamma=1.0,
|
| 656 |
+
affinity="rbf",
|
| 657 |
+
n_neighbors=10,
|
| 658 |
+
eigen_tol="auto",
|
| 659 |
+
assign_labels="kmeans",
|
| 660 |
+
degree=3,
|
| 661 |
+
coef0=1,
|
| 662 |
+
kernel_params=None,
|
| 663 |
+
n_jobs=None,
|
| 664 |
+
verbose=False,
|
| 665 |
+
):
|
| 666 |
+
self.n_clusters = n_clusters
|
| 667 |
+
self.eigen_solver = eigen_solver
|
| 668 |
+
self.n_components = n_components
|
| 669 |
+
self.random_state = random_state
|
| 670 |
+
self.n_init = n_init
|
| 671 |
+
self.gamma = gamma
|
| 672 |
+
self.affinity = affinity
|
| 673 |
+
self.n_neighbors = n_neighbors
|
| 674 |
+
self.eigen_tol = eigen_tol
|
| 675 |
+
self.assign_labels = assign_labels
|
| 676 |
+
self.degree = degree
|
| 677 |
+
self.coef0 = coef0
|
| 678 |
+
self.kernel_params = kernel_params
|
| 679 |
+
self.n_jobs = n_jobs
|
| 680 |
+
self.verbose = verbose
|
| 681 |
+
|
| 682 |
+
def fit(self, X, y=None):
|
| 683 |
+
"""Perform spectral clustering from features, or affinity matrix.
|
| 684 |
+
|
| 685 |
+
Parameters
|
| 686 |
+
----------
|
| 687 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features) or \
|
| 688 |
+
(n_samples, n_samples)
|
| 689 |
+
Training instances to cluster, similarities / affinities between
|
| 690 |
+
instances if ``affinity='precomputed'``, or distances between
|
| 691 |
+
instances if ``affinity='precomputed_nearest_neighbors``. If a
|
| 692 |
+
sparse matrix is provided in a format other than ``csr_matrix``,
|
| 693 |
+
``csc_matrix``, or ``coo_matrix``, it will be converted into a
|
| 694 |
+
sparse ``csr_matrix``.
|
| 695 |
+
|
| 696 |
+
y : Ignored
|
| 697 |
+
Not used, present here for API consistency by convention.
|
| 698 |
+
|
| 699 |
+
Returns
|
| 700 |
+
-------
|
| 701 |
+
self : object
|
| 702 |
+
A fitted instance of the estimator.
|
| 703 |
+
"""
|
| 704 |
+
self._validate_params()
|
| 705 |
+
|
| 706 |
+
X = self._validate_data(
|
| 707 |
+
X,
|
| 708 |
+
accept_sparse=["csr", "csc", "coo"],
|
| 709 |
+
dtype=np.float64,
|
| 710 |
+
ensure_min_samples=2,
|
| 711 |
+
)
|
| 712 |
+
allow_squared = self.affinity in [
|
| 713 |
+
"precomputed",
|
| 714 |
+
"precomputed_nearest_neighbors",
|
| 715 |
+
]
|
| 716 |
+
if X.shape[0] == X.shape[1] and not allow_squared:
|
| 717 |
+
warnings.warn(
|
| 718 |
+
"The spectral clustering API has changed. ``fit``"
|
| 719 |
+
"now constructs an affinity matrix from data. To use"
|
| 720 |
+
" a custom affinity matrix, "
|
| 721 |
+
"set ``affinity=precomputed``."
|
| 722 |
+
)
|
| 723 |
+
|
| 724 |
+
if self.affinity == "nearest_neighbors":
|
| 725 |
+
connectivity = kneighbors_graph(
|
| 726 |
+
X, n_neighbors=self.n_neighbors, include_self=True, n_jobs=self.n_jobs
|
| 727 |
+
)
|
| 728 |
+
self.affinity_matrix_ = 0.5 * (connectivity + connectivity.T)
|
| 729 |
+
elif self.affinity == "precomputed_nearest_neighbors":
|
| 730 |
+
estimator = NearestNeighbors(
|
| 731 |
+
n_neighbors=self.n_neighbors, n_jobs=self.n_jobs, metric="precomputed"
|
| 732 |
+
).fit(X)
|
| 733 |
+
connectivity = estimator.kneighbors_graph(X=X, mode="connectivity")
|
| 734 |
+
self.affinity_matrix_ = 0.5 * (connectivity + connectivity.T)
|
| 735 |
+
elif self.affinity == "precomputed":
|
| 736 |
+
self.affinity_matrix_ = X
|
| 737 |
+
else:
|
| 738 |
+
params = self.kernel_params
|
| 739 |
+
if params is None:
|
| 740 |
+
params = {}
|
| 741 |
+
if not callable(self.affinity):
|
| 742 |
+
params["gamma"] = self.gamma
|
| 743 |
+
params["degree"] = self.degree
|
| 744 |
+
params["coef0"] = self.coef0
|
| 745 |
+
self.affinity_matrix_ = pairwise_kernels(
|
| 746 |
+
X, metric=self.affinity, filter_params=True, **params
|
| 747 |
+
)
|
| 748 |
+
|
| 749 |
+
random_state = check_random_state(self.random_state)
|
| 750 |
+
self.labels_ = spectral_clustering(
|
| 751 |
+
self.affinity_matrix_,
|
| 752 |
+
n_clusters=self.n_clusters,
|
| 753 |
+
n_components=self.n_components,
|
| 754 |
+
eigen_solver=self.eigen_solver,
|
| 755 |
+
random_state=random_state,
|
| 756 |
+
n_init=self.n_init,
|
| 757 |
+
eigen_tol=self.eigen_tol,
|
| 758 |
+
assign_labels=self.assign_labels,
|
| 759 |
+
verbose=self.verbose,
|
| 760 |
+
)
|
| 761 |
+
return self
|
| 762 |
+
|
| 763 |
+
def fit_predict(self, X, y=None):
|
| 764 |
+
"""Perform spectral clustering on `X` and return cluster labels.
|
| 765 |
+
|
| 766 |
+
Parameters
|
| 767 |
+
----------
|
| 768 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features) or \
|
| 769 |
+
(n_samples, n_samples)
|
| 770 |
+
Training instances to cluster, similarities / affinities between
|
| 771 |
+
instances if ``affinity='precomputed'``, or distances between
|
| 772 |
+
instances if ``affinity='precomputed_nearest_neighbors``. If a
|
| 773 |
+
sparse matrix is provided in a format other than ``csr_matrix``,
|
| 774 |
+
``csc_matrix``, or ``coo_matrix``, it will be converted into a
|
| 775 |
+
sparse ``csr_matrix``.
|
| 776 |
+
|
| 777 |
+
y : Ignored
|
| 778 |
+
Not used, present here for API consistency by convention.
|
| 779 |
+
|
| 780 |
+
Returns
|
| 781 |
+
-------
|
| 782 |
+
labels : ndarray of shape (n_samples,)
|
| 783 |
+
Cluster labels.
|
| 784 |
+
"""
|
| 785 |
+
return super().fit_predict(X, y)
|
| 786 |
+
|
| 787 |
+
def _more_tags(self):
|
| 788 |
+
return {
|
| 789 |
+
"pairwise": self.affinity
|
| 790 |
+
in ["precomputed", "precomputed_nearest_neighbors"]
|
| 791 |
+
}
|
videollama2/lib/python3.10/site-packages/sklearn/ensemble/__pycache__/_base.cpython-310.pyc
ADDED
|
Binary file (11.1 kB). View file
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/ensemble/__pycache__/_iforest.cpython-310.pyc
ADDED
|
Binary file (16.7 kB). View file
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/feature_selection/__init__.py
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
The :mod:`sklearn.feature_selection` module implements feature selection
|
| 3 |
+
algorithms. It currently includes univariate filter selection methods and the
|
| 4 |
+
recursive feature elimination algorithm.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
from ._univariate_selection import chi2
|
| 8 |
+
from ._univariate_selection import f_classif
|
| 9 |
+
from ._univariate_selection import f_oneway
|
| 10 |
+
from ._univariate_selection import f_regression
|
| 11 |
+
from ._univariate_selection import r_regression
|
| 12 |
+
from ._univariate_selection import SelectPercentile
|
| 13 |
+
from ._univariate_selection import SelectKBest
|
| 14 |
+
from ._univariate_selection import SelectFpr
|
| 15 |
+
from ._univariate_selection import SelectFdr
|
| 16 |
+
from ._univariate_selection import SelectFwe
|
| 17 |
+
from ._univariate_selection import GenericUnivariateSelect
|
| 18 |
+
|
| 19 |
+
from ._variance_threshold import VarianceThreshold
|
| 20 |
+
|
| 21 |
+
from ._rfe import RFE
|
| 22 |
+
from ._rfe import RFECV
|
| 23 |
+
|
| 24 |
+
from ._from_model import SelectFromModel
|
| 25 |
+
|
| 26 |
+
from ._sequential import SequentialFeatureSelector
|
| 27 |
+
|
| 28 |
+
from ._mutual_info import mutual_info_regression, mutual_info_classif
|
| 29 |
+
|
| 30 |
+
from ._base import SelectorMixin
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
__all__ = [
|
| 34 |
+
"GenericUnivariateSelect",
|
| 35 |
+
"SequentialFeatureSelector",
|
| 36 |
+
"RFE",
|
| 37 |
+
"RFECV",
|
| 38 |
+
"SelectFdr",
|
| 39 |
+
"SelectFpr",
|
| 40 |
+
"SelectFwe",
|
| 41 |
+
"SelectKBest",
|
| 42 |
+
"SelectFromModel",
|
| 43 |
+
"SelectPercentile",
|
| 44 |
+
"VarianceThreshold",
|
| 45 |
+
"chi2",
|
| 46 |
+
"f_classif",
|
| 47 |
+
"f_oneway",
|
| 48 |
+
"f_regression",
|
| 49 |
+
"r_regression",
|
| 50 |
+
"mutual_info_classif",
|
| 51 |
+
"mutual_info_regression",
|
| 52 |
+
"SelectorMixin",
|
| 53 |
+
]
|
videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_from_model.py
ADDED
|
@@ -0,0 +1,444 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Authors: Gilles Louppe, Mathieu Blondel, Maheshakya Wijewardena
|
| 2 |
+
# License: BSD 3 clause
|
| 3 |
+
|
| 4 |
+
from copy import deepcopy
|
| 5 |
+
|
| 6 |
+
import numpy as np
|
| 7 |
+
from numbers import Integral, Real
|
| 8 |
+
|
| 9 |
+
from ._base import SelectorMixin
|
| 10 |
+
from ._base import _get_feature_importances
|
| 11 |
+
from ..base import BaseEstimator, clone, MetaEstimatorMixin
|
| 12 |
+
from ..utils._tags import _safe_tags
|
| 13 |
+
from ..utils.validation import check_is_fitted, check_scalar, _num_features
|
| 14 |
+
from ..utils._param_validation import HasMethods, Interval, Options
|
| 15 |
+
|
| 16 |
+
from ..exceptions import NotFittedError
|
| 17 |
+
from ..utils.metaestimators import available_if
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def _calculate_threshold(estimator, importances, threshold):
|
| 21 |
+
"""Interpret the threshold value"""
|
| 22 |
+
|
| 23 |
+
if threshold is None:
|
| 24 |
+
# determine default from estimator
|
| 25 |
+
est_name = estimator.__class__.__name__
|
| 26 |
+
is_l1_penalized = hasattr(estimator, "penalty") and estimator.penalty == "l1"
|
| 27 |
+
is_lasso = "Lasso" in est_name
|
| 28 |
+
is_elasticnet_l1_penalized = "ElasticNet" in est_name and (
|
| 29 |
+
(hasattr(estimator, "l1_ratio_") and np.isclose(estimator.l1_ratio_, 1.0))
|
| 30 |
+
or (hasattr(estimator, "l1_ratio") and np.isclose(estimator.l1_ratio, 1.0))
|
| 31 |
+
)
|
| 32 |
+
if is_l1_penalized or is_lasso or is_elasticnet_l1_penalized:
|
| 33 |
+
# the natural default threshold is 0 when l1 penalty was used
|
| 34 |
+
threshold = 1e-5
|
| 35 |
+
else:
|
| 36 |
+
threshold = "mean"
|
| 37 |
+
|
| 38 |
+
if isinstance(threshold, str):
|
| 39 |
+
if "*" in threshold:
|
| 40 |
+
scale, reference = threshold.split("*")
|
| 41 |
+
scale = float(scale.strip())
|
| 42 |
+
reference = reference.strip()
|
| 43 |
+
|
| 44 |
+
if reference == "median":
|
| 45 |
+
reference = np.median(importances)
|
| 46 |
+
elif reference == "mean":
|
| 47 |
+
reference = np.mean(importances)
|
| 48 |
+
else:
|
| 49 |
+
raise ValueError("Unknown reference: " + reference)
|
| 50 |
+
|
| 51 |
+
threshold = scale * reference
|
| 52 |
+
|
| 53 |
+
elif threshold == "median":
|
| 54 |
+
threshold = np.median(importances)
|
| 55 |
+
|
| 56 |
+
elif threshold == "mean":
|
| 57 |
+
threshold = np.mean(importances)
|
| 58 |
+
|
| 59 |
+
else:
|
| 60 |
+
raise ValueError(
|
| 61 |
+
"Expected threshold='mean' or threshold='median' got %s" % threshold
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
else:
|
| 65 |
+
threshold = float(threshold)
|
| 66 |
+
|
| 67 |
+
return threshold
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def _estimator_has(attr):
|
| 71 |
+
"""Check if we can delegate a method to the underlying estimator.
|
| 72 |
+
|
| 73 |
+
First, we check the fitted estimator if available, otherwise we
|
| 74 |
+
check the unfitted estimator.
|
| 75 |
+
"""
|
| 76 |
+
return lambda self: (
|
| 77 |
+
hasattr(self.estimator_, attr)
|
| 78 |
+
if hasattr(self, "estimator_")
|
| 79 |
+
else hasattr(self.estimator, attr)
|
| 80 |
+
)
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
class SelectFromModel(MetaEstimatorMixin, SelectorMixin, BaseEstimator):
|
| 84 |
+
"""Meta-transformer for selecting features based on importance weights.
|
| 85 |
+
|
| 86 |
+
.. versionadded:: 0.17
|
| 87 |
+
|
| 88 |
+
Read more in the :ref:`User Guide <select_from_model>`.
|
| 89 |
+
|
| 90 |
+
Parameters
|
| 91 |
+
----------
|
| 92 |
+
estimator : object
|
| 93 |
+
The base estimator from which the transformer is built.
|
| 94 |
+
This can be both a fitted (if ``prefit`` is set to True)
|
| 95 |
+
or a non-fitted estimator. The estimator should have a
|
| 96 |
+
``feature_importances_`` or ``coef_`` attribute after fitting.
|
| 97 |
+
Otherwise, the ``importance_getter`` parameter should be used.
|
| 98 |
+
|
| 99 |
+
threshold : str or float, default=None
|
| 100 |
+
The threshold value to use for feature selection. Features whose
|
| 101 |
+
absolute importance value is greater or equal are kept while the others
|
| 102 |
+
are discarded. If "median" (resp. "mean"), then the ``threshold`` value
|
| 103 |
+
is the median (resp. the mean) of the feature importances. A scaling
|
| 104 |
+
factor (e.g., "1.25*mean") may also be used. If None and if the
|
| 105 |
+
estimator has a parameter penalty set to l1, either explicitly
|
| 106 |
+
or implicitly (e.g, Lasso), the threshold used is 1e-5.
|
| 107 |
+
Otherwise, "mean" is used by default.
|
| 108 |
+
|
| 109 |
+
prefit : bool, default=False
|
| 110 |
+
Whether a prefit model is expected to be passed into the constructor
|
| 111 |
+
directly or not.
|
| 112 |
+
If `True`, `estimator` must be a fitted estimator.
|
| 113 |
+
If `False`, `estimator` is fitted and updated by calling
|
| 114 |
+
`fit` and `partial_fit`, respectively.
|
| 115 |
+
|
| 116 |
+
norm_order : non-zero int, inf, -inf, default=1
|
| 117 |
+
Order of the norm used to filter the vectors of coefficients below
|
| 118 |
+
``threshold`` in the case where the ``coef_`` attribute of the
|
| 119 |
+
estimator is of dimension 2.
|
| 120 |
+
|
| 121 |
+
max_features : int, callable, default=None
|
| 122 |
+
The maximum number of features to select.
|
| 123 |
+
|
| 124 |
+
- If an integer, then it specifies the maximum number of features to
|
| 125 |
+
allow.
|
| 126 |
+
- If a callable, then it specifies how to calculate the maximum number of
|
| 127 |
+
features allowed by using the output of `max_features(X)`.
|
| 128 |
+
- If `None`, then all features are kept.
|
| 129 |
+
|
| 130 |
+
To only select based on ``max_features``, set ``threshold=-np.inf``.
|
| 131 |
+
|
| 132 |
+
.. versionadded:: 0.20
|
| 133 |
+
.. versionchanged:: 1.1
|
| 134 |
+
`max_features` accepts a callable.
|
| 135 |
+
|
| 136 |
+
importance_getter : str or callable, default='auto'
|
| 137 |
+
If 'auto', uses the feature importance either through a ``coef_``
|
| 138 |
+
attribute or ``feature_importances_`` attribute of estimator.
|
| 139 |
+
|
| 140 |
+
Also accepts a string that specifies an attribute name/path
|
| 141 |
+
for extracting feature importance (implemented with `attrgetter`).
|
| 142 |
+
For example, give `regressor_.coef_` in case of
|
| 143 |
+
:class:`~sklearn.compose.TransformedTargetRegressor` or
|
| 144 |
+
`named_steps.clf.feature_importances_` in case of
|
| 145 |
+
:class:`~sklearn.pipeline.Pipeline` with its last step named `clf`.
|
| 146 |
+
|
| 147 |
+
If `callable`, overrides the default feature importance getter.
|
| 148 |
+
The callable is passed with the fitted estimator and it should
|
| 149 |
+
return importance for each feature.
|
| 150 |
+
|
| 151 |
+
.. versionadded:: 0.24
|
| 152 |
+
|
| 153 |
+
Attributes
|
| 154 |
+
----------
|
| 155 |
+
estimator_ : estimator
|
| 156 |
+
The base estimator from which the transformer is built. This attribute
|
| 157 |
+
exist only when `fit` has been called.
|
| 158 |
+
|
| 159 |
+
- If `prefit=True`, it is a deep copy of `estimator`.
|
| 160 |
+
- If `prefit=False`, it is a clone of `estimator` and fit on the data
|
| 161 |
+
passed to `fit` or `partial_fit`.
|
| 162 |
+
|
| 163 |
+
n_features_in_ : int
|
| 164 |
+
Number of features seen during :term:`fit`. Only defined if the
|
| 165 |
+
underlying estimator exposes such an attribute when fit.
|
| 166 |
+
|
| 167 |
+
.. versionadded:: 0.24
|
| 168 |
+
|
| 169 |
+
max_features_ : int
|
| 170 |
+
Maximum number of features calculated during :term:`fit`. Only defined
|
| 171 |
+
if the ``max_features`` is not `None`.
|
| 172 |
+
|
| 173 |
+
- If `max_features` is an `int`, then `max_features_ = max_features`.
|
| 174 |
+
- If `max_features` is a callable, then `max_features_ = max_features(X)`.
|
| 175 |
+
|
| 176 |
+
.. versionadded:: 1.1
|
| 177 |
+
|
| 178 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 179 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 180 |
+
has feature names that are all strings.
|
| 181 |
+
|
| 182 |
+
.. versionadded:: 1.0
|
| 183 |
+
|
| 184 |
+
threshold_ : float
|
| 185 |
+
The threshold value used for feature selection.
|
| 186 |
+
|
| 187 |
+
See Also
|
| 188 |
+
--------
|
| 189 |
+
RFE : Recursive feature elimination based on importance weights.
|
| 190 |
+
RFECV : Recursive feature elimination with built-in cross-validated
|
| 191 |
+
selection of the best number of features.
|
| 192 |
+
SequentialFeatureSelector : Sequential cross-validation based feature
|
| 193 |
+
selection. Does not rely on importance weights.
|
| 194 |
+
|
| 195 |
+
Notes
|
| 196 |
+
-----
|
| 197 |
+
Allows NaN/Inf in the input if the underlying estimator does as well.
|
| 198 |
+
|
| 199 |
+
Examples
|
| 200 |
+
--------
|
| 201 |
+
>>> from sklearn.feature_selection import SelectFromModel
|
| 202 |
+
>>> from sklearn.linear_model import LogisticRegression
|
| 203 |
+
>>> X = [[ 0.87, -1.34, 0.31 ],
|
| 204 |
+
... [-2.79, -0.02, -0.85 ],
|
| 205 |
+
... [-1.34, -0.48, -2.55 ],
|
| 206 |
+
... [ 1.92, 1.48, 0.65 ]]
|
| 207 |
+
>>> y = [0, 1, 0, 1]
|
| 208 |
+
>>> selector = SelectFromModel(estimator=LogisticRegression()).fit(X, y)
|
| 209 |
+
>>> selector.estimator_.coef_
|
| 210 |
+
array([[-0.3252302 , 0.83462377, 0.49750423]])
|
| 211 |
+
>>> selector.threshold_
|
| 212 |
+
0.55245...
|
| 213 |
+
>>> selector.get_support()
|
| 214 |
+
array([False, True, False])
|
| 215 |
+
>>> selector.transform(X)
|
| 216 |
+
array([[-1.34],
|
| 217 |
+
[-0.02],
|
| 218 |
+
[-0.48],
|
| 219 |
+
[ 1.48]])
|
| 220 |
+
|
| 221 |
+
Using a callable to create a selector that can use no more than half
|
| 222 |
+
of the input features.
|
| 223 |
+
|
| 224 |
+
>>> def half_callable(X):
|
| 225 |
+
... return round(len(X[0]) / 2)
|
| 226 |
+
>>> half_selector = SelectFromModel(estimator=LogisticRegression(),
|
| 227 |
+
... max_features=half_callable)
|
| 228 |
+
>>> _ = half_selector.fit(X, y)
|
| 229 |
+
>>> half_selector.max_features_
|
| 230 |
+
2
|
| 231 |
+
"""
|
| 232 |
+
|
| 233 |
+
_parameter_constraints: dict = {
|
| 234 |
+
"estimator": [HasMethods("fit")],
|
| 235 |
+
"threshold": [Interval(Real, None, None, closed="both"), str, None],
|
| 236 |
+
"prefit": ["boolean"],
|
| 237 |
+
"norm_order": [
|
| 238 |
+
Interval(Integral, None, -1, closed="right"),
|
| 239 |
+
Interval(Integral, 1, None, closed="left"),
|
| 240 |
+
Options(Real, {np.inf, -np.inf}),
|
| 241 |
+
],
|
| 242 |
+
"max_features": [Interval(Integral, 0, None, closed="left"), callable, None],
|
| 243 |
+
"importance_getter": [str, callable],
|
| 244 |
+
}
|
| 245 |
+
|
| 246 |
+
def __init__(
|
| 247 |
+
self,
|
| 248 |
+
estimator,
|
| 249 |
+
*,
|
| 250 |
+
threshold=None,
|
| 251 |
+
prefit=False,
|
| 252 |
+
norm_order=1,
|
| 253 |
+
max_features=None,
|
| 254 |
+
importance_getter="auto",
|
| 255 |
+
):
|
| 256 |
+
self.estimator = estimator
|
| 257 |
+
self.threshold = threshold
|
| 258 |
+
self.prefit = prefit
|
| 259 |
+
self.importance_getter = importance_getter
|
| 260 |
+
self.norm_order = norm_order
|
| 261 |
+
self.max_features = max_features
|
| 262 |
+
|
| 263 |
+
def _get_support_mask(self):
|
| 264 |
+
estimator = getattr(self, "estimator_", self.estimator)
|
| 265 |
+
max_features = getattr(self, "max_features_", self.max_features)
|
| 266 |
+
|
| 267 |
+
if self.prefit:
|
| 268 |
+
try:
|
| 269 |
+
check_is_fitted(self.estimator)
|
| 270 |
+
except NotFittedError as exc:
|
| 271 |
+
raise NotFittedError(
|
| 272 |
+
"When `prefit=True`, `estimator` is expected to be a fitted "
|
| 273 |
+
"estimator."
|
| 274 |
+
) from exc
|
| 275 |
+
if callable(max_features):
|
| 276 |
+
# This branch is executed when `transform` is called directly and thus
|
| 277 |
+
# `max_features_` is not set and we fallback using `self.max_features`
|
| 278 |
+
# that is not validated
|
| 279 |
+
raise NotFittedError(
|
| 280 |
+
"When `prefit=True` and `max_features` is a callable, call `fit` "
|
| 281 |
+
"before calling `transform`."
|
| 282 |
+
)
|
| 283 |
+
elif max_features is not None and not isinstance(max_features, Integral):
|
| 284 |
+
raise ValueError(
|
| 285 |
+
f"`max_features` must be an integer. Got `max_features={max_features}` "
|
| 286 |
+
"instead."
|
| 287 |
+
)
|
| 288 |
+
|
| 289 |
+
scores = _get_feature_importances(
|
| 290 |
+
estimator=estimator,
|
| 291 |
+
getter=self.importance_getter,
|
| 292 |
+
transform_func="norm",
|
| 293 |
+
norm_order=self.norm_order,
|
| 294 |
+
)
|
| 295 |
+
threshold = _calculate_threshold(estimator, scores, self.threshold)
|
| 296 |
+
if self.max_features is not None:
|
| 297 |
+
mask = np.zeros_like(scores, dtype=bool)
|
| 298 |
+
candidate_indices = np.argsort(-scores, kind="mergesort")[:max_features]
|
| 299 |
+
mask[candidate_indices] = True
|
| 300 |
+
else:
|
| 301 |
+
mask = np.ones_like(scores, dtype=bool)
|
| 302 |
+
mask[scores < threshold] = False
|
| 303 |
+
return mask
|
| 304 |
+
|
| 305 |
+
def _check_max_features(self, X):
|
| 306 |
+
if self.max_features is not None:
|
| 307 |
+
n_features = _num_features(X)
|
| 308 |
+
|
| 309 |
+
if callable(self.max_features):
|
| 310 |
+
max_features = self.max_features(X)
|
| 311 |
+
else: # int
|
| 312 |
+
max_features = self.max_features
|
| 313 |
+
|
| 314 |
+
check_scalar(
|
| 315 |
+
max_features,
|
| 316 |
+
"max_features",
|
| 317 |
+
Integral,
|
| 318 |
+
min_val=0,
|
| 319 |
+
max_val=n_features,
|
| 320 |
+
)
|
| 321 |
+
self.max_features_ = max_features
|
| 322 |
+
|
| 323 |
+
def fit(self, X, y=None, **fit_params):
|
| 324 |
+
"""Fit the SelectFromModel meta-transformer.
|
| 325 |
+
|
| 326 |
+
Parameters
|
| 327 |
+
----------
|
| 328 |
+
X : array-like of shape (n_samples, n_features)
|
| 329 |
+
The training input samples.
|
| 330 |
+
|
| 331 |
+
y : array-like of shape (n_samples,), default=None
|
| 332 |
+
The target values (integers that correspond to classes in
|
| 333 |
+
classification, real numbers in regression).
|
| 334 |
+
|
| 335 |
+
**fit_params : dict
|
| 336 |
+
Other estimator specific parameters.
|
| 337 |
+
|
| 338 |
+
Returns
|
| 339 |
+
-------
|
| 340 |
+
self : object
|
| 341 |
+
Fitted estimator.
|
| 342 |
+
"""
|
| 343 |
+
self._validate_params()
|
| 344 |
+
self._check_max_features(X)
|
| 345 |
+
|
| 346 |
+
if self.prefit:
|
| 347 |
+
try:
|
| 348 |
+
check_is_fitted(self.estimator)
|
| 349 |
+
except NotFittedError as exc:
|
| 350 |
+
raise NotFittedError(
|
| 351 |
+
"When `prefit=True`, `estimator` is expected to be a fitted "
|
| 352 |
+
"estimator."
|
| 353 |
+
) from exc
|
| 354 |
+
self.estimator_ = deepcopy(self.estimator)
|
| 355 |
+
else:
|
| 356 |
+
self.estimator_ = clone(self.estimator)
|
| 357 |
+
self.estimator_.fit(X, y, **fit_params)
|
| 358 |
+
|
| 359 |
+
if hasattr(self.estimator_, "feature_names_in_"):
|
| 360 |
+
self.feature_names_in_ = self.estimator_.feature_names_in_
|
| 361 |
+
else:
|
| 362 |
+
self._check_feature_names(X, reset=True)
|
| 363 |
+
|
| 364 |
+
return self
|
| 365 |
+
|
| 366 |
+
@property
|
| 367 |
+
def threshold_(self):
|
| 368 |
+
"""Threshold value used for feature selection."""
|
| 369 |
+
scores = _get_feature_importances(
|
| 370 |
+
estimator=self.estimator_,
|
| 371 |
+
getter=self.importance_getter,
|
| 372 |
+
transform_func="norm",
|
| 373 |
+
norm_order=self.norm_order,
|
| 374 |
+
)
|
| 375 |
+
return _calculate_threshold(self.estimator, scores, self.threshold)
|
| 376 |
+
|
| 377 |
+
@available_if(_estimator_has("partial_fit"))
|
| 378 |
+
def partial_fit(self, X, y=None, **fit_params):
|
| 379 |
+
"""Fit the SelectFromModel meta-transformer only once.
|
| 380 |
+
|
| 381 |
+
Parameters
|
| 382 |
+
----------
|
| 383 |
+
X : array-like of shape (n_samples, n_features)
|
| 384 |
+
The training input samples.
|
| 385 |
+
|
| 386 |
+
y : array-like of shape (n_samples,), default=None
|
| 387 |
+
The target values (integers that correspond to classes in
|
| 388 |
+
classification, real numbers in regression).
|
| 389 |
+
|
| 390 |
+
**fit_params : dict
|
| 391 |
+
Other estimator specific parameters.
|
| 392 |
+
|
| 393 |
+
Returns
|
| 394 |
+
-------
|
| 395 |
+
self : object
|
| 396 |
+
Fitted estimator.
|
| 397 |
+
"""
|
| 398 |
+
first_call = not hasattr(self, "estimator_")
|
| 399 |
+
|
| 400 |
+
if first_call:
|
| 401 |
+
self._validate_params()
|
| 402 |
+
self._check_max_features(X)
|
| 403 |
+
|
| 404 |
+
if self.prefit:
|
| 405 |
+
if first_call:
|
| 406 |
+
try:
|
| 407 |
+
check_is_fitted(self.estimator)
|
| 408 |
+
except NotFittedError as exc:
|
| 409 |
+
raise NotFittedError(
|
| 410 |
+
"When `prefit=True`, `estimator` is expected to be a fitted "
|
| 411 |
+
"estimator."
|
| 412 |
+
) from exc
|
| 413 |
+
self.estimator_ = deepcopy(self.estimator)
|
| 414 |
+
return self
|
| 415 |
+
|
| 416 |
+
if first_call:
|
| 417 |
+
self.estimator_ = clone(self.estimator)
|
| 418 |
+
self.estimator_.partial_fit(X, y, **fit_params)
|
| 419 |
+
|
| 420 |
+
if hasattr(self.estimator_, "feature_names_in_"):
|
| 421 |
+
self.feature_names_in_ = self.estimator_.feature_names_in_
|
| 422 |
+
else:
|
| 423 |
+
self._check_feature_names(X, reset=first_call)
|
| 424 |
+
|
| 425 |
+
return self
|
| 426 |
+
|
| 427 |
+
@property
|
| 428 |
+
def n_features_in_(self):
|
| 429 |
+
"""Number of features seen during `fit`."""
|
| 430 |
+
# For consistency with other estimators we raise a AttributeError so
|
| 431 |
+
# that hasattr() fails if the estimator isn't fitted.
|
| 432 |
+
try:
|
| 433 |
+
check_is_fitted(self)
|
| 434 |
+
except NotFittedError as nfe:
|
| 435 |
+
raise AttributeError(
|
| 436 |
+
"{} object has no n_features_in_ attribute.".format(
|
| 437 |
+
self.__class__.__name__
|
| 438 |
+
)
|
| 439 |
+
) from nfe
|
| 440 |
+
|
| 441 |
+
return self.estimator_.n_features_in_
|
| 442 |
+
|
| 443 |
+
def _more_tags(self):
|
| 444 |
+
return {"allow_nan": _safe_tags(self.estimator, key="allow_nan")}
|
videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_mutual_info.py
ADDED
|
@@ -0,0 +1,468 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Author: Nikolay Mayorov <n59_ru@hotmail.com>
|
| 2 |
+
# License: 3-clause BSD
|
| 3 |
+
|
| 4 |
+
import numpy as np
|
| 5 |
+
from scipy.sparse import issparse
|
| 6 |
+
from scipy.special import digamma
|
| 7 |
+
|
| 8 |
+
from ..metrics.cluster import mutual_info_score
|
| 9 |
+
from ..neighbors import NearestNeighbors, KDTree
|
| 10 |
+
from ..preprocessing import scale
|
| 11 |
+
from ..utils import check_random_state
|
| 12 |
+
from ..utils.validation import check_array, check_X_y
|
| 13 |
+
from ..utils.multiclass import check_classification_targets
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def _compute_mi_cc(x, y, n_neighbors):
|
| 17 |
+
"""Compute mutual information between two continuous variables.
|
| 18 |
+
|
| 19 |
+
Parameters
|
| 20 |
+
----------
|
| 21 |
+
x, y : ndarray, shape (n_samples,)
|
| 22 |
+
Samples of two continuous random variables, must have an identical
|
| 23 |
+
shape.
|
| 24 |
+
|
| 25 |
+
n_neighbors : int
|
| 26 |
+
Number of nearest neighbors to search for each point, see [1]_.
|
| 27 |
+
|
| 28 |
+
Returns
|
| 29 |
+
-------
|
| 30 |
+
mi : float
|
| 31 |
+
Estimated mutual information. If it turned out to be negative it is
|
| 32 |
+
replace by 0.
|
| 33 |
+
|
| 34 |
+
Notes
|
| 35 |
+
-----
|
| 36 |
+
True mutual information can't be negative. If its estimate by a numerical
|
| 37 |
+
method is negative, it means (providing the method is adequate) that the
|
| 38 |
+
mutual information is close to 0 and replacing it by 0 is a reasonable
|
| 39 |
+
strategy.
|
| 40 |
+
|
| 41 |
+
References
|
| 42 |
+
----------
|
| 43 |
+
.. [1] A. Kraskov, H. Stogbauer and P. Grassberger, "Estimating mutual
|
| 44 |
+
information". Phys. Rev. E 69, 2004.
|
| 45 |
+
"""
|
| 46 |
+
n_samples = x.size
|
| 47 |
+
|
| 48 |
+
x = x.reshape((-1, 1))
|
| 49 |
+
y = y.reshape((-1, 1))
|
| 50 |
+
xy = np.hstack((x, y))
|
| 51 |
+
|
| 52 |
+
# Here we rely on NearestNeighbors to select the fastest algorithm.
|
| 53 |
+
nn = NearestNeighbors(metric="chebyshev", n_neighbors=n_neighbors)
|
| 54 |
+
|
| 55 |
+
nn.fit(xy)
|
| 56 |
+
radius = nn.kneighbors()[0]
|
| 57 |
+
radius = np.nextafter(radius[:, -1], 0)
|
| 58 |
+
|
| 59 |
+
# KDTree is explicitly fit to allow for the querying of number of
|
| 60 |
+
# neighbors within a specified radius
|
| 61 |
+
kd = KDTree(x, metric="chebyshev")
|
| 62 |
+
nx = kd.query_radius(x, radius, count_only=True, return_distance=False)
|
| 63 |
+
nx = np.array(nx) - 1.0
|
| 64 |
+
|
| 65 |
+
kd = KDTree(y, metric="chebyshev")
|
| 66 |
+
ny = kd.query_radius(y, radius, count_only=True, return_distance=False)
|
| 67 |
+
ny = np.array(ny) - 1.0
|
| 68 |
+
|
| 69 |
+
mi = (
|
| 70 |
+
digamma(n_samples)
|
| 71 |
+
+ digamma(n_neighbors)
|
| 72 |
+
- np.mean(digamma(nx + 1))
|
| 73 |
+
- np.mean(digamma(ny + 1))
|
| 74 |
+
)
|
| 75 |
+
|
| 76 |
+
return max(0, mi)
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def _compute_mi_cd(c, d, n_neighbors):
|
| 80 |
+
"""Compute mutual information between continuous and discrete variables.
|
| 81 |
+
|
| 82 |
+
Parameters
|
| 83 |
+
----------
|
| 84 |
+
c : ndarray, shape (n_samples,)
|
| 85 |
+
Samples of a continuous random variable.
|
| 86 |
+
|
| 87 |
+
d : ndarray, shape (n_samples,)
|
| 88 |
+
Samples of a discrete random variable.
|
| 89 |
+
|
| 90 |
+
n_neighbors : int
|
| 91 |
+
Number of nearest neighbors to search for each point, see [1]_.
|
| 92 |
+
|
| 93 |
+
Returns
|
| 94 |
+
-------
|
| 95 |
+
mi : float
|
| 96 |
+
Estimated mutual information. If it turned out to be negative it is
|
| 97 |
+
replace by 0.
|
| 98 |
+
|
| 99 |
+
Notes
|
| 100 |
+
-----
|
| 101 |
+
True mutual information can't be negative. If its estimate by a numerical
|
| 102 |
+
method is negative, it means (providing the method is adequate) that the
|
| 103 |
+
mutual information is close to 0 and replacing it by 0 is a reasonable
|
| 104 |
+
strategy.
|
| 105 |
+
|
| 106 |
+
References
|
| 107 |
+
----------
|
| 108 |
+
.. [1] B. C. Ross "Mutual Information between Discrete and Continuous
|
| 109 |
+
Data Sets". PLoS ONE 9(2), 2014.
|
| 110 |
+
"""
|
| 111 |
+
n_samples = c.shape[0]
|
| 112 |
+
c = c.reshape((-1, 1))
|
| 113 |
+
|
| 114 |
+
radius = np.empty(n_samples)
|
| 115 |
+
label_counts = np.empty(n_samples)
|
| 116 |
+
k_all = np.empty(n_samples)
|
| 117 |
+
nn = NearestNeighbors()
|
| 118 |
+
for label in np.unique(d):
|
| 119 |
+
mask = d == label
|
| 120 |
+
count = np.sum(mask)
|
| 121 |
+
if count > 1:
|
| 122 |
+
k = min(n_neighbors, count - 1)
|
| 123 |
+
nn.set_params(n_neighbors=k)
|
| 124 |
+
nn.fit(c[mask])
|
| 125 |
+
r = nn.kneighbors()[0]
|
| 126 |
+
radius[mask] = np.nextafter(r[:, -1], 0)
|
| 127 |
+
k_all[mask] = k
|
| 128 |
+
label_counts[mask] = count
|
| 129 |
+
|
| 130 |
+
# Ignore points with unique labels.
|
| 131 |
+
mask = label_counts > 1
|
| 132 |
+
n_samples = np.sum(mask)
|
| 133 |
+
label_counts = label_counts[mask]
|
| 134 |
+
k_all = k_all[mask]
|
| 135 |
+
c = c[mask]
|
| 136 |
+
radius = radius[mask]
|
| 137 |
+
|
| 138 |
+
kd = KDTree(c)
|
| 139 |
+
m_all = kd.query_radius(c, radius, count_only=True, return_distance=False)
|
| 140 |
+
m_all = np.array(m_all)
|
| 141 |
+
|
| 142 |
+
mi = (
|
| 143 |
+
digamma(n_samples)
|
| 144 |
+
+ np.mean(digamma(k_all))
|
| 145 |
+
- np.mean(digamma(label_counts))
|
| 146 |
+
- np.mean(digamma(m_all))
|
| 147 |
+
)
|
| 148 |
+
|
| 149 |
+
return max(0, mi)
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
def _compute_mi(x, y, x_discrete, y_discrete, n_neighbors=3):
|
| 153 |
+
"""Compute mutual information between two variables.
|
| 154 |
+
|
| 155 |
+
This is a simple wrapper which selects a proper function to call based on
|
| 156 |
+
whether `x` and `y` are discrete or not.
|
| 157 |
+
"""
|
| 158 |
+
if x_discrete and y_discrete:
|
| 159 |
+
return mutual_info_score(x, y)
|
| 160 |
+
elif x_discrete and not y_discrete:
|
| 161 |
+
return _compute_mi_cd(y, x, n_neighbors)
|
| 162 |
+
elif not x_discrete and y_discrete:
|
| 163 |
+
return _compute_mi_cd(x, y, n_neighbors)
|
| 164 |
+
else:
|
| 165 |
+
return _compute_mi_cc(x, y, n_neighbors)
|
| 166 |
+
|
| 167 |
+
|
| 168 |
+
def _iterate_columns(X, columns=None):
|
| 169 |
+
"""Iterate over columns of a matrix.
|
| 170 |
+
|
| 171 |
+
Parameters
|
| 172 |
+
----------
|
| 173 |
+
X : ndarray or csc_matrix, shape (n_samples, n_features)
|
| 174 |
+
Matrix over which to iterate.
|
| 175 |
+
|
| 176 |
+
columns : iterable or None, default=None
|
| 177 |
+
Indices of columns to iterate over. If None, iterate over all columns.
|
| 178 |
+
|
| 179 |
+
Yields
|
| 180 |
+
------
|
| 181 |
+
x : ndarray, shape (n_samples,)
|
| 182 |
+
Columns of `X` in dense format.
|
| 183 |
+
"""
|
| 184 |
+
if columns is None:
|
| 185 |
+
columns = range(X.shape[1])
|
| 186 |
+
|
| 187 |
+
if issparse(X):
|
| 188 |
+
for i in columns:
|
| 189 |
+
x = np.zeros(X.shape[0])
|
| 190 |
+
start_ptr, end_ptr = X.indptr[i], X.indptr[i + 1]
|
| 191 |
+
x[X.indices[start_ptr:end_ptr]] = X.data[start_ptr:end_ptr]
|
| 192 |
+
yield x
|
| 193 |
+
else:
|
| 194 |
+
for i in columns:
|
| 195 |
+
yield X[:, i]
|
| 196 |
+
|
| 197 |
+
|
| 198 |
+
def _estimate_mi(
|
| 199 |
+
X,
|
| 200 |
+
y,
|
| 201 |
+
discrete_features="auto",
|
| 202 |
+
discrete_target=False,
|
| 203 |
+
n_neighbors=3,
|
| 204 |
+
copy=True,
|
| 205 |
+
random_state=None,
|
| 206 |
+
):
|
| 207 |
+
"""Estimate mutual information between the features and the target.
|
| 208 |
+
|
| 209 |
+
Parameters
|
| 210 |
+
----------
|
| 211 |
+
X : array-like or sparse matrix, shape (n_samples, n_features)
|
| 212 |
+
Feature matrix.
|
| 213 |
+
|
| 214 |
+
y : array-like of shape (n_samples,)
|
| 215 |
+
Target vector.
|
| 216 |
+
|
| 217 |
+
discrete_features : {'auto', bool, array-like}, default='auto'
|
| 218 |
+
If bool, then determines whether to consider all features discrete
|
| 219 |
+
or continuous. If array, then it should be either a boolean mask
|
| 220 |
+
with shape (n_features,) or array with indices of discrete features.
|
| 221 |
+
If 'auto', it is assigned to False for dense `X` and to True for
|
| 222 |
+
sparse `X`.
|
| 223 |
+
|
| 224 |
+
discrete_target : bool, default=False
|
| 225 |
+
Whether to consider `y` as a discrete variable.
|
| 226 |
+
|
| 227 |
+
n_neighbors : int, default=3
|
| 228 |
+
Number of neighbors to use for MI estimation for continuous variables,
|
| 229 |
+
see [1]_ and [2]_. Higher values reduce variance of the estimation, but
|
| 230 |
+
could introduce a bias.
|
| 231 |
+
|
| 232 |
+
copy : bool, default=True
|
| 233 |
+
Whether to make a copy of the given data. If set to False, the initial
|
| 234 |
+
data will be overwritten.
|
| 235 |
+
|
| 236 |
+
random_state : int, RandomState instance or None, default=None
|
| 237 |
+
Determines random number generation for adding small noise to
|
| 238 |
+
continuous variables in order to remove repeated values.
|
| 239 |
+
Pass an int for reproducible results across multiple function calls.
|
| 240 |
+
See :term:`Glossary <random_state>`.
|
| 241 |
+
|
| 242 |
+
Returns
|
| 243 |
+
-------
|
| 244 |
+
mi : ndarray, shape (n_features,)
|
| 245 |
+
Estimated mutual information between each feature and the target.
|
| 246 |
+
A negative value will be replaced by 0.
|
| 247 |
+
|
| 248 |
+
References
|
| 249 |
+
----------
|
| 250 |
+
.. [1] A. Kraskov, H. Stogbauer and P. Grassberger, "Estimating mutual
|
| 251 |
+
information". Phys. Rev. E 69, 2004.
|
| 252 |
+
.. [2] B. C. Ross "Mutual Information between Discrete and Continuous
|
| 253 |
+
Data Sets". PLoS ONE 9(2), 2014.
|
| 254 |
+
"""
|
| 255 |
+
X, y = check_X_y(X, y, accept_sparse="csc", y_numeric=not discrete_target)
|
| 256 |
+
n_samples, n_features = X.shape
|
| 257 |
+
|
| 258 |
+
if isinstance(discrete_features, (str, bool)):
|
| 259 |
+
if isinstance(discrete_features, str):
|
| 260 |
+
if discrete_features == "auto":
|
| 261 |
+
discrete_features = issparse(X)
|
| 262 |
+
else:
|
| 263 |
+
raise ValueError("Invalid string value for discrete_features.")
|
| 264 |
+
discrete_mask = np.empty(n_features, dtype=bool)
|
| 265 |
+
discrete_mask.fill(discrete_features)
|
| 266 |
+
else:
|
| 267 |
+
discrete_features = check_array(discrete_features, ensure_2d=False)
|
| 268 |
+
if discrete_features.dtype != "bool":
|
| 269 |
+
discrete_mask = np.zeros(n_features, dtype=bool)
|
| 270 |
+
discrete_mask[discrete_features] = True
|
| 271 |
+
else:
|
| 272 |
+
discrete_mask = discrete_features
|
| 273 |
+
|
| 274 |
+
continuous_mask = ~discrete_mask
|
| 275 |
+
if np.any(continuous_mask) and issparse(X):
|
| 276 |
+
raise ValueError("Sparse matrix `X` can't have continuous features.")
|
| 277 |
+
|
| 278 |
+
rng = check_random_state(random_state)
|
| 279 |
+
if np.any(continuous_mask):
|
| 280 |
+
if copy:
|
| 281 |
+
X = X.copy()
|
| 282 |
+
|
| 283 |
+
X[:, continuous_mask] = scale(
|
| 284 |
+
X[:, continuous_mask], with_mean=False, copy=False
|
| 285 |
+
)
|
| 286 |
+
|
| 287 |
+
# Add small noise to continuous features as advised in Kraskov et. al.
|
| 288 |
+
X = X.astype(np.float64, copy=False)
|
| 289 |
+
means = np.maximum(1, np.mean(np.abs(X[:, continuous_mask]), axis=0))
|
| 290 |
+
X[:, continuous_mask] += (
|
| 291 |
+
1e-10
|
| 292 |
+
* means
|
| 293 |
+
* rng.standard_normal(size=(n_samples, np.sum(continuous_mask)))
|
| 294 |
+
)
|
| 295 |
+
|
| 296 |
+
if not discrete_target:
|
| 297 |
+
y = scale(y, with_mean=False)
|
| 298 |
+
y += (
|
| 299 |
+
1e-10
|
| 300 |
+
* np.maximum(1, np.mean(np.abs(y)))
|
| 301 |
+
* rng.standard_normal(size=n_samples)
|
| 302 |
+
)
|
| 303 |
+
|
| 304 |
+
mi = [
|
| 305 |
+
_compute_mi(x, y, discrete_feature, discrete_target, n_neighbors)
|
| 306 |
+
for x, discrete_feature in zip(_iterate_columns(X), discrete_mask)
|
| 307 |
+
]
|
| 308 |
+
|
| 309 |
+
return np.array(mi)
|
| 310 |
+
|
| 311 |
+
|
| 312 |
+
def mutual_info_regression(
|
| 313 |
+
X, y, *, discrete_features="auto", n_neighbors=3, copy=True, random_state=None
|
| 314 |
+
):
|
| 315 |
+
"""Estimate mutual information for a continuous target variable.
|
| 316 |
+
|
| 317 |
+
Mutual information (MI) [1]_ between two random variables is a non-negative
|
| 318 |
+
value, which measures the dependency between the variables. It is equal
|
| 319 |
+
to zero if and only if two random variables are independent, and higher
|
| 320 |
+
values mean higher dependency.
|
| 321 |
+
|
| 322 |
+
The function relies on nonparametric methods based on entropy estimation
|
| 323 |
+
from k-nearest neighbors distances as described in [2]_ and [3]_. Both
|
| 324 |
+
methods are based on the idea originally proposed in [4]_.
|
| 325 |
+
|
| 326 |
+
It can be used for univariate features selection, read more in the
|
| 327 |
+
:ref:`User Guide <univariate_feature_selection>`.
|
| 328 |
+
|
| 329 |
+
Parameters
|
| 330 |
+
----------
|
| 331 |
+
X : array-like or sparse matrix, shape (n_samples, n_features)
|
| 332 |
+
Feature matrix.
|
| 333 |
+
|
| 334 |
+
y : array-like of shape (n_samples,)
|
| 335 |
+
Target vector.
|
| 336 |
+
|
| 337 |
+
discrete_features : {'auto', bool, array-like}, default='auto'
|
| 338 |
+
If bool, then determines whether to consider all features discrete
|
| 339 |
+
or continuous. If array, then it should be either a boolean mask
|
| 340 |
+
with shape (n_features,) or array with indices of discrete features.
|
| 341 |
+
If 'auto', it is assigned to False for dense `X` and to True for
|
| 342 |
+
sparse `X`.
|
| 343 |
+
|
| 344 |
+
n_neighbors : int, default=3
|
| 345 |
+
Number of neighbors to use for MI estimation for continuous variables,
|
| 346 |
+
see [2]_ and [3]_. Higher values reduce variance of the estimation, but
|
| 347 |
+
could introduce a bias.
|
| 348 |
+
|
| 349 |
+
copy : bool, default=True
|
| 350 |
+
Whether to make a copy of the given data. If set to False, the initial
|
| 351 |
+
data will be overwritten.
|
| 352 |
+
|
| 353 |
+
random_state : int, RandomState instance or None, default=None
|
| 354 |
+
Determines random number generation for adding small noise to
|
| 355 |
+
continuous variables in order to remove repeated values.
|
| 356 |
+
Pass an int for reproducible results across multiple function calls.
|
| 357 |
+
See :term:`Glossary <random_state>`.
|
| 358 |
+
|
| 359 |
+
Returns
|
| 360 |
+
-------
|
| 361 |
+
mi : ndarray, shape (n_features,)
|
| 362 |
+
Estimated mutual information between each feature and the target.
|
| 363 |
+
|
| 364 |
+
Notes
|
| 365 |
+
-----
|
| 366 |
+
1. The term "discrete features" is used instead of naming them
|
| 367 |
+
"categorical", because it describes the essence more accurately.
|
| 368 |
+
For example, pixel intensities of an image are discrete features
|
| 369 |
+
(but hardly categorical) and you will get better results if mark them
|
| 370 |
+
as such. Also note, that treating a continuous variable as discrete and
|
| 371 |
+
vice versa will usually give incorrect results, so be attentive about
|
| 372 |
+
that.
|
| 373 |
+
2. True mutual information can't be negative. If its estimate turns out
|
| 374 |
+
to be negative, it is replaced by zero.
|
| 375 |
+
|
| 376 |
+
References
|
| 377 |
+
----------
|
| 378 |
+
.. [1] `Mutual Information
|
| 379 |
+
<https://en.wikipedia.org/wiki/Mutual_information>`_
|
| 380 |
+
on Wikipedia.
|
| 381 |
+
.. [2] A. Kraskov, H. Stogbauer and P. Grassberger, "Estimating mutual
|
| 382 |
+
information". Phys. Rev. E 69, 2004.
|
| 383 |
+
.. [3] B. C. Ross "Mutual Information between Discrete and Continuous
|
| 384 |
+
Data Sets". PLoS ONE 9(2), 2014.
|
| 385 |
+
.. [4] L. F. Kozachenko, N. N. Leonenko, "Sample Estimate of the Entropy
|
| 386 |
+
of a Random Vector", Probl. Peredachi Inf., 23:2 (1987), 9-16
|
| 387 |
+
"""
|
| 388 |
+
return _estimate_mi(X, y, discrete_features, False, n_neighbors, copy, random_state)
|
| 389 |
+
|
| 390 |
+
|
| 391 |
+
def mutual_info_classif(
|
| 392 |
+
X, y, *, discrete_features="auto", n_neighbors=3, copy=True, random_state=None
|
| 393 |
+
):
|
| 394 |
+
"""Estimate mutual information for a discrete target variable.
|
| 395 |
+
|
| 396 |
+
Mutual information (MI) [1]_ between two random variables is a non-negative
|
| 397 |
+
value, which measures the dependency between the variables. It is equal
|
| 398 |
+
to zero if and only if two random variables are independent, and higher
|
| 399 |
+
values mean higher dependency.
|
| 400 |
+
|
| 401 |
+
The function relies on nonparametric methods based on entropy estimation
|
| 402 |
+
from k-nearest neighbors distances as described in [2]_ and [3]_. Both
|
| 403 |
+
methods are based on the idea originally proposed in [4]_.
|
| 404 |
+
|
| 405 |
+
It can be used for univariate features selection, read more in the
|
| 406 |
+
:ref:`User Guide <univariate_feature_selection>`.
|
| 407 |
+
|
| 408 |
+
Parameters
|
| 409 |
+
----------
|
| 410 |
+
X : array-like or sparse matrix, shape (n_samples, n_features)
|
| 411 |
+
Feature matrix.
|
| 412 |
+
|
| 413 |
+
y : array-like of shape (n_samples,)
|
| 414 |
+
Target vector.
|
| 415 |
+
|
| 416 |
+
discrete_features : {'auto', bool, array-like}, default='auto'
|
| 417 |
+
If bool, then determines whether to consider all features discrete
|
| 418 |
+
or continuous. If array, then it should be either a boolean mask
|
| 419 |
+
with shape (n_features,) or array with indices of discrete features.
|
| 420 |
+
If 'auto', it is assigned to False for dense `X` and to True for
|
| 421 |
+
sparse `X`.
|
| 422 |
+
|
| 423 |
+
n_neighbors : int, default=3
|
| 424 |
+
Number of neighbors to use for MI estimation for continuous variables,
|
| 425 |
+
see [2]_ and [3]_. Higher values reduce variance of the estimation, but
|
| 426 |
+
could introduce a bias.
|
| 427 |
+
|
| 428 |
+
copy : bool, default=True
|
| 429 |
+
Whether to make a copy of the given data. If set to False, the initial
|
| 430 |
+
data will be overwritten.
|
| 431 |
+
|
| 432 |
+
random_state : int, RandomState instance or None, default=None
|
| 433 |
+
Determines random number generation for adding small noise to
|
| 434 |
+
continuous variables in order to remove repeated values.
|
| 435 |
+
Pass an int for reproducible results across multiple function calls.
|
| 436 |
+
See :term:`Glossary <random_state>`.
|
| 437 |
+
|
| 438 |
+
Returns
|
| 439 |
+
-------
|
| 440 |
+
mi : ndarray, shape (n_features,)
|
| 441 |
+
Estimated mutual information between each feature and the target.
|
| 442 |
+
|
| 443 |
+
Notes
|
| 444 |
+
-----
|
| 445 |
+
1. The term "discrete features" is used instead of naming them
|
| 446 |
+
"categorical", because it describes the essence more accurately.
|
| 447 |
+
For example, pixel intensities of an image are discrete features
|
| 448 |
+
(but hardly categorical) and you will get better results if mark them
|
| 449 |
+
as such. Also note, that treating a continuous variable as discrete and
|
| 450 |
+
vice versa will usually give incorrect results, so be attentive about
|
| 451 |
+
that.
|
| 452 |
+
2. True mutual information can't be negative. If its estimate turns out
|
| 453 |
+
to be negative, it is replaced by zero.
|
| 454 |
+
|
| 455 |
+
References
|
| 456 |
+
----------
|
| 457 |
+
.. [1] `Mutual Information
|
| 458 |
+
<https://en.wikipedia.org/wiki/Mutual_information>`_
|
| 459 |
+
on Wikipedia.
|
| 460 |
+
.. [2] A. Kraskov, H. Stogbauer and P. Grassberger, "Estimating mutual
|
| 461 |
+
information". Phys. Rev. E 69, 2004.
|
| 462 |
+
.. [3] B. C. Ross "Mutual Information between Discrete and Continuous
|
| 463 |
+
Data Sets". PLoS ONE 9(2), 2014.
|
| 464 |
+
.. [4] L. F. Kozachenko, N. N. Leonenko, "Sample Estimate of the Entropy
|
| 465 |
+
of a Random Vector:, Probl. Peredachi Inf., 23:2 (1987), 9-16
|
| 466 |
+
"""
|
| 467 |
+
check_classification_targets(y)
|
| 468 |
+
return _estimate_mi(X, y, discrete_features, True, n_neighbors, copy, random_state)
|
videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_rfe.py
ADDED
|
@@ -0,0 +1,765 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Authors: Alexandre Gramfort <alexandre.gramfort@inria.fr>
|
| 2 |
+
# Vincent Michel <vincent.michel@inria.fr>
|
| 3 |
+
# Gilles Louppe <g.louppe@gmail.com>
|
| 4 |
+
#
|
| 5 |
+
# License: BSD 3 clause
|
| 6 |
+
|
| 7 |
+
"""Recursive feature elimination for feature ranking"""
|
| 8 |
+
|
| 9 |
+
import numpy as np
|
| 10 |
+
from numbers import Integral, Real
|
| 11 |
+
from joblib import effective_n_jobs
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
from ..utils.metaestimators import available_if
|
| 15 |
+
from ..utils.metaestimators import _safe_split
|
| 16 |
+
from ..utils._param_validation import HasMethods, Interval
|
| 17 |
+
from ..utils._tags import _safe_tags
|
| 18 |
+
from ..utils.validation import check_is_fitted
|
| 19 |
+
from ..utils.parallel import delayed, Parallel
|
| 20 |
+
from ..base import BaseEstimator
|
| 21 |
+
from ..base import MetaEstimatorMixin
|
| 22 |
+
from ..base import clone
|
| 23 |
+
from ..base import is_classifier
|
| 24 |
+
from ..model_selection import check_cv
|
| 25 |
+
from ..model_selection._validation import _score
|
| 26 |
+
from ..metrics import check_scoring
|
| 27 |
+
from ._base import SelectorMixin
|
| 28 |
+
from ._base import _get_feature_importances
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
def _rfe_single_fit(rfe, estimator, X, y, train, test, scorer):
|
| 32 |
+
"""
|
| 33 |
+
Return the score for a fit across one fold.
|
| 34 |
+
"""
|
| 35 |
+
X_train, y_train = _safe_split(estimator, X, y, train)
|
| 36 |
+
X_test, y_test = _safe_split(estimator, X, y, test, train)
|
| 37 |
+
return rfe._fit(
|
| 38 |
+
X_train,
|
| 39 |
+
y_train,
|
| 40 |
+
lambda estimator, features: _score(
|
| 41 |
+
estimator, X_test[:, features], y_test, scorer
|
| 42 |
+
),
|
| 43 |
+
).scores_
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def _estimator_has(attr):
|
| 47 |
+
"""Check if we can delegate a method to the underlying estimator.
|
| 48 |
+
|
| 49 |
+
First, we check the first fitted estimator if available, otherwise we
|
| 50 |
+
check the unfitted estimator.
|
| 51 |
+
"""
|
| 52 |
+
return lambda self: (
|
| 53 |
+
hasattr(self.estimator_, attr)
|
| 54 |
+
if hasattr(self, "estimator_")
|
| 55 |
+
else hasattr(self.estimator, attr)
|
| 56 |
+
)
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
class RFE(SelectorMixin, MetaEstimatorMixin, BaseEstimator):
|
| 60 |
+
"""Feature ranking with recursive feature elimination.
|
| 61 |
+
|
| 62 |
+
Given an external estimator that assigns weights to features (e.g., the
|
| 63 |
+
coefficients of a linear model), the goal of recursive feature elimination
|
| 64 |
+
(RFE) is to select features by recursively considering smaller and smaller
|
| 65 |
+
sets of features. First, the estimator is trained on the initial set of
|
| 66 |
+
features and the importance of each feature is obtained either through
|
| 67 |
+
any specific attribute or callable.
|
| 68 |
+
Then, the least important features are pruned from current set of features.
|
| 69 |
+
That procedure is recursively repeated on the pruned set until the desired
|
| 70 |
+
number of features to select is eventually reached.
|
| 71 |
+
|
| 72 |
+
Read more in the :ref:`User Guide <rfe>`.
|
| 73 |
+
|
| 74 |
+
Parameters
|
| 75 |
+
----------
|
| 76 |
+
estimator : ``Estimator`` instance
|
| 77 |
+
A supervised learning estimator with a ``fit`` method that provides
|
| 78 |
+
information about feature importance
|
| 79 |
+
(e.g. `coef_`, `feature_importances_`).
|
| 80 |
+
|
| 81 |
+
n_features_to_select : int or float, default=None
|
| 82 |
+
The number of features to select. If `None`, half of the features are
|
| 83 |
+
selected. If integer, the parameter is the absolute number of features
|
| 84 |
+
to select. If float between 0 and 1, it is the fraction of features to
|
| 85 |
+
select.
|
| 86 |
+
|
| 87 |
+
.. versionchanged:: 0.24
|
| 88 |
+
Added float values for fractions.
|
| 89 |
+
|
| 90 |
+
step : int or float, default=1
|
| 91 |
+
If greater than or equal to 1, then ``step`` corresponds to the
|
| 92 |
+
(integer) number of features to remove at each iteration.
|
| 93 |
+
If within (0.0, 1.0), then ``step`` corresponds to the percentage
|
| 94 |
+
(rounded down) of features to remove at each iteration.
|
| 95 |
+
|
| 96 |
+
verbose : int, default=0
|
| 97 |
+
Controls verbosity of output.
|
| 98 |
+
|
| 99 |
+
importance_getter : str or callable, default='auto'
|
| 100 |
+
If 'auto', uses the feature importance either through a `coef_`
|
| 101 |
+
or `feature_importances_` attributes of estimator.
|
| 102 |
+
|
| 103 |
+
Also accepts a string that specifies an attribute name/path
|
| 104 |
+
for extracting feature importance (implemented with `attrgetter`).
|
| 105 |
+
For example, give `regressor_.coef_` in case of
|
| 106 |
+
:class:`~sklearn.compose.TransformedTargetRegressor` or
|
| 107 |
+
`named_steps.clf.feature_importances_` in case of
|
| 108 |
+
class:`~sklearn.pipeline.Pipeline` with its last step named `clf`.
|
| 109 |
+
|
| 110 |
+
If `callable`, overrides the default feature importance getter.
|
| 111 |
+
The callable is passed with the fitted estimator and it should
|
| 112 |
+
return importance for each feature.
|
| 113 |
+
|
| 114 |
+
.. versionadded:: 0.24
|
| 115 |
+
|
| 116 |
+
Attributes
|
| 117 |
+
----------
|
| 118 |
+
classes_ : ndarray of shape (n_classes,)
|
| 119 |
+
The classes labels. Only available when `estimator` is a classifier.
|
| 120 |
+
|
| 121 |
+
estimator_ : ``Estimator`` instance
|
| 122 |
+
The fitted estimator used to select features.
|
| 123 |
+
|
| 124 |
+
n_features_ : int
|
| 125 |
+
The number of selected features.
|
| 126 |
+
|
| 127 |
+
n_features_in_ : int
|
| 128 |
+
Number of features seen during :term:`fit`. Only defined if the
|
| 129 |
+
underlying estimator exposes such an attribute when fit.
|
| 130 |
+
|
| 131 |
+
.. versionadded:: 0.24
|
| 132 |
+
|
| 133 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 134 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 135 |
+
has feature names that are all strings.
|
| 136 |
+
|
| 137 |
+
.. versionadded:: 1.0
|
| 138 |
+
|
| 139 |
+
ranking_ : ndarray of shape (n_features,)
|
| 140 |
+
The feature ranking, such that ``ranking_[i]`` corresponds to the
|
| 141 |
+
ranking position of the i-th feature. Selected (i.e., estimated
|
| 142 |
+
best) features are assigned rank 1.
|
| 143 |
+
|
| 144 |
+
support_ : ndarray of shape (n_features,)
|
| 145 |
+
The mask of selected features.
|
| 146 |
+
|
| 147 |
+
See Also
|
| 148 |
+
--------
|
| 149 |
+
RFECV : Recursive feature elimination with built-in cross-validated
|
| 150 |
+
selection of the best number of features.
|
| 151 |
+
SelectFromModel : Feature selection based on thresholds of importance
|
| 152 |
+
weights.
|
| 153 |
+
SequentialFeatureSelector : Sequential cross-validation based feature
|
| 154 |
+
selection. Does not rely on importance weights.
|
| 155 |
+
|
| 156 |
+
Notes
|
| 157 |
+
-----
|
| 158 |
+
Allows NaN/Inf in the input if the underlying estimator does as well.
|
| 159 |
+
|
| 160 |
+
References
|
| 161 |
+
----------
|
| 162 |
+
|
| 163 |
+
.. [1] Guyon, I., Weston, J., Barnhill, S., & Vapnik, V., "Gene selection
|
| 164 |
+
for cancer classification using support vector machines",
|
| 165 |
+
Mach. Learn., 46(1-3), 389--422, 2002.
|
| 166 |
+
|
| 167 |
+
Examples
|
| 168 |
+
--------
|
| 169 |
+
The following example shows how to retrieve the 5 most informative
|
| 170 |
+
features in the Friedman #1 dataset.
|
| 171 |
+
|
| 172 |
+
>>> from sklearn.datasets import make_friedman1
|
| 173 |
+
>>> from sklearn.feature_selection import RFE
|
| 174 |
+
>>> from sklearn.svm import SVR
|
| 175 |
+
>>> X, y = make_friedman1(n_samples=50, n_features=10, random_state=0)
|
| 176 |
+
>>> estimator = SVR(kernel="linear")
|
| 177 |
+
>>> selector = RFE(estimator, n_features_to_select=5, step=1)
|
| 178 |
+
>>> selector = selector.fit(X, y)
|
| 179 |
+
>>> selector.support_
|
| 180 |
+
array([ True, True, True, True, True, False, False, False, False,
|
| 181 |
+
False])
|
| 182 |
+
>>> selector.ranking_
|
| 183 |
+
array([1, 1, 1, 1, 1, 6, 4, 3, 2, 5])
|
| 184 |
+
"""
|
| 185 |
+
|
| 186 |
+
_parameter_constraints: dict = {
|
| 187 |
+
"estimator": [HasMethods(["fit"])],
|
| 188 |
+
"n_features_to_select": [
|
| 189 |
+
None,
|
| 190 |
+
Interval(Real, 0, 1, closed="right"),
|
| 191 |
+
Interval(Integral, 0, None, closed="neither"),
|
| 192 |
+
],
|
| 193 |
+
"step": [
|
| 194 |
+
Interval(Integral, 0, None, closed="neither"),
|
| 195 |
+
Interval(Real, 0, 1, closed="neither"),
|
| 196 |
+
],
|
| 197 |
+
"verbose": ["verbose"],
|
| 198 |
+
"importance_getter": [str, callable],
|
| 199 |
+
}
|
| 200 |
+
|
| 201 |
+
def __init__(
|
| 202 |
+
self,
|
| 203 |
+
estimator,
|
| 204 |
+
*,
|
| 205 |
+
n_features_to_select=None,
|
| 206 |
+
step=1,
|
| 207 |
+
verbose=0,
|
| 208 |
+
importance_getter="auto",
|
| 209 |
+
):
|
| 210 |
+
self.estimator = estimator
|
| 211 |
+
self.n_features_to_select = n_features_to_select
|
| 212 |
+
self.step = step
|
| 213 |
+
self.importance_getter = importance_getter
|
| 214 |
+
self.verbose = verbose
|
| 215 |
+
|
| 216 |
+
@property
|
| 217 |
+
def _estimator_type(self):
|
| 218 |
+
return self.estimator._estimator_type
|
| 219 |
+
|
| 220 |
+
@property
|
| 221 |
+
def classes_(self):
|
| 222 |
+
"""Classes labels available when `estimator` is a classifier.
|
| 223 |
+
|
| 224 |
+
Returns
|
| 225 |
+
-------
|
| 226 |
+
ndarray of shape (n_classes,)
|
| 227 |
+
"""
|
| 228 |
+
return self.estimator_.classes_
|
| 229 |
+
|
| 230 |
+
def fit(self, X, y, **fit_params):
|
| 231 |
+
"""Fit the RFE model and then the underlying estimator on the selected features.
|
| 232 |
+
|
| 233 |
+
Parameters
|
| 234 |
+
----------
|
| 235 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 236 |
+
The training input samples.
|
| 237 |
+
|
| 238 |
+
y : array-like of shape (n_samples,)
|
| 239 |
+
The target values.
|
| 240 |
+
|
| 241 |
+
**fit_params : dict
|
| 242 |
+
Additional parameters passed to the `fit` method of the underlying
|
| 243 |
+
estimator.
|
| 244 |
+
|
| 245 |
+
Returns
|
| 246 |
+
-------
|
| 247 |
+
self : object
|
| 248 |
+
Fitted estimator.
|
| 249 |
+
"""
|
| 250 |
+
self._validate_params()
|
| 251 |
+
return self._fit(X, y, **fit_params)
|
| 252 |
+
|
| 253 |
+
def _fit(self, X, y, step_score=None, **fit_params):
|
| 254 |
+
# Parameter step_score controls the calculation of self.scores_
|
| 255 |
+
# step_score is not exposed to users
|
| 256 |
+
# and is used when implementing RFECV
|
| 257 |
+
# self.scores_ will not be calculated when calling _fit through fit
|
| 258 |
+
|
| 259 |
+
tags = self._get_tags()
|
| 260 |
+
X, y = self._validate_data(
|
| 261 |
+
X,
|
| 262 |
+
y,
|
| 263 |
+
accept_sparse="csc",
|
| 264 |
+
ensure_min_features=2,
|
| 265 |
+
force_all_finite=not tags.get("allow_nan", True),
|
| 266 |
+
multi_output=True,
|
| 267 |
+
)
|
| 268 |
+
|
| 269 |
+
# Initialization
|
| 270 |
+
n_features = X.shape[1]
|
| 271 |
+
if self.n_features_to_select is None:
|
| 272 |
+
n_features_to_select = n_features // 2
|
| 273 |
+
elif isinstance(self.n_features_to_select, Integral): # int
|
| 274 |
+
n_features_to_select = self.n_features_to_select
|
| 275 |
+
else: # float
|
| 276 |
+
n_features_to_select = int(n_features * self.n_features_to_select)
|
| 277 |
+
|
| 278 |
+
if 0.0 < self.step < 1.0:
|
| 279 |
+
step = int(max(1, self.step * n_features))
|
| 280 |
+
else:
|
| 281 |
+
step = int(self.step)
|
| 282 |
+
|
| 283 |
+
support_ = np.ones(n_features, dtype=bool)
|
| 284 |
+
ranking_ = np.ones(n_features, dtype=int)
|
| 285 |
+
|
| 286 |
+
if step_score:
|
| 287 |
+
self.scores_ = []
|
| 288 |
+
|
| 289 |
+
# Elimination
|
| 290 |
+
while np.sum(support_) > n_features_to_select:
|
| 291 |
+
# Remaining features
|
| 292 |
+
features = np.arange(n_features)[support_]
|
| 293 |
+
|
| 294 |
+
# Rank the remaining features
|
| 295 |
+
estimator = clone(self.estimator)
|
| 296 |
+
if self.verbose > 0:
|
| 297 |
+
print("Fitting estimator with %d features." % np.sum(support_))
|
| 298 |
+
|
| 299 |
+
estimator.fit(X[:, features], y, **fit_params)
|
| 300 |
+
|
| 301 |
+
# Get importance and rank them
|
| 302 |
+
importances = _get_feature_importances(
|
| 303 |
+
estimator,
|
| 304 |
+
self.importance_getter,
|
| 305 |
+
transform_func="square",
|
| 306 |
+
)
|
| 307 |
+
ranks = np.argsort(importances)
|
| 308 |
+
|
| 309 |
+
# for sparse case ranks is matrix
|
| 310 |
+
ranks = np.ravel(ranks)
|
| 311 |
+
|
| 312 |
+
# Eliminate the worse features
|
| 313 |
+
threshold = min(step, np.sum(support_) - n_features_to_select)
|
| 314 |
+
|
| 315 |
+
# Compute step score on the previous selection iteration
|
| 316 |
+
# because 'estimator' must use features
|
| 317 |
+
# that have not been eliminated yet
|
| 318 |
+
if step_score:
|
| 319 |
+
self.scores_.append(step_score(estimator, features))
|
| 320 |
+
support_[features[ranks][:threshold]] = False
|
| 321 |
+
ranking_[np.logical_not(support_)] += 1
|
| 322 |
+
|
| 323 |
+
# Set final attributes
|
| 324 |
+
features = np.arange(n_features)[support_]
|
| 325 |
+
self.estimator_ = clone(self.estimator)
|
| 326 |
+
self.estimator_.fit(X[:, features], y, **fit_params)
|
| 327 |
+
|
| 328 |
+
# Compute step score when only n_features_to_select features left
|
| 329 |
+
if step_score:
|
| 330 |
+
self.scores_.append(step_score(self.estimator_, features))
|
| 331 |
+
self.n_features_ = support_.sum()
|
| 332 |
+
self.support_ = support_
|
| 333 |
+
self.ranking_ = ranking_
|
| 334 |
+
|
| 335 |
+
return self
|
| 336 |
+
|
| 337 |
+
@available_if(_estimator_has("predict"))
|
| 338 |
+
def predict(self, X):
|
| 339 |
+
"""Reduce X to the selected features and predict using the estimator.
|
| 340 |
+
|
| 341 |
+
Parameters
|
| 342 |
+
----------
|
| 343 |
+
X : array of shape [n_samples, n_features]
|
| 344 |
+
The input samples.
|
| 345 |
+
|
| 346 |
+
Returns
|
| 347 |
+
-------
|
| 348 |
+
y : array of shape [n_samples]
|
| 349 |
+
The predicted target values.
|
| 350 |
+
"""
|
| 351 |
+
check_is_fitted(self)
|
| 352 |
+
return self.estimator_.predict(self.transform(X))
|
| 353 |
+
|
| 354 |
+
@available_if(_estimator_has("score"))
|
| 355 |
+
def score(self, X, y, **fit_params):
|
| 356 |
+
"""Reduce X to the selected features and return the score of the estimator.
|
| 357 |
+
|
| 358 |
+
Parameters
|
| 359 |
+
----------
|
| 360 |
+
X : array of shape [n_samples, n_features]
|
| 361 |
+
The input samples.
|
| 362 |
+
|
| 363 |
+
y : array of shape [n_samples]
|
| 364 |
+
The target values.
|
| 365 |
+
|
| 366 |
+
**fit_params : dict
|
| 367 |
+
Parameters to pass to the `score` method of the underlying
|
| 368 |
+
estimator.
|
| 369 |
+
|
| 370 |
+
.. versionadded:: 1.0
|
| 371 |
+
|
| 372 |
+
Returns
|
| 373 |
+
-------
|
| 374 |
+
score : float
|
| 375 |
+
Score of the underlying base estimator computed with the selected
|
| 376 |
+
features returned by `rfe.transform(X)` and `y`.
|
| 377 |
+
"""
|
| 378 |
+
check_is_fitted(self)
|
| 379 |
+
return self.estimator_.score(self.transform(X), y, **fit_params)
|
| 380 |
+
|
| 381 |
+
def _get_support_mask(self):
|
| 382 |
+
check_is_fitted(self)
|
| 383 |
+
return self.support_
|
| 384 |
+
|
| 385 |
+
@available_if(_estimator_has("decision_function"))
|
| 386 |
+
def decision_function(self, X):
|
| 387 |
+
"""Compute the decision function of ``X``.
|
| 388 |
+
|
| 389 |
+
Parameters
|
| 390 |
+
----------
|
| 391 |
+
X : {array-like or sparse matrix} of shape (n_samples, n_features)
|
| 392 |
+
The input samples. Internally, it will be converted to
|
| 393 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 394 |
+
to a sparse ``csr_matrix``.
|
| 395 |
+
|
| 396 |
+
Returns
|
| 397 |
+
-------
|
| 398 |
+
score : array, shape = [n_samples, n_classes] or [n_samples]
|
| 399 |
+
The decision function of the input samples. The order of the
|
| 400 |
+
classes corresponds to that in the attribute :term:`classes_`.
|
| 401 |
+
Regression and binary classification produce an array of shape
|
| 402 |
+
[n_samples].
|
| 403 |
+
"""
|
| 404 |
+
check_is_fitted(self)
|
| 405 |
+
return self.estimator_.decision_function(self.transform(X))
|
| 406 |
+
|
| 407 |
+
@available_if(_estimator_has("predict_proba"))
|
| 408 |
+
def predict_proba(self, X):
|
| 409 |
+
"""Predict class probabilities for X.
|
| 410 |
+
|
| 411 |
+
Parameters
|
| 412 |
+
----------
|
| 413 |
+
X : {array-like or sparse matrix} of shape (n_samples, n_features)
|
| 414 |
+
The input samples. Internally, it will be converted to
|
| 415 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 416 |
+
to a sparse ``csr_matrix``.
|
| 417 |
+
|
| 418 |
+
Returns
|
| 419 |
+
-------
|
| 420 |
+
p : array of shape (n_samples, n_classes)
|
| 421 |
+
The class probabilities of the input samples. The order of the
|
| 422 |
+
classes corresponds to that in the attribute :term:`classes_`.
|
| 423 |
+
"""
|
| 424 |
+
check_is_fitted(self)
|
| 425 |
+
return self.estimator_.predict_proba(self.transform(X))
|
| 426 |
+
|
| 427 |
+
@available_if(_estimator_has("predict_log_proba"))
|
| 428 |
+
def predict_log_proba(self, X):
|
| 429 |
+
"""Predict class log-probabilities for X.
|
| 430 |
+
|
| 431 |
+
Parameters
|
| 432 |
+
----------
|
| 433 |
+
X : array of shape [n_samples, n_features]
|
| 434 |
+
The input samples.
|
| 435 |
+
|
| 436 |
+
Returns
|
| 437 |
+
-------
|
| 438 |
+
p : array of shape (n_samples, n_classes)
|
| 439 |
+
The class log-probabilities of the input samples. The order of the
|
| 440 |
+
classes corresponds to that in the attribute :term:`classes_`.
|
| 441 |
+
"""
|
| 442 |
+
check_is_fitted(self)
|
| 443 |
+
return self.estimator_.predict_log_proba(self.transform(X))
|
| 444 |
+
|
| 445 |
+
def _more_tags(self):
|
| 446 |
+
return {
|
| 447 |
+
"poor_score": True,
|
| 448 |
+
"allow_nan": _safe_tags(self.estimator, key="allow_nan"),
|
| 449 |
+
"requires_y": True,
|
| 450 |
+
}
|
| 451 |
+
|
| 452 |
+
|
| 453 |
+
class RFECV(RFE):
|
| 454 |
+
"""Recursive feature elimination with cross-validation to select features.
|
| 455 |
+
|
| 456 |
+
See glossary entry for :term:`cross-validation estimator`.
|
| 457 |
+
|
| 458 |
+
Read more in the :ref:`User Guide <rfe>`.
|
| 459 |
+
|
| 460 |
+
Parameters
|
| 461 |
+
----------
|
| 462 |
+
estimator : ``Estimator`` instance
|
| 463 |
+
A supervised learning estimator with a ``fit`` method that provides
|
| 464 |
+
information about feature importance either through a ``coef_``
|
| 465 |
+
attribute or through a ``feature_importances_`` attribute.
|
| 466 |
+
|
| 467 |
+
step : int or float, default=1
|
| 468 |
+
If greater than or equal to 1, then ``step`` corresponds to the
|
| 469 |
+
(integer) number of features to remove at each iteration.
|
| 470 |
+
If within (0.0, 1.0), then ``step`` corresponds to the percentage
|
| 471 |
+
(rounded down) of features to remove at each iteration.
|
| 472 |
+
Note that the last iteration may remove fewer than ``step`` features in
|
| 473 |
+
order to reach ``min_features_to_select``.
|
| 474 |
+
|
| 475 |
+
min_features_to_select : int, default=1
|
| 476 |
+
The minimum number of features to be selected. This number of features
|
| 477 |
+
will always be scored, even if the difference between the original
|
| 478 |
+
feature count and ``min_features_to_select`` isn't divisible by
|
| 479 |
+
``step``.
|
| 480 |
+
|
| 481 |
+
.. versionadded:: 0.20
|
| 482 |
+
|
| 483 |
+
cv : int, cross-validation generator or an iterable, default=None
|
| 484 |
+
Determines the cross-validation splitting strategy.
|
| 485 |
+
Possible inputs for cv are:
|
| 486 |
+
|
| 487 |
+
- None, to use the default 5-fold cross-validation,
|
| 488 |
+
- integer, to specify the number of folds.
|
| 489 |
+
- :term:`CV splitter`,
|
| 490 |
+
- An iterable yielding (train, test) splits as arrays of indices.
|
| 491 |
+
|
| 492 |
+
For integer/None inputs, if ``y`` is binary or multiclass,
|
| 493 |
+
:class:`~sklearn.model_selection.StratifiedKFold` is used. If the
|
| 494 |
+
estimator is a classifier or if ``y`` is neither binary nor multiclass,
|
| 495 |
+
:class:`~sklearn.model_selection.KFold` is used.
|
| 496 |
+
|
| 497 |
+
Refer :ref:`User Guide <cross_validation>` for the various
|
| 498 |
+
cross-validation strategies that can be used here.
|
| 499 |
+
|
| 500 |
+
.. versionchanged:: 0.22
|
| 501 |
+
``cv`` default value of None changed from 3-fold to 5-fold.
|
| 502 |
+
|
| 503 |
+
scoring : str, callable or None, default=None
|
| 504 |
+
A string (see model evaluation documentation) or
|
| 505 |
+
a scorer callable object / function with signature
|
| 506 |
+
``scorer(estimator, X, y)``.
|
| 507 |
+
|
| 508 |
+
verbose : int, default=0
|
| 509 |
+
Controls verbosity of output.
|
| 510 |
+
|
| 511 |
+
n_jobs : int or None, default=None
|
| 512 |
+
Number of cores to run in parallel while fitting across folds.
|
| 513 |
+
``None`` means 1 unless in a :obj:`joblib.parallel_backend` context.
|
| 514 |
+
``-1`` means using all processors. See :term:`Glossary <n_jobs>`
|
| 515 |
+
for more details.
|
| 516 |
+
|
| 517 |
+
.. versionadded:: 0.18
|
| 518 |
+
|
| 519 |
+
importance_getter : str or callable, default='auto'
|
| 520 |
+
If 'auto', uses the feature importance either through a `coef_`
|
| 521 |
+
or `feature_importances_` attributes of estimator.
|
| 522 |
+
|
| 523 |
+
Also accepts a string that specifies an attribute name/path
|
| 524 |
+
for extracting feature importance.
|
| 525 |
+
For example, give `regressor_.coef_` in case of
|
| 526 |
+
:class:`~sklearn.compose.TransformedTargetRegressor` or
|
| 527 |
+
`named_steps.clf.feature_importances_` in case of
|
| 528 |
+
:class:`~sklearn.pipeline.Pipeline` with its last step named `clf`.
|
| 529 |
+
|
| 530 |
+
If `callable`, overrides the default feature importance getter.
|
| 531 |
+
The callable is passed with the fitted estimator and it should
|
| 532 |
+
return importance for each feature.
|
| 533 |
+
|
| 534 |
+
.. versionadded:: 0.24
|
| 535 |
+
|
| 536 |
+
Attributes
|
| 537 |
+
----------
|
| 538 |
+
classes_ : ndarray of shape (n_classes,)
|
| 539 |
+
The classes labels. Only available when `estimator` is a classifier.
|
| 540 |
+
|
| 541 |
+
estimator_ : ``Estimator`` instance
|
| 542 |
+
The fitted estimator used to select features.
|
| 543 |
+
|
| 544 |
+
cv_results_ : dict of ndarrays
|
| 545 |
+
A dict with keys:
|
| 546 |
+
|
| 547 |
+
split(k)_test_score : ndarray of shape (n_subsets_of_features,)
|
| 548 |
+
The cross-validation scores across (k)th fold.
|
| 549 |
+
|
| 550 |
+
mean_test_score : ndarray of shape (n_subsets_of_features,)
|
| 551 |
+
Mean of scores over the folds.
|
| 552 |
+
|
| 553 |
+
std_test_score : ndarray of shape (n_subsets_of_features,)
|
| 554 |
+
Standard deviation of scores over the folds.
|
| 555 |
+
|
| 556 |
+
.. versionadded:: 1.0
|
| 557 |
+
|
| 558 |
+
n_features_ : int
|
| 559 |
+
The number of selected features with cross-validation.
|
| 560 |
+
|
| 561 |
+
n_features_in_ : int
|
| 562 |
+
Number of features seen during :term:`fit`. Only defined if the
|
| 563 |
+
underlying estimator exposes such an attribute when fit.
|
| 564 |
+
|
| 565 |
+
.. versionadded:: 0.24
|
| 566 |
+
|
| 567 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 568 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 569 |
+
has feature names that are all strings.
|
| 570 |
+
|
| 571 |
+
.. versionadded:: 1.0
|
| 572 |
+
|
| 573 |
+
ranking_ : narray of shape (n_features,)
|
| 574 |
+
The feature ranking, such that `ranking_[i]`
|
| 575 |
+
corresponds to the ranking
|
| 576 |
+
position of the i-th feature.
|
| 577 |
+
Selected (i.e., estimated best)
|
| 578 |
+
features are assigned rank 1.
|
| 579 |
+
|
| 580 |
+
support_ : ndarray of shape (n_features,)
|
| 581 |
+
The mask of selected features.
|
| 582 |
+
|
| 583 |
+
See Also
|
| 584 |
+
--------
|
| 585 |
+
RFE : Recursive feature elimination.
|
| 586 |
+
|
| 587 |
+
Notes
|
| 588 |
+
-----
|
| 589 |
+
The size of all values in ``cv_results_`` is equal to
|
| 590 |
+
``ceil((n_features - min_features_to_select) / step) + 1``,
|
| 591 |
+
where step is the number of features removed at each iteration.
|
| 592 |
+
|
| 593 |
+
Allows NaN/Inf in the input if the underlying estimator does as well.
|
| 594 |
+
|
| 595 |
+
References
|
| 596 |
+
----------
|
| 597 |
+
|
| 598 |
+
.. [1] Guyon, I., Weston, J., Barnhill, S., & Vapnik, V., "Gene selection
|
| 599 |
+
for cancer classification using support vector machines",
|
| 600 |
+
Mach. Learn., 46(1-3), 389--422, 2002.
|
| 601 |
+
|
| 602 |
+
Examples
|
| 603 |
+
--------
|
| 604 |
+
The following example shows how to retrieve the a-priori not known 5
|
| 605 |
+
informative features in the Friedman #1 dataset.
|
| 606 |
+
|
| 607 |
+
>>> from sklearn.datasets import make_friedman1
|
| 608 |
+
>>> from sklearn.feature_selection import RFECV
|
| 609 |
+
>>> from sklearn.svm import SVR
|
| 610 |
+
>>> X, y = make_friedman1(n_samples=50, n_features=10, random_state=0)
|
| 611 |
+
>>> estimator = SVR(kernel="linear")
|
| 612 |
+
>>> selector = RFECV(estimator, step=1, cv=5)
|
| 613 |
+
>>> selector = selector.fit(X, y)
|
| 614 |
+
>>> selector.support_
|
| 615 |
+
array([ True, True, True, True, True, False, False, False, False,
|
| 616 |
+
False])
|
| 617 |
+
>>> selector.ranking_
|
| 618 |
+
array([1, 1, 1, 1, 1, 6, 4, 3, 2, 5])
|
| 619 |
+
"""
|
| 620 |
+
|
| 621 |
+
_parameter_constraints: dict = {
|
| 622 |
+
**RFE._parameter_constraints,
|
| 623 |
+
"min_features_to_select": [Interval(Integral, 0, None, closed="neither")],
|
| 624 |
+
"cv": ["cv_object"],
|
| 625 |
+
"scoring": [None, str, callable],
|
| 626 |
+
"n_jobs": [None, Integral],
|
| 627 |
+
}
|
| 628 |
+
_parameter_constraints.pop("n_features_to_select")
|
| 629 |
+
|
| 630 |
+
def __init__(
|
| 631 |
+
self,
|
| 632 |
+
estimator,
|
| 633 |
+
*,
|
| 634 |
+
step=1,
|
| 635 |
+
min_features_to_select=1,
|
| 636 |
+
cv=None,
|
| 637 |
+
scoring=None,
|
| 638 |
+
verbose=0,
|
| 639 |
+
n_jobs=None,
|
| 640 |
+
importance_getter="auto",
|
| 641 |
+
):
|
| 642 |
+
self.estimator = estimator
|
| 643 |
+
self.step = step
|
| 644 |
+
self.importance_getter = importance_getter
|
| 645 |
+
self.cv = cv
|
| 646 |
+
self.scoring = scoring
|
| 647 |
+
self.verbose = verbose
|
| 648 |
+
self.n_jobs = n_jobs
|
| 649 |
+
self.min_features_to_select = min_features_to_select
|
| 650 |
+
|
| 651 |
+
def fit(self, X, y, groups=None):
|
| 652 |
+
"""Fit the RFE model and automatically tune the number of selected features.
|
| 653 |
+
|
| 654 |
+
Parameters
|
| 655 |
+
----------
|
| 656 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 657 |
+
Training vector, where `n_samples` is the number of samples and
|
| 658 |
+
`n_features` is the total number of features.
|
| 659 |
+
|
| 660 |
+
y : array-like of shape (n_samples,)
|
| 661 |
+
Target values (integers for classification, real numbers for
|
| 662 |
+
regression).
|
| 663 |
+
|
| 664 |
+
groups : array-like of shape (n_samples,) or None, default=None
|
| 665 |
+
Group labels for the samples used while splitting the dataset into
|
| 666 |
+
train/test set. Only used in conjunction with a "Group" :term:`cv`
|
| 667 |
+
instance (e.g., :class:`~sklearn.model_selection.GroupKFold`).
|
| 668 |
+
|
| 669 |
+
.. versionadded:: 0.20
|
| 670 |
+
|
| 671 |
+
Returns
|
| 672 |
+
-------
|
| 673 |
+
self : object
|
| 674 |
+
Fitted estimator.
|
| 675 |
+
"""
|
| 676 |
+
self._validate_params()
|
| 677 |
+
tags = self._get_tags()
|
| 678 |
+
X, y = self._validate_data(
|
| 679 |
+
X,
|
| 680 |
+
y,
|
| 681 |
+
accept_sparse="csr",
|
| 682 |
+
ensure_min_features=2,
|
| 683 |
+
force_all_finite=not tags.get("allow_nan", True),
|
| 684 |
+
multi_output=True,
|
| 685 |
+
)
|
| 686 |
+
|
| 687 |
+
# Initialization
|
| 688 |
+
cv = check_cv(self.cv, y, classifier=is_classifier(self.estimator))
|
| 689 |
+
scorer = check_scoring(self.estimator, scoring=self.scoring)
|
| 690 |
+
n_features = X.shape[1]
|
| 691 |
+
|
| 692 |
+
if 0.0 < self.step < 1.0:
|
| 693 |
+
step = int(max(1, self.step * n_features))
|
| 694 |
+
else:
|
| 695 |
+
step = int(self.step)
|
| 696 |
+
|
| 697 |
+
# Build an RFE object, which will evaluate and score each possible
|
| 698 |
+
# feature count, down to self.min_features_to_select
|
| 699 |
+
rfe = RFE(
|
| 700 |
+
estimator=self.estimator,
|
| 701 |
+
n_features_to_select=self.min_features_to_select,
|
| 702 |
+
importance_getter=self.importance_getter,
|
| 703 |
+
step=self.step,
|
| 704 |
+
verbose=self.verbose,
|
| 705 |
+
)
|
| 706 |
+
|
| 707 |
+
# Determine the number of subsets of features by fitting across
|
| 708 |
+
# the train folds and choosing the "features_to_select" parameter
|
| 709 |
+
# that gives the least averaged error across all folds.
|
| 710 |
+
|
| 711 |
+
# Note that joblib raises a non-picklable error for bound methods
|
| 712 |
+
# even if n_jobs is set to 1 with the default multiprocessing
|
| 713 |
+
# backend.
|
| 714 |
+
# This branching is done so that to
|
| 715 |
+
# make sure that user code that sets n_jobs to 1
|
| 716 |
+
# and provides bound methods as scorers is not broken with the
|
| 717 |
+
# addition of n_jobs parameter in version 0.18.
|
| 718 |
+
|
| 719 |
+
if effective_n_jobs(self.n_jobs) == 1:
|
| 720 |
+
parallel, func = list, _rfe_single_fit
|
| 721 |
+
else:
|
| 722 |
+
parallel = Parallel(n_jobs=self.n_jobs)
|
| 723 |
+
func = delayed(_rfe_single_fit)
|
| 724 |
+
|
| 725 |
+
scores = parallel(
|
| 726 |
+
func(rfe, self.estimator, X, y, train, test, scorer)
|
| 727 |
+
for train, test in cv.split(X, y, groups)
|
| 728 |
+
)
|
| 729 |
+
|
| 730 |
+
scores = np.array(scores)
|
| 731 |
+
scores_sum = np.sum(scores, axis=0)
|
| 732 |
+
scores_sum_rev = scores_sum[::-1]
|
| 733 |
+
argmax_idx = len(scores_sum) - np.argmax(scores_sum_rev) - 1
|
| 734 |
+
n_features_to_select = max(
|
| 735 |
+
n_features - (argmax_idx * step), self.min_features_to_select
|
| 736 |
+
)
|
| 737 |
+
|
| 738 |
+
# Re-execute an elimination with best_k over the whole set
|
| 739 |
+
rfe = RFE(
|
| 740 |
+
estimator=self.estimator,
|
| 741 |
+
n_features_to_select=n_features_to_select,
|
| 742 |
+
step=self.step,
|
| 743 |
+
importance_getter=self.importance_getter,
|
| 744 |
+
verbose=self.verbose,
|
| 745 |
+
)
|
| 746 |
+
|
| 747 |
+
rfe.fit(X, y)
|
| 748 |
+
|
| 749 |
+
# Set final attributes
|
| 750 |
+
self.support_ = rfe.support_
|
| 751 |
+
self.n_features_ = rfe.n_features_
|
| 752 |
+
self.ranking_ = rfe.ranking_
|
| 753 |
+
self.estimator_ = clone(self.estimator)
|
| 754 |
+
self.estimator_.fit(self._transform(X), y)
|
| 755 |
+
|
| 756 |
+
# reverse to stay consistent with before
|
| 757 |
+
scores_rev = scores[:, ::-1]
|
| 758 |
+
self.cv_results_ = {}
|
| 759 |
+
self.cv_results_["mean_test_score"] = np.mean(scores_rev, axis=0)
|
| 760 |
+
self.cv_results_["std_test_score"] = np.std(scores_rev, axis=0)
|
| 761 |
+
|
| 762 |
+
for i in range(scores.shape[0]):
|
| 763 |
+
self.cv_results_[f"split{i}_test_score"] = scores_rev[i]
|
| 764 |
+
|
| 765 |
+
return self
|
videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_sequential.py
ADDED
|
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Sequential feature selection
|
| 3 |
+
"""
|
| 4 |
+
from numbers import Integral, Real
|
| 5 |
+
|
| 6 |
+
import numpy as np
|
| 7 |
+
|
| 8 |
+
import warnings
|
| 9 |
+
|
| 10 |
+
from ._base import SelectorMixin
|
| 11 |
+
from ..base import BaseEstimator, MetaEstimatorMixin, clone
|
| 12 |
+
from ..utils._param_validation import HasMethods, Hidden, Interval, StrOptions
|
| 13 |
+
from ..utils._tags import _safe_tags
|
| 14 |
+
from ..utils.validation import check_is_fitted
|
| 15 |
+
from ..model_selection import cross_val_score
|
| 16 |
+
from ..metrics import get_scorer_names
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
class SequentialFeatureSelector(SelectorMixin, MetaEstimatorMixin, BaseEstimator):
|
| 20 |
+
"""Transformer that performs Sequential Feature Selection.
|
| 21 |
+
|
| 22 |
+
This Sequential Feature Selector adds (forward selection) or
|
| 23 |
+
removes (backward selection) features to form a feature subset in a
|
| 24 |
+
greedy fashion. At each stage, this estimator chooses the best feature to
|
| 25 |
+
add or remove based on the cross-validation score of an estimator. In
|
| 26 |
+
the case of unsupervised learning, this Sequential Feature Selector
|
| 27 |
+
looks only at the features (X), not the desired outputs (y).
|
| 28 |
+
|
| 29 |
+
Read more in the :ref:`User Guide <sequential_feature_selection>`.
|
| 30 |
+
|
| 31 |
+
.. versionadded:: 0.24
|
| 32 |
+
|
| 33 |
+
Parameters
|
| 34 |
+
----------
|
| 35 |
+
estimator : estimator instance
|
| 36 |
+
An unfitted estimator.
|
| 37 |
+
|
| 38 |
+
n_features_to_select : "auto", int or float, default='warn'
|
| 39 |
+
If `"auto"`, the behaviour depends on the `tol` parameter:
|
| 40 |
+
|
| 41 |
+
- if `tol` is not `None`, then features are selected until the score
|
| 42 |
+
improvement does not exceed `tol`.
|
| 43 |
+
- otherwise, half of the features are selected.
|
| 44 |
+
|
| 45 |
+
If integer, the parameter is the absolute number of features to select.
|
| 46 |
+
If float between 0 and 1, it is the fraction of features to select.
|
| 47 |
+
|
| 48 |
+
.. versionadded:: 1.1
|
| 49 |
+
The option `"auto"` was added in version 1.1.
|
| 50 |
+
|
| 51 |
+
.. deprecated:: 1.1
|
| 52 |
+
The default changed from `None` to `"warn"` in 1.1 and will become
|
| 53 |
+
`"auto"` in 1.3. `None` and `'warn'` will be removed in 1.3.
|
| 54 |
+
To keep the same behaviour as `None`, set
|
| 55 |
+
`n_features_to_select="auto" and `tol=None`.
|
| 56 |
+
|
| 57 |
+
tol : float, default=None
|
| 58 |
+
If the score is not incremented by at least `tol` between two
|
| 59 |
+
consecutive feature additions or removals, stop adding or removing.
|
| 60 |
+
|
| 61 |
+
`tol` can be negative when removing features using `direction="backward"`.
|
| 62 |
+
It can be useful to reduce the number of features at the cost of a small
|
| 63 |
+
decrease in the score.
|
| 64 |
+
|
| 65 |
+
`tol` is enabled only when `n_features_to_select` is `"auto"`.
|
| 66 |
+
|
| 67 |
+
.. versionadded:: 1.1
|
| 68 |
+
|
| 69 |
+
direction : {'forward', 'backward'}, default='forward'
|
| 70 |
+
Whether to perform forward selection or backward selection.
|
| 71 |
+
|
| 72 |
+
scoring : str or callable, default=None
|
| 73 |
+
A single str (see :ref:`scoring_parameter`) or a callable
|
| 74 |
+
(see :ref:`scoring`) to evaluate the predictions on the test set.
|
| 75 |
+
|
| 76 |
+
NOTE that when using a custom scorer, it should return a single
|
| 77 |
+
value.
|
| 78 |
+
|
| 79 |
+
If None, the estimator's score method is used.
|
| 80 |
+
|
| 81 |
+
cv : int, cross-validation generator or an iterable, default=None
|
| 82 |
+
Determines the cross-validation splitting strategy.
|
| 83 |
+
Possible inputs for cv are:
|
| 84 |
+
|
| 85 |
+
- None, to use the default 5-fold cross validation,
|
| 86 |
+
- integer, to specify the number of folds in a `(Stratified)KFold`,
|
| 87 |
+
- :term:`CV splitter`,
|
| 88 |
+
- An iterable yielding (train, test) splits as arrays of indices.
|
| 89 |
+
|
| 90 |
+
For integer/None inputs, if the estimator is a classifier and ``y`` is
|
| 91 |
+
either binary or multiclass, :class:`StratifiedKFold` is used. In all
|
| 92 |
+
other cases, :class:`KFold` is used. These splitters are instantiated
|
| 93 |
+
with `shuffle=False` so the splits will be the same across calls.
|
| 94 |
+
|
| 95 |
+
Refer :ref:`User Guide <cross_validation>` for the various
|
| 96 |
+
cross-validation strategies that can be used here.
|
| 97 |
+
|
| 98 |
+
n_jobs : int, default=None
|
| 99 |
+
Number of jobs to run in parallel. When evaluating a new feature to
|
| 100 |
+
add or remove, the cross-validation procedure is parallel over the
|
| 101 |
+
folds.
|
| 102 |
+
``None`` means 1 unless in a :obj:`joblib.parallel_backend` context.
|
| 103 |
+
``-1`` means using all processors. See :term:`Glossary <n_jobs>`
|
| 104 |
+
for more details.
|
| 105 |
+
|
| 106 |
+
Attributes
|
| 107 |
+
----------
|
| 108 |
+
n_features_in_ : int
|
| 109 |
+
Number of features seen during :term:`fit`. Only defined if the
|
| 110 |
+
underlying estimator exposes such an attribute when fit.
|
| 111 |
+
|
| 112 |
+
.. versionadded:: 0.24
|
| 113 |
+
|
| 114 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 115 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 116 |
+
has feature names that are all strings.
|
| 117 |
+
|
| 118 |
+
.. versionadded:: 1.0
|
| 119 |
+
|
| 120 |
+
n_features_to_select_ : int
|
| 121 |
+
The number of features that were selected.
|
| 122 |
+
|
| 123 |
+
support_ : ndarray of shape (n_features,), dtype=bool
|
| 124 |
+
The mask of selected features.
|
| 125 |
+
|
| 126 |
+
See Also
|
| 127 |
+
--------
|
| 128 |
+
GenericUnivariateSelect : Univariate feature selector with configurable
|
| 129 |
+
strategy.
|
| 130 |
+
RFE : Recursive feature elimination based on importance weights.
|
| 131 |
+
RFECV : Recursive feature elimination based on importance weights, with
|
| 132 |
+
automatic selection of the number of features.
|
| 133 |
+
SelectFromModel : Feature selection based on thresholds of importance
|
| 134 |
+
weights.
|
| 135 |
+
|
| 136 |
+
Examples
|
| 137 |
+
--------
|
| 138 |
+
>>> from sklearn.feature_selection import SequentialFeatureSelector
|
| 139 |
+
>>> from sklearn.neighbors import KNeighborsClassifier
|
| 140 |
+
>>> from sklearn.datasets import load_iris
|
| 141 |
+
>>> X, y = load_iris(return_X_y=True)
|
| 142 |
+
>>> knn = KNeighborsClassifier(n_neighbors=3)
|
| 143 |
+
>>> sfs = SequentialFeatureSelector(knn, n_features_to_select=3)
|
| 144 |
+
>>> sfs.fit(X, y)
|
| 145 |
+
SequentialFeatureSelector(estimator=KNeighborsClassifier(n_neighbors=3),
|
| 146 |
+
n_features_to_select=3)
|
| 147 |
+
>>> sfs.get_support()
|
| 148 |
+
array([ True, False, True, True])
|
| 149 |
+
>>> sfs.transform(X).shape
|
| 150 |
+
(150, 3)
|
| 151 |
+
"""
|
| 152 |
+
|
| 153 |
+
_parameter_constraints: dict = {
|
| 154 |
+
"estimator": [HasMethods(["fit"])],
|
| 155 |
+
"n_features_to_select": [
|
| 156 |
+
StrOptions({"auto", "warn"}, deprecated={"warn"}),
|
| 157 |
+
Interval(Real, 0, 1, closed="right"),
|
| 158 |
+
Interval(Integral, 0, None, closed="neither"),
|
| 159 |
+
Hidden(None),
|
| 160 |
+
],
|
| 161 |
+
"tol": [None, Interval(Real, None, None, closed="neither")],
|
| 162 |
+
"direction": [StrOptions({"forward", "backward"})],
|
| 163 |
+
"scoring": [None, StrOptions(set(get_scorer_names())), callable],
|
| 164 |
+
"cv": ["cv_object"],
|
| 165 |
+
"n_jobs": [None, Integral],
|
| 166 |
+
}
|
| 167 |
+
|
| 168 |
+
def __init__(
|
| 169 |
+
self,
|
| 170 |
+
estimator,
|
| 171 |
+
*,
|
| 172 |
+
n_features_to_select="warn",
|
| 173 |
+
tol=None,
|
| 174 |
+
direction="forward",
|
| 175 |
+
scoring=None,
|
| 176 |
+
cv=5,
|
| 177 |
+
n_jobs=None,
|
| 178 |
+
):
|
| 179 |
+
|
| 180 |
+
self.estimator = estimator
|
| 181 |
+
self.n_features_to_select = n_features_to_select
|
| 182 |
+
self.tol = tol
|
| 183 |
+
self.direction = direction
|
| 184 |
+
self.scoring = scoring
|
| 185 |
+
self.cv = cv
|
| 186 |
+
self.n_jobs = n_jobs
|
| 187 |
+
|
| 188 |
+
def fit(self, X, y=None):
|
| 189 |
+
"""Learn the features to select from X.
|
| 190 |
+
|
| 191 |
+
Parameters
|
| 192 |
+
----------
|
| 193 |
+
X : array-like of shape (n_samples, n_features)
|
| 194 |
+
Training vectors, where `n_samples` is the number of samples and
|
| 195 |
+
`n_features` is the number of predictors.
|
| 196 |
+
|
| 197 |
+
y : array-like of shape (n_samples,), default=None
|
| 198 |
+
Target values. This parameter may be ignored for
|
| 199 |
+
unsupervised learning.
|
| 200 |
+
|
| 201 |
+
Returns
|
| 202 |
+
-------
|
| 203 |
+
self : object
|
| 204 |
+
Returns the instance itself.
|
| 205 |
+
"""
|
| 206 |
+
self._validate_params()
|
| 207 |
+
|
| 208 |
+
# FIXME: to be removed in 1.3
|
| 209 |
+
if self.n_features_to_select in ("warn", None):
|
| 210 |
+
# for backwards compatibility
|
| 211 |
+
warnings.warn(
|
| 212 |
+
"Leaving `n_features_to_select` to "
|
| 213 |
+
"None is deprecated in 1.0 and will become 'auto' "
|
| 214 |
+
"in 1.3. To keep the same behaviour as with None "
|
| 215 |
+
"(i.e. select half of the features) and avoid "
|
| 216 |
+
"this warning, you should manually set "
|
| 217 |
+
"`n_features_to_select='auto'` and set tol=None "
|
| 218 |
+
"when creating an instance.",
|
| 219 |
+
FutureWarning,
|
| 220 |
+
)
|
| 221 |
+
|
| 222 |
+
tags = self._get_tags()
|
| 223 |
+
X = self._validate_data(
|
| 224 |
+
X,
|
| 225 |
+
accept_sparse="csc",
|
| 226 |
+
ensure_min_features=2,
|
| 227 |
+
force_all_finite=not tags.get("allow_nan", True),
|
| 228 |
+
)
|
| 229 |
+
n_features = X.shape[1]
|
| 230 |
+
|
| 231 |
+
# FIXME: to be fixed in 1.3
|
| 232 |
+
error_msg = (
|
| 233 |
+
"n_features_to_select must be either 'auto', 'warn', "
|
| 234 |
+
"None, an integer in [1, n_features - 1] "
|
| 235 |
+
"representing the absolute "
|
| 236 |
+
"number of features, or a float in (0, 1] "
|
| 237 |
+
"representing a percentage of features to "
|
| 238 |
+
f"select. Got {self.n_features_to_select}"
|
| 239 |
+
)
|
| 240 |
+
if self.n_features_to_select in ("warn", None):
|
| 241 |
+
if self.tol is not None:
|
| 242 |
+
raise ValueError("tol is only enabled if `n_features_to_select='auto'`")
|
| 243 |
+
self.n_features_to_select_ = n_features // 2
|
| 244 |
+
elif self.n_features_to_select == "auto":
|
| 245 |
+
if self.tol is not None:
|
| 246 |
+
# With auto feature selection, `n_features_to_select_` will be updated
|
| 247 |
+
# to `support_.sum()` after features are selected.
|
| 248 |
+
self.n_features_to_select_ = n_features - 1
|
| 249 |
+
else:
|
| 250 |
+
self.n_features_to_select_ = n_features // 2
|
| 251 |
+
elif isinstance(self.n_features_to_select, Integral):
|
| 252 |
+
if not 0 < self.n_features_to_select < n_features:
|
| 253 |
+
raise ValueError(error_msg)
|
| 254 |
+
self.n_features_to_select_ = self.n_features_to_select
|
| 255 |
+
elif isinstance(self.n_features_to_select, Real):
|
| 256 |
+
self.n_features_to_select_ = int(n_features * self.n_features_to_select)
|
| 257 |
+
|
| 258 |
+
if self.tol is not None and self.tol < 0 and self.direction == "forward":
|
| 259 |
+
raise ValueError("tol must be positive when doing forward selection")
|
| 260 |
+
|
| 261 |
+
cloned_estimator = clone(self.estimator)
|
| 262 |
+
|
| 263 |
+
# the current mask corresponds to the set of features:
|
| 264 |
+
# - that we have already *selected* if we do forward selection
|
| 265 |
+
# - that we have already *excluded* if we do backward selection
|
| 266 |
+
current_mask = np.zeros(shape=n_features, dtype=bool)
|
| 267 |
+
n_iterations = (
|
| 268 |
+
self.n_features_to_select_
|
| 269 |
+
if self.n_features_to_select == "auto" or self.direction == "forward"
|
| 270 |
+
else n_features - self.n_features_to_select_
|
| 271 |
+
)
|
| 272 |
+
|
| 273 |
+
old_score = -np.inf
|
| 274 |
+
is_auto_select = self.tol is not None and self.n_features_to_select == "auto"
|
| 275 |
+
for _ in range(n_iterations):
|
| 276 |
+
new_feature_idx, new_score = self._get_best_new_feature_score(
|
| 277 |
+
cloned_estimator, X, y, current_mask
|
| 278 |
+
)
|
| 279 |
+
if is_auto_select and ((new_score - old_score) < self.tol):
|
| 280 |
+
break
|
| 281 |
+
|
| 282 |
+
old_score = new_score
|
| 283 |
+
current_mask[new_feature_idx] = True
|
| 284 |
+
|
| 285 |
+
if self.direction == "backward":
|
| 286 |
+
current_mask = ~current_mask
|
| 287 |
+
|
| 288 |
+
self.support_ = current_mask
|
| 289 |
+
self.n_features_to_select_ = self.support_.sum()
|
| 290 |
+
|
| 291 |
+
return self
|
| 292 |
+
|
| 293 |
+
def _get_best_new_feature_score(self, estimator, X, y, current_mask):
|
| 294 |
+
# Return the best new feature and its score to add to the current_mask,
|
| 295 |
+
# i.e. return the best new feature and its score to add (resp. remove)
|
| 296 |
+
# when doing forward selection (resp. backward selection).
|
| 297 |
+
# Feature will be added if the current score and past score are greater
|
| 298 |
+
# than tol when n_feature is auto,
|
| 299 |
+
candidate_feature_indices = np.flatnonzero(~current_mask)
|
| 300 |
+
scores = {}
|
| 301 |
+
for feature_idx in candidate_feature_indices:
|
| 302 |
+
candidate_mask = current_mask.copy()
|
| 303 |
+
candidate_mask[feature_idx] = True
|
| 304 |
+
if self.direction == "backward":
|
| 305 |
+
candidate_mask = ~candidate_mask
|
| 306 |
+
X_new = X[:, candidate_mask]
|
| 307 |
+
scores[feature_idx] = cross_val_score(
|
| 308 |
+
estimator,
|
| 309 |
+
X_new,
|
| 310 |
+
y,
|
| 311 |
+
cv=self.cv,
|
| 312 |
+
scoring=self.scoring,
|
| 313 |
+
n_jobs=self.n_jobs,
|
| 314 |
+
).mean()
|
| 315 |
+
new_feature_idx = max(scores, key=lambda feature_idx: scores[feature_idx])
|
| 316 |
+
return new_feature_idx, scores[new_feature_idx]
|
| 317 |
+
|
| 318 |
+
def _get_support_mask(self):
|
| 319 |
+
check_is_fitted(self)
|
| 320 |
+
return self.support_
|
| 321 |
+
|
| 322 |
+
def _more_tags(self):
|
| 323 |
+
return {
|
| 324 |
+
"allow_nan": _safe_tags(self.estimator, key="allow_nan"),
|
| 325 |
+
}
|
videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_univariate_selection.py
ADDED
|
@@ -0,0 +1,1052 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Univariate features selection."""
|
| 2 |
+
|
| 3 |
+
# Authors: V. Michel, B. Thirion, G. Varoquaux, A. Gramfort, E. Duchesnay.
|
| 4 |
+
# L. Buitinck, A. Joly
|
| 5 |
+
# License: BSD 3 clause
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
import numpy as np
|
| 9 |
+
import warnings
|
| 10 |
+
|
| 11 |
+
from numbers import Integral, Real
|
| 12 |
+
from scipy import special, stats
|
| 13 |
+
from scipy.sparse import issparse
|
| 14 |
+
|
| 15 |
+
from ..base import BaseEstimator
|
| 16 |
+
from ..preprocessing import LabelBinarizer
|
| 17 |
+
from ..utils import as_float_array, check_array, check_X_y, safe_sqr, safe_mask
|
| 18 |
+
from ..utils.extmath import safe_sparse_dot, row_norms
|
| 19 |
+
from ..utils.validation import check_is_fitted
|
| 20 |
+
from ..utils._param_validation import Interval, StrOptions
|
| 21 |
+
from ._base import SelectorMixin
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def _clean_nans(scores):
|
| 25 |
+
"""
|
| 26 |
+
Fixes Issue #1240: NaNs can't be properly compared, so change them to the
|
| 27 |
+
smallest value of scores's dtype. -inf seems to be unreliable.
|
| 28 |
+
"""
|
| 29 |
+
# XXX where should this function be called? fit? scoring functions
|
| 30 |
+
# themselves?
|
| 31 |
+
scores = as_float_array(scores, copy=True)
|
| 32 |
+
scores[np.isnan(scores)] = np.finfo(scores.dtype).min
|
| 33 |
+
return scores
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
######################################################################
|
| 37 |
+
# Scoring functions
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
# The following function is a rewriting of scipy.stats.f_oneway
|
| 41 |
+
# Contrary to the scipy.stats.f_oneway implementation it does not
|
| 42 |
+
# copy the data while keeping the inputs unchanged.
|
| 43 |
+
def f_oneway(*args):
|
| 44 |
+
"""Perform a 1-way ANOVA.
|
| 45 |
+
|
| 46 |
+
The one-way ANOVA tests the null hypothesis that 2 or more groups have
|
| 47 |
+
the same population mean. The test is applied to samples from two or
|
| 48 |
+
more groups, possibly with differing sizes.
|
| 49 |
+
|
| 50 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 51 |
+
|
| 52 |
+
Parameters
|
| 53 |
+
----------
|
| 54 |
+
*args : {array-like, sparse matrix}
|
| 55 |
+
Sample1, sample2... The sample measurements should be given as
|
| 56 |
+
arguments.
|
| 57 |
+
|
| 58 |
+
Returns
|
| 59 |
+
-------
|
| 60 |
+
f_statistic : float
|
| 61 |
+
The computed F-value of the test.
|
| 62 |
+
p_value : float
|
| 63 |
+
The associated p-value from the F-distribution.
|
| 64 |
+
|
| 65 |
+
Notes
|
| 66 |
+
-----
|
| 67 |
+
The ANOVA test has important assumptions that must be satisfied in order
|
| 68 |
+
for the associated p-value to be valid.
|
| 69 |
+
|
| 70 |
+
1. The samples are independent
|
| 71 |
+
2. Each sample is from a normally distributed population
|
| 72 |
+
3. The population standard deviations of the groups are all equal. This
|
| 73 |
+
property is known as homoscedasticity.
|
| 74 |
+
|
| 75 |
+
If these assumptions are not true for a given set of data, it may still be
|
| 76 |
+
possible to use the Kruskal-Wallis H-test (`scipy.stats.kruskal`_) although
|
| 77 |
+
with some loss of power.
|
| 78 |
+
|
| 79 |
+
The algorithm is from Heiman[2], pp.394-7.
|
| 80 |
+
|
| 81 |
+
See ``scipy.stats.f_oneway`` that should give the same results while
|
| 82 |
+
being less efficient.
|
| 83 |
+
|
| 84 |
+
References
|
| 85 |
+
----------
|
| 86 |
+
.. [1] Lowry, Richard. "Concepts and Applications of Inferential
|
| 87 |
+
Statistics". Chapter 14.
|
| 88 |
+
http://faculty.vassar.edu/lowry/ch14pt1.html
|
| 89 |
+
|
| 90 |
+
.. [2] Heiman, G.W. Research Methods in Statistics. 2002.
|
| 91 |
+
"""
|
| 92 |
+
n_classes = len(args)
|
| 93 |
+
args = [as_float_array(a) for a in args]
|
| 94 |
+
n_samples_per_class = np.array([a.shape[0] for a in args])
|
| 95 |
+
n_samples = np.sum(n_samples_per_class)
|
| 96 |
+
ss_alldata = sum(safe_sqr(a).sum(axis=0) for a in args)
|
| 97 |
+
sums_args = [np.asarray(a.sum(axis=0)) for a in args]
|
| 98 |
+
square_of_sums_alldata = sum(sums_args) ** 2
|
| 99 |
+
square_of_sums_args = [s**2 for s in sums_args]
|
| 100 |
+
sstot = ss_alldata - square_of_sums_alldata / float(n_samples)
|
| 101 |
+
ssbn = 0.0
|
| 102 |
+
for k, _ in enumerate(args):
|
| 103 |
+
ssbn += square_of_sums_args[k] / n_samples_per_class[k]
|
| 104 |
+
ssbn -= square_of_sums_alldata / float(n_samples)
|
| 105 |
+
sswn = sstot - ssbn
|
| 106 |
+
dfbn = n_classes - 1
|
| 107 |
+
dfwn = n_samples - n_classes
|
| 108 |
+
msb = ssbn / float(dfbn)
|
| 109 |
+
msw = sswn / float(dfwn)
|
| 110 |
+
constant_features_idx = np.where(msw == 0.0)[0]
|
| 111 |
+
if np.nonzero(msb)[0].size != msb.size and constant_features_idx.size:
|
| 112 |
+
warnings.warn("Features %s are constant." % constant_features_idx, UserWarning)
|
| 113 |
+
f = msb / msw
|
| 114 |
+
# flatten matrix to vector in sparse case
|
| 115 |
+
f = np.asarray(f).ravel()
|
| 116 |
+
prob = special.fdtrc(dfbn, dfwn, f)
|
| 117 |
+
return f, prob
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
def f_classif(X, y):
|
| 121 |
+
"""Compute the ANOVA F-value for the provided sample.
|
| 122 |
+
|
| 123 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 124 |
+
|
| 125 |
+
Parameters
|
| 126 |
+
----------
|
| 127 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 128 |
+
The set of regressors that will be tested sequentially.
|
| 129 |
+
|
| 130 |
+
y : ndarray of shape (n_samples,)
|
| 131 |
+
The target vector.
|
| 132 |
+
|
| 133 |
+
Returns
|
| 134 |
+
-------
|
| 135 |
+
f_statistic : ndarray of shape (n_features,)
|
| 136 |
+
F-statistic for each feature.
|
| 137 |
+
|
| 138 |
+
p_values : ndarray of shape (n_features,)
|
| 139 |
+
P-values associated with the F-statistic.
|
| 140 |
+
|
| 141 |
+
See Also
|
| 142 |
+
--------
|
| 143 |
+
chi2 : Chi-squared stats of non-negative features for classification tasks.
|
| 144 |
+
f_regression : F-value between label/feature for regression tasks.
|
| 145 |
+
"""
|
| 146 |
+
X, y = check_X_y(X, y, accept_sparse=["csr", "csc", "coo"])
|
| 147 |
+
args = [X[safe_mask(X, y == k)] for k in np.unique(y)]
|
| 148 |
+
return f_oneway(*args)
|
| 149 |
+
|
| 150 |
+
|
| 151 |
+
def _chisquare(f_obs, f_exp):
|
| 152 |
+
"""Fast replacement for scipy.stats.chisquare.
|
| 153 |
+
|
| 154 |
+
Version from https://github.com/scipy/scipy/pull/2525 with additional
|
| 155 |
+
optimizations.
|
| 156 |
+
"""
|
| 157 |
+
f_obs = np.asarray(f_obs, dtype=np.float64)
|
| 158 |
+
|
| 159 |
+
k = len(f_obs)
|
| 160 |
+
# Reuse f_obs for chi-squared statistics
|
| 161 |
+
chisq = f_obs
|
| 162 |
+
chisq -= f_exp
|
| 163 |
+
chisq **= 2
|
| 164 |
+
with np.errstate(invalid="ignore"):
|
| 165 |
+
chisq /= f_exp
|
| 166 |
+
chisq = chisq.sum(axis=0)
|
| 167 |
+
return chisq, special.chdtrc(k - 1, chisq)
|
| 168 |
+
|
| 169 |
+
|
| 170 |
+
def chi2(X, y):
|
| 171 |
+
"""Compute chi-squared stats between each non-negative feature and class.
|
| 172 |
+
|
| 173 |
+
This score can be used to select the `n_features` features with the
|
| 174 |
+
highest values for the test chi-squared statistic from X, which must
|
| 175 |
+
contain only **non-negative features** such as booleans or frequencies
|
| 176 |
+
(e.g., term counts in document classification), relative to the classes.
|
| 177 |
+
|
| 178 |
+
Recall that the chi-square test measures dependence between stochastic
|
| 179 |
+
variables, so using this function "weeds out" the features that are the
|
| 180 |
+
most likely to be independent of class and therefore irrelevant for
|
| 181 |
+
classification.
|
| 182 |
+
|
| 183 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 184 |
+
|
| 185 |
+
Parameters
|
| 186 |
+
----------
|
| 187 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 188 |
+
Sample vectors.
|
| 189 |
+
|
| 190 |
+
y : array-like of shape (n_samples,)
|
| 191 |
+
Target vector (class labels).
|
| 192 |
+
|
| 193 |
+
Returns
|
| 194 |
+
-------
|
| 195 |
+
chi2 : ndarray of shape (n_features,)
|
| 196 |
+
Chi2 statistics for each feature.
|
| 197 |
+
|
| 198 |
+
p_values : ndarray of shape (n_features,)
|
| 199 |
+
P-values for each feature.
|
| 200 |
+
|
| 201 |
+
See Also
|
| 202 |
+
--------
|
| 203 |
+
f_classif : ANOVA F-value between label/feature for classification tasks.
|
| 204 |
+
f_regression : F-value between label/feature for regression tasks.
|
| 205 |
+
|
| 206 |
+
Notes
|
| 207 |
+
-----
|
| 208 |
+
Complexity of this algorithm is O(n_classes * n_features).
|
| 209 |
+
"""
|
| 210 |
+
|
| 211 |
+
# XXX: we might want to do some of the following in logspace instead for
|
| 212 |
+
# numerical stability.
|
| 213 |
+
# Converting X to float allows getting better performance for the
|
| 214 |
+
# safe_sparse_dot call made below.
|
| 215 |
+
X = check_array(X, accept_sparse="csr", dtype=(np.float64, np.float32))
|
| 216 |
+
if np.any((X.data if issparse(X) else X) < 0):
|
| 217 |
+
raise ValueError("Input X must be non-negative.")
|
| 218 |
+
|
| 219 |
+
# Use a sparse representation for Y by default to reduce memory usage when
|
| 220 |
+
# y has many unique classes.
|
| 221 |
+
Y = LabelBinarizer(sparse_output=True).fit_transform(y)
|
| 222 |
+
if Y.shape[1] == 1:
|
| 223 |
+
Y = Y.toarray()
|
| 224 |
+
Y = np.append(1 - Y, Y, axis=1)
|
| 225 |
+
|
| 226 |
+
observed = safe_sparse_dot(Y.T, X) # n_classes * n_features
|
| 227 |
+
|
| 228 |
+
if issparse(observed):
|
| 229 |
+
# convert back to a dense array before calling _chisquare
|
| 230 |
+
# XXX: could _chisquare be reimplement to accept sparse matrices for
|
| 231 |
+
# cases where both n_classes and n_features are large (and X is
|
| 232 |
+
# sparse)?
|
| 233 |
+
observed = observed.toarray()
|
| 234 |
+
|
| 235 |
+
feature_count = X.sum(axis=0).reshape(1, -1)
|
| 236 |
+
class_prob = Y.mean(axis=0).reshape(1, -1)
|
| 237 |
+
expected = np.dot(class_prob.T, feature_count)
|
| 238 |
+
|
| 239 |
+
return _chisquare(observed, expected)
|
| 240 |
+
|
| 241 |
+
|
| 242 |
+
def r_regression(X, y, *, center=True, force_finite=True):
|
| 243 |
+
"""Compute Pearson's r for each features and the target.
|
| 244 |
+
|
| 245 |
+
Pearson's r is also known as the Pearson correlation coefficient.
|
| 246 |
+
|
| 247 |
+
Linear model for testing the individual effect of each of many regressors.
|
| 248 |
+
This is a scoring function to be used in a feature selection procedure, not
|
| 249 |
+
a free standing feature selection procedure.
|
| 250 |
+
|
| 251 |
+
The cross correlation between each regressor and the target is computed
|
| 252 |
+
as::
|
| 253 |
+
|
| 254 |
+
E[(X[:, i] - mean(X[:, i])) * (y - mean(y))] / (std(X[:, i]) * std(y))
|
| 255 |
+
|
| 256 |
+
For more on usage see the :ref:`User Guide <univariate_feature_selection>`.
|
| 257 |
+
|
| 258 |
+
.. versionadded:: 1.0
|
| 259 |
+
|
| 260 |
+
Parameters
|
| 261 |
+
----------
|
| 262 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 263 |
+
The data matrix.
|
| 264 |
+
|
| 265 |
+
y : array-like of shape (n_samples,)
|
| 266 |
+
The target vector.
|
| 267 |
+
|
| 268 |
+
center : bool, default=True
|
| 269 |
+
Whether or not to center the data matrix `X` and the target vector `y`.
|
| 270 |
+
By default, `X` and `y` will be centered.
|
| 271 |
+
|
| 272 |
+
force_finite : bool, default=True
|
| 273 |
+
Whether or not to force the Pearson's R correlation to be finite.
|
| 274 |
+
In the particular case where some features in `X` or the target `y`
|
| 275 |
+
are constant, the Pearson's R correlation is not defined. When
|
| 276 |
+
`force_finite=False`, a correlation of `np.nan` is returned to
|
| 277 |
+
acknowledge this case. When `force_finite=True`, this value will be
|
| 278 |
+
forced to a minimal correlation of `0.0`.
|
| 279 |
+
|
| 280 |
+
.. versionadded:: 1.1
|
| 281 |
+
|
| 282 |
+
Returns
|
| 283 |
+
-------
|
| 284 |
+
correlation_coefficient : ndarray of shape (n_features,)
|
| 285 |
+
Pearson's R correlation coefficients of features.
|
| 286 |
+
|
| 287 |
+
See Also
|
| 288 |
+
--------
|
| 289 |
+
f_regression: Univariate linear regression tests returning f-statistic
|
| 290 |
+
and p-values.
|
| 291 |
+
mutual_info_regression: Mutual information for a continuous target.
|
| 292 |
+
f_classif: ANOVA F-value between label/feature for classification tasks.
|
| 293 |
+
chi2: Chi-squared stats of non-negative features for classification tasks.
|
| 294 |
+
"""
|
| 295 |
+
X, y = check_X_y(X, y, accept_sparse=["csr", "csc", "coo"], dtype=np.float64)
|
| 296 |
+
n_samples = X.shape[0]
|
| 297 |
+
|
| 298 |
+
# Compute centered values
|
| 299 |
+
# Note that E[(x - mean(x))*(y - mean(y))] = E[x*(y - mean(y))], so we
|
| 300 |
+
# need not center X
|
| 301 |
+
if center:
|
| 302 |
+
y = y - np.mean(y)
|
| 303 |
+
if issparse(X):
|
| 304 |
+
X_means = X.mean(axis=0).getA1()
|
| 305 |
+
else:
|
| 306 |
+
X_means = X.mean(axis=0)
|
| 307 |
+
# Compute the scaled standard deviations via moments
|
| 308 |
+
X_norms = np.sqrt(row_norms(X.T, squared=True) - n_samples * X_means**2)
|
| 309 |
+
else:
|
| 310 |
+
X_norms = row_norms(X.T)
|
| 311 |
+
|
| 312 |
+
correlation_coefficient = safe_sparse_dot(y, X)
|
| 313 |
+
with np.errstate(divide="ignore", invalid="ignore"):
|
| 314 |
+
correlation_coefficient /= X_norms
|
| 315 |
+
correlation_coefficient /= np.linalg.norm(y)
|
| 316 |
+
|
| 317 |
+
if force_finite and not np.isfinite(correlation_coefficient).all():
|
| 318 |
+
# case where the target or some features are constant
|
| 319 |
+
# the correlation coefficient(s) is/are set to the minimum (i.e. 0.0)
|
| 320 |
+
nan_mask = np.isnan(correlation_coefficient)
|
| 321 |
+
correlation_coefficient[nan_mask] = 0.0
|
| 322 |
+
return correlation_coefficient
|
| 323 |
+
|
| 324 |
+
|
| 325 |
+
def f_regression(X, y, *, center=True, force_finite=True):
|
| 326 |
+
"""Univariate linear regression tests returning F-statistic and p-values.
|
| 327 |
+
|
| 328 |
+
Quick linear model for testing the effect of a single regressor,
|
| 329 |
+
sequentially for many regressors.
|
| 330 |
+
|
| 331 |
+
This is done in 2 steps:
|
| 332 |
+
|
| 333 |
+
1. The cross correlation between each regressor and the target is computed
|
| 334 |
+
using :func:`r_regression` as::
|
| 335 |
+
|
| 336 |
+
E[(X[:, i] - mean(X[:, i])) * (y - mean(y))] / (std(X[:, i]) * std(y))
|
| 337 |
+
|
| 338 |
+
2. It is converted to an F score and then to a p-value.
|
| 339 |
+
|
| 340 |
+
:func:`f_regression` is derived from :func:`r_regression` and will rank
|
| 341 |
+
features in the same order if all the features are positively correlated
|
| 342 |
+
with the target.
|
| 343 |
+
|
| 344 |
+
Note however that contrary to :func:`f_regression`, :func:`r_regression`
|
| 345 |
+
values lie in [-1, 1] and can thus be negative. :func:`f_regression` is
|
| 346 |
+
therefore recommended as a feature selection criterion to identify
|
| 347 |
+
potentially predictive feature for a downstream classifier, irrespective of
|
| 348 |
+
the sign of the association with the target variable.
|
| 349 |
+
|
| 350 |
+
Furthermore :func:`f_regression` returns p-values while
|
| 351 |
+
:func:`r_regression` does not.
|
| 352 |
+
|
| 353 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 354 |
+
|
| 355 |
+
Parameters
|
| 356 |
+
----------
|
| 357 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 358 |
+
The data matrix.
|
| 359 |
+
|
| 360 |
+
y : array-like of shape (n_samples,)
|
| 361 |
+
The target vector.
|
| 362 |
+
|
| 363 |
+
center : bool, default=True
|
| 364 |
+
Whether or not to center the data matrix `X` and the target vector `y`.
|
| 365 |
+
By default, `X` and `y` will be centered.
|
| 366 |
+
|
| 367 |
+
force_finite : bool, default=True
|
| 368 |
+
Whether or not to force the F-statistics and associated p-values to
|
| 369 |
+
be finite. There are two cases where the F-statistic is expected to not
|
| 370 |
+
be finite:
|
| 371 |
+
|
| 372 |
+
- when the target `y` or some features in `X` are constant. In this
|
| 373 |
+
case, the Pearson's R correlation is not defined leading to obtain
|
| 374 |
+
`np.nan` values in the F-statistic and p-value. When
|
| 375 |
+
`force_finite=True`, the F-statistic is set to `0.0` and the
|
| 376 |
+
associated p-value is set to `1.0`.
|
| 377 |
+
- when a feature in `X` is perfectly correlated (or
|
| 378 |
+
anti-correlated) with the target `y`. In this case, the F-statistic
|
| 379 |
+
is expected to be `np.inf`. When `force_finite=True`, the F-statistic
|
| 380 |
+
is set to `np.finfo(dtype).max` and the associated p-value is set to
|
| 381 |
+
`0.0`.
|
| 382 |
+
|
| 383 |
+
.. versionadded:: 1.1
|
| 384 |
+
|
| 385 |
+
Returns
|
| 386 |
+
-------
|
| 387 |
+
f_statistic : ndarray of shape (n_features,)
|
| 388 |
+
F-statistic for each feature.
|
| 389 |
+
|
| 390 |
+
p_values : ndarray of shape (n_features,)
|
| 391 |
+
P-values associated with the F-statistic.
|
| 392 |
+
|
| 393 |
+
See Also
|
| 394 |
+
--------
|
| 395 |
+
r_regression: Pearson's R between label/feature for regression tasks.
|
| 396 |
+
f_classif: ANOVA F-value between label/feature for classification tasks.
|
| 397 |
+
chi2: Chi-squared stats of non-negative features for classification tasks.
|
| 398 |
+
SelectKBest: Select features based on the k highest scores.
|
| 399 |
+
SelectFpr: Select features based on a false positive rate test.
|
| 400 |
+
SelectFdr: Select features based on an estimated false discovery rate.
|
| 401 |
+
SelectFwe: Select features based on family-wise error rate.
|
| 402 |
+
SelectPercentile: Select features based on percentile of the highest
|
| 403 |
+
scores.
|
| 404 |
+
"""
|
| 405 |
+
correlation_coefficient = r_regression(
|
| 406 |
+
X, y, center=center, force_finite=force_finite
|
| 407 |
+
)
|
| 408 |
+
deg_of_freedom = y.size - (2 if center else 1)
|
| 409 |
+
|
| 410 |
+
corr_coef_squared = correlation_coefficient**2
|
| 411 |
+
|
| 412 |
+
with np.errstate(divide="ignore", invalid="ignore"):
|
| 413 |
+
f_statistic = corr_coef_squared / (1 - corr_coef_squared) * deg_of_freedom
|
| 414 |
+
p_values = stats.f.sf(f_statistic, 1, deg_of_freedom)
|
| 415 |
+
|
| 416 |
+
if force_finite and not np.isfinite(f_statistic).all():
|
| 417 |
+
# case where there is a perfect (anti-)correlation
|
| 418 |
+
# f-statistics can be set to the maximum and p-values to zero
|
| 419 |
+
mask_inf = np.isinf(f_statistic)
|
| 420 |
+
f_statistic[mask_inf] = np.finfo(f_statistic.dtype).max
|
| 421 |
+
# case where the target or some features are constant
|
| 422 |
+
# f-statistics would be minimum and thus p-values large
|
| 423 |
+
mask_nan = np.isnan(f_statistic)
|
| 424 |
+
f_statistic[mask_nan] = 0.0
|
| 425 |
+
p_values[mask_nan] = 1.0
|
| 426 |
+
return f_statistic, p_values
|
| 427 |
+
|
| 428 |
+
|
| 429 |
+
######################################################################
|
| 430 |
+
# Base classes
|
| 431 |
+
|
| 432 |
+
|
| 433 |
+
class _BaseFilter(SelectorMixin, BaseEstimator):
|
| 434 |
+
"""Initialize the univariate feature selection.
|
| 435 |
+
|
| 436 |
+
Parameters
|
| 437 |
+
----------
|
| 438 |
+
score_func : callable
|
| 439 |
+
Function taking two arrays X and y, and returning a pair of arrays
|
| 440 |
+
(scores, pvalues) or a single array with scores.
|
| 441 |
+
"""
|
| 442 |
+
|
| 443 |
+
_parameter_constraints: dict = {"score_func": [callable]}
|
| 444 |
+
|
| 445 |
+
def __init__(self, score_func):
|
| 446 |
+
self.score_func = score_func
|
| 447 |
+
|
| 448 |
+
def fit(self, X, y):
|
| 449 |
+
"""Run score function on (X, y) and get the appropriate features.
|
| 450 |
+
|
| 451 |
+
Parameters
|
| 452 |
+
----------
|
| 453 |
+
X : array-like of shape (n_samples, n_features)
|
| 454 |
+
The training input samples.
|
| 455 |
+
|
| 456 |
+
y : array-like of shape (n_samples,)
|
| 457 |
+
The target values (class labels in classification, real numbers in
|
| 458 |
+
regression).
|
| 459 |
+
|
| 460 |
+
Returns
|
| 461 |
+
-------
|
| 462 |
+
self : object
|
| 463 |
+
Returns the instance itself.
|
| 464 |
+
"""
|
| 465 |
+
self._validate_params()
|
| 466 |
+
|
| 467 |
+
X, y = self._validate_data(
|
| 468 |
+
X, y, accept_sparse=["csr", "csc"], multi_output=True
|
| 469 |
+
)
|
| 470 |
+
|
| 471 |
+
self._check_params(X, y)
|
| 472 |
+
score_func_ret = self.score_func(X, y)
|
| 473 |
+
if isinstance(score_func_ret, (list, tuple)):
|
| 474 |
+
self.scores_, self.pvalues_ = score_func_ret
|
| 475 |
+
self.pvalues_ = np.asarray(self.pvalues_)
|
| 476 |
+
else:
|
| 477 |
+
self.scores_ = score_func_ret
|
| 478 |
+
self.pvalues_ = None
|
| 479 |
+
|
| 480 |
+
self.scores_ = np.asarray(self.scores_)
|
| 481 |
+
|
| 482 |
+
return self
|
| 483 |
+
|
| 484 |
+
def _check_params(self, X, y):
|
| 485 |
+
pass
|
| 486 |
+
|
| 487 |
+
def _more_tags(self):
|
| 488 |
+
return {"requires_y": True}
|
| 489 |
+
|
| 490 |
+
|
| 491 |
+
######################################################################
|
| 492 |
+
# Specific filters
|
| 493 |
+
######################################################################
|
| 494 |
+
class SelectPercentile(_BaseFilter):
|
| 495 |
+
"""Select features according to a percentile of the highest scores.
|
| 496 |
+
|
| 497 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 498 |
+
|
| 499 |
+
Parameters
|
| 500 |
+
----------
|
| 501 |
+
score_func : callable, default=f_classif
|
| 502 |
+
Function taking two arrays X and y, and returning a pair of arrays
|
| 503 |
+
(scores, pvalues) or a single array with scores.
|
| 504 |
+
Default is f_classif (see below "See Also"). The default function only
|
| 505 |
+
works with classification tasks.
|
| 506 |
+
|
| 507 |
+
.. versionadded:: 0.18
|
| 508 |
+
|
| 509 |
+
percentile : int, default=10
|
| 510 |
+
Percent of features to keep.
|
| 511 |
+
|
| 512 |
+
Attributes
|
| 513 |
+
----------
|
| 514 |
+
scores_ : array-like of shape (n_features,)
|
| 515 |
+
Scores of features.
|
| 516 |
+
|
| 517 |
+
pvalues_ : array-like of shape (n_features,)
|
| 518 |
+
p-values of feature scores, None if `score_func` returned only scores.
|
| 519 |
+
|
| 520 |
+
n_features_in_ : int
|
| 521 |
+
Number of features seen during :term:`fit`.
|
| 522 |
+
|
| 523 |
+
.. versionadded:: 0.24
|
| 524 |
+
|
| 525 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 526 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 527 |
+
has feature names that are all strings.
|
| 528 |
+
|
| 529 |
+
.. versionadded:: 1.0
|
| 530 |
+
|
| 531 |
+
See Also
|
| 532 |
+
--------
|
| 533 |
+
f_classif : ANOVA F-value between label/feature for classification tasks.
|
| 534 |
+
mutual_info_classif : Mutual information for a discrete target.
|
| 535 |
+
chi2 : Chi-squared stats of non-negative features for classification tasks.
|
| 536 |
+
f_regression : F-value between label/feature for regression tasks.
|
| 537 |
+
mutual_info_regression : Mutual information for a continuous target.
|
| 538 |
+
SelectKBest : Select features based on the k highest scores.
|
| 539 |
+
SelectFpr : Select features based on a false positive rate test.
|
| 540 |
+
SelectFdr : Select features based on an estimated false discovery rate.
|
| 541 |
+
SelectFwe : Select features based on family-wise error rate.
|
| 542 |
+
GenericUnivariateSelect : Univariate feature selector with configurable
|
| 543 |
+
mode.
|
| 544 |
+
|
| 545 |
+
Notes
|
| 546 |
+
-----
|
| 547 |
+
Ties between features with equal scores will be broken in an unspecified
|
| 548 |
+
way.
|
| 549 |
+
|
| 550 |
+
Examples
|
| 551 |
+
--------
|
| 552 |
+
>>> from sklearn.datasets import load_digits
|
| 553 |
+
>>> from sklearn.feature_selection import SelectPercentile, chi2
|
| 554 |
+
>>> X, y = load_digits(return_X_y=True)
|
| 555 |
+
>>> X.shape
|
| 556 |
+
(1797, 64)
|
| 557 |
+
>>> X_new = SelectPercentile(chi2, percentile=10).fit_transform(X, y)
|
| 558 |
+
>>> X_new.shape
|
| 559 |
+
(1797, 7)
|
| 560 |
+
"""
|
| 561 |
+
|
| 562 |
+
_parameter_constraints: dict = {
|
| 563 |
+
**_BaseFilter._parameter_constraints,
|
| 564 |
+
"percentile": [Interval(Real, 0, 100, closed="both")],
|
| 565 |
+
}
|
| 566 |
+
|
| 567 |
+
def __init__(self, score_func=f_classif, *, percentile=10):
|
| 568 |
+
super().__init__(score_func=score_func)
|
| 569 |
+
self.percentile = percentile
|
| 570 |
+
|
| 571 |
+
def _get_support_mask(self):
|
| 572 |
+
check_is_fitted(self)
|
| 573 |
+
|
| 574 |
+
# Cater for NaNs
|
| 575 |
+
if self.percentile == 100:
|
| 576 |
+
return np.ones(len(self.scores_), dtype=bool)
|
| 577 |
+
elif self.percentile == 0:
|
| 578 |
+
return np.zeros(len(self.scores_), dtype=bool)
|
| 579 |
+
|
| 580 |
+
scores = _clean_nans(self.scores_)
|
| 581 |
+
threshold = np.percentile(scores, 100 - self.percentile)
|
| 582 |
+
mask = scores > threshold
|
| 583 |
+
ties = np.where(scores == threshold)[0]
|
| 584 |
+
if len(ties):
|
| 585 |
+
max_feats = int(len(scores) * self.percentile / 100)
|
| 586 |
+
kept_ties = ties[: max_feats - mask.sum()]
|
| 587 |
+
mask[kept_ties] = True
|
| 588 |
+
return mask
|
| 589 |
+
|
| 590 |
+
|
| 591 |
+
class SelectKBest(_BaseFilter):
|
| 592 |
+
"""Select features according to the k highest scores.
|
| 593 |
+
|
| 594 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 595 |
+
|
| 596 |
+
Parameters
|
| 597 |
+
----------
|
| 598 |
+
score_func : callable, default=f_classif
|
| 599 |
+
Function taking two arrays X and y, and returning a pair of arrays
|
| 600 |
+
(scores, pvalues) or a single array with scores.
|
| 601 |
+
Default is f_classif (see below "See Also"). The default function only
|
| 602 |
+
works with classification tasks.
|
| 603 |
+
|
| 604 |
+
.. versionadded:: 0.18
|
| 605 |
+
|
| 606 |
+
k : int or "all", default=10
|
| 607 |
+
Number of top features to select.
|
| 608 |
+
The "all" option bypasses selection, for use in a parameter search.
|
| 609 |
+
|
| 610 |
+
Attributes
|
| 611 |
+
----------
|
| 612 |
+
scores_ : array-like of shape (n_features,)
|
| 613 |
+
Scores of features.
|
| 614 |
+
|
| 615 |
+
pvalues_ : array-like of shape (n_features,)
|
| 616 |
+
p-values of feature scores, None if `score_func` returned only scores.
|
| 617 |
+
|
| 618 |
+
n_features_in_ : int
|
| 619 |
+
Number of features seen during :term:`fit`.
|
| 620 |
+
|
| 621 |
+
.. versionadded:: 0.24
|
| 622 |
+
|
| 623 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 624 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 625 |
+
has feature names that are all strings.
|
| 626 |
+
|
| 627 |
+
.. versionadded:: 1.0
|
| 628 |
+
|
| 629 |
+
See Also
|
| 630 |
+
--------
|
| 631 |
+
f_classif: ANOVA F-value between label/feature for classification tasks.
|
| 632 |
+
mutual_info_classif: Mutual information for a discrete target.
|
| 633 |
+
chi2: Chi-squared stats of non-negative features for classification tasks.
|
| 634 |
+
f_regression: F-value between label/feature for regression tasks.
|
| 635 |
+
mutual_info_regression: Mutual information for a continuous target.
|
| 636 |
+
SelectPercentile: Select features based on percentile of the highest
|
| 637 |
+
scores.
|
| 638 |
+
SelectFpr : Select features based on a false positive rate test.
|
| 639 |
+
SelectFdr : Select features based on an estimated false discovery rate.
|
| 640 |
+
SelectFwe : Select features based on family-wise error rate.
|
| 641 |
+
GenericUnivariateSelect : Univariate feature selector with configurable
|
| 642 |
+
mode.
|
| 643 |
+
|
| 644 |
+
Notes
|
| 645 |
+
-----
|
| 646 |
+
Ties between features with equal scores will be broken in an unspecified
|
| 647 |
+
way.
|
| 648 |
+
|
| 649 |
+
Examples
|
| 650 |
+
--------
|
| 651 |
+
>>> from sklearn.datasets import load_digits
|
| 652 |
+
>>> from sklearn.feature_selection import SelectKBest, chi2
|
| 653 |
+
>>> X, y = load_digits(return_X_y=True)
|
| 654 |
+
>>> X.shape
|
| 655 |
+
(1797, 64)
|
| 656 |
+
>>> X_new = SelectKBest(chi2, k=20).fit_transform(X, y)
|
| 657 |
+
>>> X_new.shape
|
| 658 |
+
(1797, 20)
|
| 659 |
+
"""
|
| 660 |
+
|
| 661 |
+
_parameter_constraints: dict = {
|
| 662 |
+
**_BaseFilter._parameter_constraints,
|
| 663 |
+
"k": [StrOptions({"all"}), Interval(Integral, 0, None, closed="left")],
|
| 664 |
+
}
|
| 665 |
+
|
| 666 |
+
def __init__(self, score_func=f_classif, *, k=10):
|
| 667 |
+
super().__init__(score_func=score_func)
|
| 668 |
+
self.k = k
|
| 669 |
+
|
| 670 |
+
def _check_params(self, X, y):
|
| 671 |
+
if not isinstance(self.k, str) and self.k > X.shape[1]:
|
| 672 |
+
raise ValueError(
|
| 673 |
+
f"k should be <= n_features = {X.shape[1]}; "
|
| 674 |
+
f"got {self.k}. Use k='all' to return all features."
|
| 675 |
+
)
|
| 676 |
+
|
| 677 |
+
def _get_support_mask(self):
|
| 678 |
+
check_is_fitted(self)
|
| 679 |
+
|
| 680 |
+
if self.k == "all":
|
| 681 |
+
return np.ones(self.scores_.shape, dtype=bool)
|
| 682 |
+
elif self.k == 0:
|
| 683 |
+
return np.zeros(self.scores_.shape, dtype=bool)
|
| 684 |
+
else:
|
| 685 |
+
scores = _clean_nans(self.scores_)
|
| 686 |
+
mask = np.zeros(scores.shape, dtype=bool)
|
| 687 |
+
|
| 688 |
+
# Request a stable sort. Mergesort takes more memory (~40MB per
|
| 689 |
+
# megafeature on x86-64).
|
| 690 |
+
mask[np.argsort(scores, kind="mergesort")[-self.k :]] = 1
|
| 691 |
+
return mask
|
| 692 |
+
|
| 693 |
+
|
| 694 |
+
class SelectFpr(_BaseFilter):
|
| 695 |
+
"""Filter: Select the pvalues below alpha based on a FPR test.
|
| 696 |
+
|
| 697 |
+
FPR test stands for False Positive Rate test. It controls the total
|
| 698 |
+
amount of false detections.
|
| 699 |
+
|
| 700 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 701 |
+
|
| 702 |
+
Parameters
|
| 703 |
+
----------
|
| 704 |
+
score_func : callable, default=f_classif
|
| 705 |
+
Function taking two arrays X and y, and returning a pair of arrays
|
| 706 |
+
(scores, pvalues).
|
| 707 |
+
Default is f_classif (see below "See Also"). The default function only
|
| 708 |
+
works with classification tasks.
|
| 709 |
+
|
| 710 |
+
alpha : float, default=5e-2
|
| 711 |
+
Features with p-values less than `alpha` are selected.
|
| 712 |
+
|
| 713 |
+
Attributes
|
| 714 |
+
----------
|
| 715 |
+
scores_ : array-like of shape (n_features,)
|
| 716 |
+
Scores of features.
|
| 717 |
+
|
| 718 |
+
pvalues_ : array-like of shape (n_features,)
|
| 719 |
+
p-values of feature scores.
|
| 720 |
+
|
| 721 |
+
n_features_in_ : int
|
| 722 |
+
Number of features seen during :term:`fit`.
|
| 723 |
+
|
| 724 |
+
.. versionadded:: 0.24
|
| 725 |
+
|
| 726 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 727 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 728 |
+
has feature names that are all strings.
|
| 729 |
+
|
| 730 |
+
.. versionadded:: 1.0
|
| 731 |
+
|
| 732 |
+
See Also
|
| 733 |
+
--------
|
| 734 |
+
f_classif : ANOVA F-value between label/feature for classification tasks.
|
| 735 |
+
chi2 : Chi-squared stats of non-negative features for classification tasks.
|
| 736 |
+
mutual_info_classif: Mutual information for a discrete target.
|
| 737 |
+
f_regression : F-value between label/feature for regression tasks.
|
| 738 |
+
mutual_info_regression : Mutual information for a continuous target.
|
| 739 |
+
SelectPercentile : Select features based on percentile of the highest
|
| 740 |
+
scores.
|
| 741 |
+
SelectKBest : Select features based on the k highest scores.
|
| 742 |
+
SelectFdr : Select features based on an estimated false discovery rate.
|
| 743 |
+
SelectFwe : Select features based on family-wise error rate.
|
| 744 |
+
GenericUnivariateSelect : Univariate feature selector with configurable
|
| 745 |
+
mode.
|
| 746 |
+
|
| 747 |
+
Examples
|
| 748 |
+
--------
|
| 749 |
+
>>> from sklearn.datasets import load_breast_cancer
|
| 750 |
+
>>> from sklearn.feature_selection import SelectFpr, chi2
|
| 751 |
+
>>> X, y = load_breast_cancer(return_X_y=True)
|
| 752 |
+
>>> X.shape
|
| 753 |
+
(569, 30)
|
| 754 |
+
>>> X_new = SelectFpr(chi2, alpha=0.01).fit_transform(X, y)
|
| 755 |
+
>>> X_new.shape
|
| 756 |
+
(569, 16)
|
| 757 |
+
"""
|
| 758 |
+
|
| 759 |
+
_parameter_constraints: dict = {
|
| 760 |
+
**_BaseFilter._parameter_constraints,
|
| 761 |
+
"alpha": [Interval(Real, 0, 1, closed="both")],
|
| 762 |
+
}
|
| 763 |
+
|
| 764 |
+
def __init__(self, score_func=f_classif, *, alpha=5e-2):
|
| 765 |
+
super().__init__(score_func=score_func)
|
| 766 |
+
self.alpha = alpha
|
| 767 |
+
|
| 768 |
+
def _get_support_mask(self):
|
| 769 |
+
check_is_fitted(self)
|
| 770 |
+
|
| 771 |
+
return self.pvalues_ < self.alpha
|
| 772 |
+
|
| 773 |
+
|
| 774 |
+
class SelectFdr(_BaseFilter):
|
| 775 |
+
"""Filter: Select the p-values for an estimated false discovery rate.
|
| 776 |
+
|
| 777 |
+
This uses the Benjamini-Hochberg procedure. ``alpha`` is an upper bound
|
| 778 |
+
on the expected false discovery rate.
|
| 779 |
+
|
| 780 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 781 |
+
|
| 782 |
+
Parameters
|
| 783 |
+
----------
|
| 784 |
+
score_func : callable, default=f_classif
|
| 785 |
+
Function taking two arrays X and y, and returning a pair of arrays
|
| 786 |
+
(scores, pvalues).
|
| 787 |
+
Default is f_classif (see below "See Also"). The default function only
|
| 788 |
+
works with classification tasks.
|
| 789 |
+
|
| 790 |
+
alpha : float, default=5e-2
|
| 791 |
+
The highest uncorrected p-value for features to keep.
|
| 792 |
+
|
| 793 |
+
Attributes
|
| 794 |
+
----------
|
| 795 |
+
scores_ : array-like of shape (n_features,)
|
| 796 |
+
Scores of features.
|
| 797 |
+
|
| 798 |
+
pvalues_ : array-like of shape (n_features,)
|
| 799 |
+
p-values of feature scores.
|
| 800 |
+
|
| 801 |
+
n_features_in_ : int
|
| 802 |
+
Number of features seen during :term:`fit`.
|
| 803 |
+
|
| 804 |
+
.. versionadded:: 0.24
|
| 805 |
+
|
| 806 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 807 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 808 |
+
has feature names that are all strings.
|
| 809 |
+
|
| 810 |
+
.. versionadded:: 1.0
|
| 811 |
+
|
| 812 |
+
See Also
|
| 813 |
+
--------
|
| 814 |
+
f_classif : ANOVA F-value between label/feature for classification tasks.
|
| 815 |
+
mutual_info_classif : Mutual information for a discrete target.
|
| 816 |
+
chi2 : Chi-squared stats of non-negative features for classification tasks.
|
| 817 |
+
f_regression : F-value between label/feature for regression tasks.
|
| 818 |
+
mutual_info_regression : Mutual information for a continuous target.
|
| 819 |
+
SelectPercentile : Select features based on percentile of the highest
|
| 820 |
+
scores.
|
| 821 |
+
SelectKBest : Select features based on the k highest scores.
|
| 822 |
+
SelectFpr : Select features based on a false positive rate test.
|
| 823 |
+
SelectFwe : Select features based on family-wise error rate.
|
| 824 |
+
GenericUnivariateSelect : Univariate feature selector with configurable
|
| 825 |
+
mode.
|
| 826 |
+
|
| 827 |
+
References
|
| 828 |
+
----------
|
| 829 |
+
https://en.wikipedia.org/wiki/False_discovery_rate
|
| 830 |
+
|
| 831 |
+
Examples
|
| 832 |
+
--------
|
| 833 |
+
>>> from sklearn.datasets import load_breast_cancer
|
| 834 |
+
>>> from sklearn.feature_selection import SelectFdr, chi2
|
| 835 |
+
>>> X, y = load_breast_cancer(return_X_y=True)
|
| 836 |
+
>>> X.shape
|
| 837 |
+
(569, 30)
|
| 838 |
+
>>> X_new = SelectFdr(chi2, alpha=0.01).fit_transform(X, y)
|
| 839 |
+
>>> X_new.shape
|
| 840 |
+
(569, 16)
|
| 841 |
+
"""
|
| 842 |
+
|
| 843 |
+
_parameter_constraints: dict = {
|
| 844 |
+
**_BaseFilter._parameter_constraints,
|
| 845 |
+
"alpha": [Interval(Real, 0, 1, closed="both")],
|
| 846 |
+
}
|
| 847 |
+
|
| 848 |
+
def __init__(self, score_func=f_classif, *, alpha=5e-2):
|
| 849 |
+
super().__init__(score_func=score_func)
|
| 850 |
+
self.alpha = alpha
|
| 851 |
+
|
| 852 |
+
def _get_support_mask(self):
|
| 853 |
+
check_is_fitted(self)
|
| 854 |
+
|
| 855 |
+
n_features = len(self.pvalues_)
|
| 856 |
+
sv = np.sort(self.pvalues_)
|
| 857 |
+
selected = sv[
|
| 858 |
+
sv <= float(self.alpha) / n_features * np.arange(1, n_features + 1)
|
| 859 |
+
]
|
| 860 |
+
if selected.size == 0:
|
| 861 |
+
return np.zeros_like(self.pvalues_, dtype=bool)
|
| 862 |
+
return self.pvalues_ <= selected.max()
|
| 863 |
+
|
| 864 |
+
|
| 865 |
+
class SelectFwe(_BaseFilter):
|
| 866 |
+
"""Filter: Select the p-values corresponding to Family-wise error rate.
|
| 867 |
+
|
| 868 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 869 |
+
|
| 870 |
+
Parameters
|
| 871 |
+
----------
|
| 872 |
+
score_func : callable, default=f_classif
|
| 873 |
+
Function taking two arrays X and y, and returning a pair of arrays
|
| 874 |
+
(scores, pvalues).
|
| 875 |
+
Default is f_classif (see below "See Also"). The default function only
|
| 876 |
+
works with classification tasks.
|
| 877 |
+
|
| 878 |
+
alpha : float, default=5e-2
|
| 879 |
+
The highest uncorrected p-value for features to keep.
|
| 880 |
+
|
| 881 |
+
Attributes
|
| 882 |
+
----------
|
| 883 |
+
scores_ : array-like of shape (n_features,)
|
| 884 |
+
Scores of features.
|
| 885 |
+
|
| 886 |
+
pvalues_ : array-like of shape (n_features,)
|
| 887 |
+
p-values of feature scores.
|
| 888 |
+
|
| 889 |
+
n_features_in_ : int
|
| 890 |
+
Number of features seen during :term:`fit`.
|
| 891 |
+
|
| 892 |
+
.. versionadded:: 0.24
|
| 893 |
+
|
| 894 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 895 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 896 |
+
has feature names that are all strings.
|
| 897 |
+
|
| 898 |
+
.. versionadded:: 1.0
|
| 899 |
+
|
| 900 |
+
See Also
|
| 901 |
+
--------
|
| 902 |
+
f_classif : ANOVA F-value between label/feature for classification tasks.
|
| 903 |
+
chi2 : Chi-squared stats of non-negative features for classification tasks.
|
| 904 |
+
f_regression : F-value between label/feature for regression tasks.
|
| 905 |
+
SelectPercentile : Select features based on percentile of the highest
|
| 906 |
+
scores.
|
| 907 |
+
SelectKBest : Select features based on the k highest scores.
|
| 908 |
+
SelectFpr : Select features based on a false positive rate test.
|
| 909 |
+
SelectFdr : Select features based on an estimated false discovery rate.
|
| 910 |
+
GenericUnivariateSelect : Univariate feature selector with configurable
|
| 911 |
+
mode.
|
| 912 |
+
|
| 913 |
+
Examples
|
| 914 |
+
--------
|
| 915 |
+
>>> from sklearn.datasets import load_breast_cancer
|
| 916 |
+
>>> from sklearn.feature_selection import SelectFwe, chi2
|
| 917 |
+
>>> X, y = load_breast_cancer(return_X_y=True)
|
| 918 |
+
>>> X.shape
|
| 919 |
+
(569, 30)
|
| 920 |
+
>>> X_new = SelectFwe(chi2, alpha=0.01).fit_transform(X, y)
|
| 921 |
+
>>> X_new.shape
|
| 922 |
+
(569, 15)
|
| 923 |
+
"""
|
| 924 |
+
|
| 925 |
+
_parameter_constraints: dict = {
|
| 926 |
+
**_BaseFilter._parameter_constraints,
|
| 927 |
+
"alpha": [Interval(Real, 0, 1, closed="both")],
|
| 928 |
+
}
|
| 929 |
+
|
| 930 |
+
def __init__(self, score_func=f_classif, *, alpha=5e-2):
|
| 931 |
+
super().__init__(score_func=score_func)
|
| 932 |
+
self.alpha = alpha
|
| 933 |
+
|
| 934 |
+
def _get_support_mask(self):
|
| 935 |
+
check_is_fitted(self)
|
| 936 |
+
|
| 937 |
+
return self.pvalues_ < self.alpha / len(self.pvalues_)
|
| 938 |
+
|
| 939 |
+
|
| 940 |
+
######################################################################
|
| 941 |
+
# Generic filter
|
| 942 |
+
######################################################################
|
| 943 |
+
|
| 944 |
+
# TODO this class should fit on either p-values or scores,
|
| 945 |
+
# depending on the mode.
|
| 946 |
+
class GenericUnivariateSelect(_BaseFilter):
|
| 947 |
+
"""Univariate feature selector with configurable strategy.
|
| 948 |
+
|
| 949 |
+
Read more in the :ref:`User Guide <univariate_feature_selection>`.
|
| 950 |
+
|
| 951 |
+
Parameters
|
| 952 |
+
----------
|
| 953 |
+
score_func : callable, default=f_classif
|
| 954 |
+
Function taking two arrays X and y, and returning a pair of arrays
|
| 955 |
+
(scores, pvalues). For modes 'percentile' or 'kbest' it can return
|
| 956 |
+
a single array scores.
|
| 957 |
+
|
| 958 |
+
mode : {'percentile', 'k_best', 'fpr', 'fdr', 'fwe'}, default='percentile'
|
| 959 |
+
Feature selection mode.
|
| 960 |
+
|
| 961 |
+
param : "all", float or int, default=1e-5
|
| 962 |
+
Parameter of the corresponding mode.
|
| 963 |
+
|
| 964 |
+
Attributes
|
| 965 |
+
----------
|
| 966 |
+
scores_ : array-like of shape (n_features,)
|
| 967 |
+
Scores of features.
|
| 968 |
+
|
| 969 |
+
pvalues_ : array-like of shape (n_features,)
|
| 970 |
+
p-values of feature scores, None if `score_func` returned scores only.
|
| 971 |
+
|
| 972 |
+
n_features_in_ : int
|
| 973 |
+
Number of features seen during :term:`fit`.
|
| 974 |
+
|
| 975 |
+
.. versionadded:: 0.24
|
| 976 |
+
|
| 977 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 978 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 979 |
+
has feature names that are all strings.
|
| 980 |
+
|
| 981 |
+
.. versionadded:: 1.0
|
| 982 |
+
|
| 983 |
+
See Also
|
| 984 |
+
--------
|
| 985 |
+
f_classif : ANOVA F-value between label/feature for classification tasks.
|
| 986 |
+
mutual_info_classif : Mutual information for a discrete target.
|
| 987 |
+
chi2 : Chi-squared stats of non-negative features for classification tasks.
|
| 988 |
+
f_regression : F-value between label/feature for regression tasks.
|
| 989 |
+
mutual_info_regression : Mutual information for a continuous target.
|
| 990 |
+
SelectPercentile : Select features based on percentile of the highest
|
| 991 |
+
scores.
|
| 992 |
+
SelectKBest : Select features based on the k highest scores.
|
| 993 |
+
SelectFpr : Select features based on a false positive rate test.
|
| 994 |
+
SelectFdr : Select features based on an estimated false discovery rate.
|
| 995 |
+
SelectFwe : Select features based on family-wise error rate.
|
| 996 |
+
|
| 997 |
+
Examples
|
| 998 |
+
--------
|
| 999 |
+
>>> from sklearn.datasets import load_breast_cancer
|
| 1000 |
+
>>> from sklearn.feature_selection import GenericUnivariateSelect, chi2
|
| 1001 |
+
>>> X, y = load_breast_cancer(return_X_y=True)
|
| 1002 |
+
>>> X.shape
|
| 1003 |
+
(569, 30)
|
| 1004 |
+
>>> transformer = GenericUnivariateSelect(chi2, mode='k_best', param=20)
|
| 1005 |
+
>>> X_new = transformer.fit_transform(X, y)
|
| 1006 |
+
>>> X_new.shape
|
| 1007 |
+
(569, 20)
|
| 1008 |
+
"""
|
| 1009 |
+
|
| 1010 |
+
_selection_modes: dict = {
|
| 1011 |
+
"percentile": SelectPercentile,
|
| 1012 |
+
"k_best": SelectKBest,
|
| 1013 |
+
"fpr": SelectFpr,
|
| 1014 |
+
"fdr": SelectFdr,
|
| 1015 |
+
"fwe": SelectFwe,
|
| 1016 |
+
}
|
| 1017 |
+
|
| 1018 |
+
_parameter_constraints: dict = {
|
| 1019 |
+
**_BaseFilter._parameter_constraints,
|
| 1020 |
+
"mode": [StrOptions(set(_selection_modes.keys()))],
|
| 1021 |
+
"param": [Interval(Real, 0, None, closed="left"), StrOptions({"all"})],
|
| 1022 |
+
}
|
| 1023 |
+
|
| 1024 |
+
def __init__(self, score_func=f_classif, *, mode="percentile", param=1e-5):
|
| 1025 |
+
super().__init__(score_func=score_func)
|
| 1026 |
+
self.mode = mode
|
| 1027 |
+
self.param = param
|
| 1028 |
+
|
| 1029 |
+
def _make_selector(self):
|
| 1030 |
+
selector = self._selection_modes[self.mode](score_func=self.score_func)
|
| 1031 |
+
|
| 1032 |
+
# Now perform some acrobatics to set the right named parameter in
|
| 1033 |
+
# the selector
|
| 1034 |
+
possible_params = selector._get_param_names()
|
| 1035 |
+
possible_params.remove("score_func")
|
| 1036 |
+
selector.set_params(**{possible_params[0]: self.param})
|
| 1037 |
+
|
| 1038 |
+
return selector
|
| 1039 |
+
|
| 1040 |
+
def _more_tags(self):
|
| 1041 |
+
return {"preserves_dtype": [np.float64, np.float32]}
|
| 1042 |
+
|
| 1043 |
+
def _check_params(self, X, y):
|
| 1044 |
+
self._make_selector()._check_params(X, y)
|
| 1045 |
+
|
| 1046 |
+
def _get_support_mask(self):
|
| 1047 |
+
check_is_fitted(self)
|
| 1048 |
+
|
| 1049 |
+
selector = self._make_selector()
|
| 1050 |
+
selector.pvalues_ = self.pvalues_
|
| 1051 |
+
selector.scores_ = self.scores_
|
| 1052 |
+
return selector._get_support_mask()
|
videollama2/lib/python3.10/site-packages/sklearn/feature_selection/_variance_threshold.py
ADDED
|
@@ -0,0 +1,135 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Author: Lars Buitinck
|
| 2 |
+
# License: 3-clause BSD
|
| 3 |
+
from numbers import Real
|
| 4 |
+
|
| 5 |
+
import numpy as np
|
| 6 |
+
from ..base import BaseEstimator
|
| 7 |
+
from ._base import SelectorMixin
|
| 8 |
+
from ..utils.sparsefuncs import mean_variance_axis, min_max_axis
|
| 9 |
+
from ..utils.validation import check_is_fitted
|
| 10 |
+
from ..utils._param_validation import Interval
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
class VarianceThreshold(SelectorMixin, BaseEstimator):
|
| 14 |
+
"""Feature selector that removes all low-variance features.
|
| 15 |
+
|
| 16 |
+
This feature selection algorithm looks only at the features (X), not the
|
| 17 |
+
desired outputs (y), and can thus be used for unsupervised learning.
|
| 18 |
+
|
| 19 |
+
Read more in the :ref:`User Guide <variance_threshold>`.
|
| 20 |
+
|
| 21 |
+
Parameters
|
| 22 |
+
----------
|
| 23 |
+
threshold : float, default=0
|
| 24 |
+
Features with a training-set variance lower than this threshold will
|
| 25 |
+
be removed. The default is to keep all features with non-zero variance,
|
| 26 |
+
i.e. remove the features that have the same value in all samples.
|
| 27 |
+
|
| 28 |
+
Attributes
|
| 29 |
+
----------
|
| 30 |
+
variances_ : array, shape (n_features,)
|
| 31 |
+
Variances of individual features.
|
| 32 |
+
|
| 33 |
+
n_features_in_ : int
|
| 34 |
+
Number of features seen during :term:`fit`.
|
| 35 |
+
|
| 36 |
+
.. versionadded:: 0.24
|
| 37 |
+
|
| 38 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 39 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 40 |
+
has feature names that are all strings.
|
| 41 |
+
|
| 42 |
+
.. versionadded:: 1.0
|
| 43 |
+
|
| 44 |
+
See Also
|
| 45 |
+
--------
|
| 46 |
+
SelectFromModel: Meta-transformer for selecting features based on
|
| 47 |
+
importance weights.
|
| 48 |
+
SelectPercentile : Select features according to a percentile of the highest
|
| 49 |
+
scores.
|
| 50 |
+
SequentialFeatureSelector : Transformer that performs Sequential Feature
|
| 51 |
+
Selection.
|
| 52 |
+
|
| 53 |
+
Notes
|
| 54 |
+
-----
|
| 55 |
+
Allows NaN in the input.
|
| 56 |
+
Raises ValueError if no feature in X meets the variance threshold.
|
| 57 |
+
|
| 58 |
+
Examples
|
| 59 |
+
--------
|
| 60 |
+
The following dataset has integer features, two of which are the same
|
| 61 |
+
in every sample. These are removed with the default setting for threshold::
|
| 62 |
+
|
| 63 |
+
>>> from sklearn.feature_selection import VarianceThreshold
|
| 64 |
+
>>> X = [[0, 2, 0, 3], [0, 1, 4, 3], [0, 1, 1, 3]]
|
| 65 |
+
>>> selector = VarianceThreshold()
|
| 66 |
+
>>> selector.fit_transform(X)
|
| 67 |
+
array([[2, 0],
|
| 68 |
+
[1, 4],
|
| 69 |
+
[1, 1]])
|
| 70 |
+
"""
|
| 71 |
+
|
| 72 |
+
_parameter_constraints: dict = {
|
| 73 |
+
"threshold": [Interval(Real, 0, None, closed="left")]
|
| 74 |
+
}
|
| 75 |
+
|
| 76 |
+
def __init__(self, threshold=0.0):
|
| 77 |
+
self.threshold = threshold
|
| 78 |
+
|
| 79 |
+
def fit(self, X, y=None):
|
| 80 |
+
"""Learn empirical variances from X.
|
| 81 |
+
|
| 82 |
+
Parameters
|
| 83 |
+
----------
|
| 84 |
+
X : {array-like, sparse matrix}, shape (n_samples, n_features)
|
| 85 |
+
Data from which to compute variances, where `n_samples` is
|
| 86 |
+
the number of samples and `n_features` is the number of features.
|
| 87 |
+
|
| 88 |
+
y : any, default=None
|
| 89 |
+
Ignored. This parameter exists only for compatibility with
|
| 90 |
+
sklearn.pipeline.Pipeline.
|
| 91 |
+
|
| 92 |
+
Returns
|
| 93 |
+
-------
|
| 94 |
+
self : object
|
| 95 |
+
Returns the instance itself.
|
| 96 |
+
"""
|
| 97 |
+
self._validate_params()
|
| 98 |
+
X = self._validate_data(
|
| 99 |
+
X,
|
| 100 |
+
accept_sparse=("csr", "csc"),
|
| 101 |
+
dtype=np.float64,
|
| 102 |
+
force_all_finite="allow-nan",
|
| 103 |
+
)
|
| 104 |
+
|
| 105 |
+
if hasattr(X, "toarray"): # sparse matrix
|
| 106 |
+
_, self.variances_ = mean_variance_axis(X, axis=0)
|
| 107 |
+
if self.threshold == 0:
|
| 108 |
+
mins, maxes = min_max_axis(X, axis=0)
|
| 109 |
+
peak_to_peaks = maxes - mins
|
| 110 |
+
else:
|
| 111 |
+
self.variances_ = np.nanvar(X, axis=0)
|
| 112 |
+
if self.threshold == 0:
|
| 113 |
+
peak_to_peaks = np.ptp(X, axis=0)
|
| 114 |
+
|
| 115 |
+
if self.threshold == 0:
|
| 116 |
+
# Use peak-to-peak to avoid numeric precision issues
|
| 117 |
+
# for constant features
|
| 118 |
+
compare_arr = np.array([self.variances_, peak_to_peaks])
|
| 119 |
+
self.variances_ = np.nanmin(compare_arr, axis=0)
|
| 120 |
+
|
| 121 |
+
if np.all(~np.isfinite(self.variances_) | (self.variances_ <= self.threshold)):
|
| 122 |
+
msg = "No feature in X meets the variance threshold {0:.5f}"
|
| 123 |
+
if X.shape[0] == 1:
|
| 124 |
+
msg += " (X contains only one sample)"
|
| 125 |
+
raise ValueError(msg.format(self.threshold))
|
| 126 |
+
|
| 127 |
+
return self
|
| 128 |
+
|
| 129 |
+
def _get_support_mask(self):
|
| 130 |
+
check_is_fitted(self)
|
| 131 |
+
|
| 132 |
+
return self.variances_ > self.threshold
|
| 133 |
+
|
| 134 |
+
def _more_tags(self):
|
| 135 |
+
return {"allow_nan": True}
|
videollama2/lib/python3.10/site-packages/sklearn/feature_selection/tests/test_feature_select.py
ADDED
|
@@ -0,0 +1,946 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Todo: cross-check the F-value with stats model
|
| 3 |
+
"""
|
| 4 |
+
import itertools
|
| 5 |
+
import warnings
|
| 6 |
+
import numpy as np
|
| 7 |
+
from numpy.testing import assert_allclose
|
| 8 |
+
from scipy import stats, sparse
|
| 9 |
+
|
| 10 |
+
import pytest
|
| 11 |
+
|
| 12 |
+
from sklearn.utils._testing import assert_almost_equal, _convert_container
|
| 13 |
+
from sklearn.utils._testing import assert_array_equal
|
| 14 |
+
from sklearn.utils._testing import assert_array_almost_equal
|
| 15 |
+
from sklearn.utils._testing import ignore_warnings
|
| 16 |
+
from sklearn.utils import safe_mask
|
| 17 |
+
|
| 18 |
+
from sklearn.datasets import make_classification, make_regression
|
| 19 |
+
from sklearn.feature_selection import (
|
| 20 |
+
chi2,
|
| 21 |
+
f_classif,
|
| 22 |
+
f_oneway,
|
| 23 |
+
f_regression,
|
| 24 |
+
GenericUnivariateSelect,
|
| 25 |
+
mutual_info_classif,
|
| 26 |
+
mutual_info_regression,
|
| 27 |
+
r_regression,
|
| 28 |
+
SelectPercentile,
|
| 29 |
+
SelectKBest,
|
| 30 |
+
SelectFpr,
|
| 31 |
+
SelectFdr,
|
| 32 |
+
SelectFwe,
|
| 33 |
+
)
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
##############################################################################
|
| 37 |
+
# Test the score functions
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
def test_f_oneway_vs_scipy_stats():
|
| 41 |
+
# Test that our f_oneway gives the same result as scipy.stats
|
| 42 |
+
rng = np.random.RandomState(0)
|
| 43 |
+
X1 = rng.randn(10, 3)
|
| 44 |
+
X2 = 1 + rng.randn(10, 3)
|
| 45 |
+
f, pv = stats.f_oneway(X1, X2)
|
| 46 |
+
f2, pv2 = f_oneway(X1, X2)
|
| 47 |
+
assert np.allclose(f, f2)
|
| 48 |
+
assert np.allclose(pv, pv2)
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def test_f_oneway_ints():
|
| 52 |
+
# Smoke test f_oneway on integers: that it does raise casting errors
|
| 53 |
+
# with recent numpys
|
| 54 |
+
rng = np.random.RandomState(0)
|
| 55 |
+
X = rng.randint(10, size=(10, 10))
|
| 56 |
+
y = np.arange(10)
|
| 57 |
+
fint, pint = f_oneway(X, y)
|
| 58 |
+
|
| 59 |
+
# test that is gives the same result as with float
|
| 60 |
+
f, p = f_oneway(X.astype(float), y)
|
| 61 |
+
assert_array_almost_equal(f, fint, decimal=4)
|
| 62 |
+
assert_array_almost_equal(p, pint, decimal=4)
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def test_f_classif():
|
| 66 |
+
# Test whether the F test yields meaningful results
|
| 67 |
+
# on a simple simulated classification problem
|
| 68 |
+
X, y = make_classification(
|
| 69 |
+
n_samples=200,
|
| 70 |
+
n_features=20,
|
| 71 |
+
n_informative=3,
|
| 72 |
+
n_redundant=2,
|
| 73 |
+
n_repeated=0,
|
| 74 |
+
n_classes=8,
|
| 75 |
+
n_clusters_per_class=1,
|
| 76 |
+
flip_y=0.0,
|
| 77 |
+
class_sep=10,
|
| 78 |
+
shuffle=False,
|
| 79 |
+
random_state=0,
|
| 80 |
+
)
|
| 81 |
+
|
| 82 |
+
F, pv = f_classif(X, y)
|
| 83 |
+
F_sparse, pv_sparse = f_classif(sparse.csr_matrix(X), y)
|
| 84 |
+
assert (F > 0).all()
|
| 85 |
+
assert (pv > 0).all()
|
| 86 |
+
assert (pv < 1).all()
|
| 87 |
+
assert (pv[:5] < 0.05).all()
|
| 88 |
+
assert (pv[5:] > 1.0e-4).all()
|
| 89 |
+
assert_array_almost_equal(F_sparse, F)
|
| 90 |
+
assert_array_almost_equal(pv_sparse, pv)
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
@pytest.mark.parametrize("center", [True, False])
|
| 94 |
+
def test_r_regression(center):
|
| 95 |
+
X, y = make_regression(
|
| 96 |
+
n_samples=2000, n_features=20, n_informative=5, shuffle=False, random_state=0
|
| 97 |
+
)
|
| 98 |
+
|
| 99 |
+
corr_coeffs = r_regression(X, y, center=center)
|
| 100 |
+
assert (-1 < corr_coeffs).all()
|
| 101 |
+
assert (corr_coeffs < 1).all()
|
| 102 |
+
|
| 103 |
+
sparse_X = _convert_container(X, "sparse")
|
| 104 |
+
|
| 105 |
+
sparse_corr_coeffs = r_regression(sparse_X, y, center=center)
|
| 106 |
+
assert_allclose(sparse_corr_coeffs, corr_coeffs)
|
| 107 |
+
|
| 108 |
+
# Testing against numpy for reference
|
| 109 |
+
Z = np.hstack((X, y[:, np.newaxis]))
|
| 110 |
+
correlation_matrix = np.corrcoef(Z, rowvar=False)
|
| 111 |
+
np_corr_coeffs = correlation_matrix[:-1, -1]
|
| 112 |
+
assert_array_almost_equal(np_corr_coeffs, corr_coeffs, decimal=3)
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def test_f_regression():
|
| 116 |
+
# Test whether the F test yields meaningful results
|
| 117 |
+
# on a simple simulated regression problem
|
| 118 |
+
X, y = make_regression(
|
| 119 |
+
n_samples=200, n_features=20, n_informative=5, shuffle=False, random_state=0
|
| 120 |
+
)
|
| 121 |
+
|
| 122 |
+
F, pv = f_regression(X, y)
|
| 123 |
+
assert (F > 0).all()
|
| 124 |
+
assert (pv > 0).all()
|
| 125 |
+
assert (pv < 1).all()
|
| 126 |
+
assert (pv[:5] < 0.05).all()
|
| 127 |
+
assert (pv[5:] > 1.0e-4).all()
|
| 128 |
+
|
| 129 |
+
# with centering, compare with sparse
|
| 130 |
+
F, pv = f_regression(X, y, center=True)
|
| 131 |
+
F_sparse, pv_sparse = f_regression(sparse.csr_matrix(X), y, center=True)
|
| 132 |
+
assert_allclose(F_sparse, F)
|
| 133 |
+
assert_allclose(pv_sparse, pv)
|
| 134 |
+
|
| 135 |
+
# again without centering, compare with sparse
|
| 136 |
+
F, pv = f_regression(X, y, center=False)
|
| 137 |
+
F_sparse, pv_sparse = f_regression(sparse.csr_matrix(X), y, center=False)
|
| 138 |
+
assert_allclose(F_sparse, F)
|
| 139 |
+
assert_allclose(pv_sparse, pv)
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
def test_f_regression_input_dtype():
|
| 143 |
+
# Test whether f_regression returns the same value
|
| 144 |
+
# for any numeric data_type
|
| 145 |
+
rng = np.random.RandomState(0)
|
| 146 |
+
X = rng.rand(10, 20)
|
| 147 |
+
y = np.arange(10).astype(int)
|
| 148 |
+
|
| 149 |
+
F1, pv1 = f_regression(X, y)
|
| 150 |
+
F2, pv2 = f_regression(X, y.astype(float))
|
| 151 |
+
assert_allclose(F1, F2, 5)
|
| 152 |
+
assert_allclose(pv1, pv2, 5)
|
| 153 |
+
|
| 154 |
+
|
| 155 |
+
def test_f_regression_center():
|
| 156 |
+
# Test whether f_regression preserves dof according to 'center' argument
|
| 157 |
+
# We use two centered variates so we have a simple relationship between
|
| 158 |
+
# F-score with variates centering and F-score without variates centering.
|
| 159 |
+
# Create toy example
|
| 160 |
+
X = np.arange(-5, 6).reshape(-1, 1) # X has zero mean
|
| 161 |
+
n_samples = X.size
|
| 162 |
+
Y = np.ones(n_samples)
|
| 163 |
+
Y[::2] *= -1.0
|
| 164 |
+
Y[0] = 0.0 # have Y mean being null
|
| 165 |
+
|
| 166 |
+
F1, _ = f_regression(X, Y, center=True)
|
| 167 |
+
F2, _ = f_regression(X, Y, center=False)
|
| 168 |
+
assert_allclose(F1 * (n_samples - 1.0) / (n_samples - 2.0), F2)
|
| 169 |
+
assert_almost_equal(F2[0], 0.232558139) # value from statsmodels OLS
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
@pytest.mark.parametrize(
|
| 173 |
+
"X, y, expected_corr_coef, force_finite",
|
| 174 |
+
[
|
| 175 |
+
(
|
| 176 |
+
# A feature in X is constant - forcing finite
|
| 177 |
+
np.array([[2, 1], [2, 0], [2, 10], [2, 4]]),
|
| 178 |
+
np.array([0, 1, 1, 0]),
|
| 179 |
+
np.array([0.0, 0.32075]),
|
| 180 |
+
True,
|
| 181 |
+
),
|
| 182 |
+
(
|
| 183 |
+
# The target y is constant - forcing finite
|
| 184 |
+
np.array([[5, 1], [3, 0], [2, 10], [8, 4]]),
|
| 185 |
+
np.array([0, 0, 0, 0]),
|
| 186 |
+
np.array([0.0, 0.0]),
|
| 187 |
+
True,
|
| 188 |
+
),
|
| 189 |
+
(
|
| 190 |
+
# A feature in X is constant - not forcing finite
|
| 191 |
+
np.array([[2, 1], [2, 0], [2, 10], [2, 4]]),
|
| 192 |
+
np.array([0, 1, 1, 0]),
|
| 193 |
+
np.array([np.nan, 0.32075]),
|
| 194 |
+
False,
|
| 195 |
+
),
|
| 196 |
+
(
|
| 197 |
+
# The target y is constant - not forcing finite
|
| 198 |
+
np.array([[5, 1], [3, 0], [2, 10], [8, 4]]),
|
| 199 |
+
np.array([0, 0, 0, 0]),
|
| 200 |
+
np.array([np.nan, np.nan]),
|
| 201 |
+
False,
|
| 202 |
+
),
|
| 203 |
+
],
|
| 204 |
+
)
|
| 205 |
+
def test_r_regression_force_finite(X, y, expected_corr_coef, force_finite):
|
| 206 |
+
"""Check the behaviour of `force_finite` for some corner cases with `r_regression`.
|
| 207 |
+
|
| 208 |
+
Non-regression test for:
|
| 209 |
+
https://github.com/scikit-learn/scikit-learn/issues/15672
|
| 210 |
+
"""
|
| 211 |
+
with warnings.catch_warnings():
|
| 212 |
+
warnings.simplefilter("error", RuntimeWarning)
|
| 213 |
+
corr_coef = r_regression(X, y, force_finite=force_finite)
|
| 214 |
+
np.testing.assert_array_almost_equal(corr_coef, expected_corr_coef)
|
| 215 |
+
|
| 216 |
+
|
| 217 |
+
@pytest.mark.parametrize(
|
| 218 |
+
"X, y, expected_f_statistic, expected_p_values, force_finite",
|
| 219 |
+
[
|
| 220 |
+
(
|
| 221 |
+
# A feature in X is constant - forcing finite
|
| 222 |
+
np.array([[2, 1], [2, 0], [2, 10], [2, 4]]),
|
| 223 |
+
np.array([0, 1, 1, 0]),
|
| 224 |
+
np.array([0.0, 0.2293578]),
|
| 225 |
+
np.array([1.0, 0.67924985]),
|
| 226 |
+
True,
|
| 227 |
+
),
|
| 228 |
+
(
|
| 229 |
+
# The target y is constant - forcing finite
|
| 230 |
+
np.array([[5, 1], [3, 0], [2, 10], [8, 4]]),
|
| 231 |
+
np.array([0, 0, 0, 0]),
|
| 232 |
+
np.array([0.0, 0.0]),
|
| 233 |
+
np.array([1.0, 1.0]),
|
| 234 |
+
True,
|
| 235 |
+
),
|
| 236 |
+
(
|
| 237 |
+
# Feature in X correlated with y - forcing finite
|
| 238 |
+
np.array([[0, 1], [1, 0], [2, 10], [3, 4]]),
|
| 239 |
+
np.array([0, 1, 2, 3]),
|
| 240 |
+
np.array([np.finfo(np.float64).max, 0.845433]),
|
| 241 |
+
np.array([0.0, 0.454913]),
|
| 242 |
+
True,
|
| 243 |
+
),
|
| 244 |
+
(
|
| 245 |
+
# Feature in X anti-correlated with y - forcing finite
|
| 246 |
+
np.array([[3, 1], [2, 0], [1, 10], [0, 4]]),
|
| 247 |
+
np.array([0, 1, 2, 3]),
|
| 248 |
+
np.array([np.finfo(np.float64).max, 0.845433]),
|
| 249 |
+
np.array([0.0, 0.454913]),
|
| 250 |
+
True,
|
| 251 |
+
),
|
| 252 |
+
(
|
| 253 |
+
# A feature in X is constant - not forcing finite
|
| 254 |
+
np.array([[2, 1], [2, 0], [2, 10], [2, 4]]),
|
| 255 |
+
np.array([0, 1, 1, 0]),
|
| 256 |
+
np.array([np.nan, 0.2293578]),
|
| 257 |
+
np.array([np.nan, 0.67924985]),
|
| 258 |
+
False,
|
| 259 |
+
),
|
| 260 |
+
(
|
| 261 |
+
# The target y is constant - not forcing finite
|
| 262 |
+
np.array([[5, 1], [3, 0], [2, 10], [8, 4]]),
|
| 263 |
+
np.array([0, 0, 0, 0]),
|
| 264 |
+
np.array([np.nan, np.nan]),
|
| 265 |
+
np.array([np.nan, np.nan]),
|
| 266 |
+
False,
|
| 267 |
+
),
|
| 268 |
+
(
|
| 269 |
+
# Feature in X correlated with y - not forcing finite
|
| 270 |
+
np.array([[0, 1], [1, 0], [2, 10], [3, 4]]),
|
| 271 |
+
np.array([0, 1, 2, 3]),
|
| 272 |
+
np.array([np.inf, 0.845433]),
|
| 273 |
+
np.array([0.0, 0.454913]),
|
| 274 |
+
False,
|
| 275 |
+
),
|
| 276 |
+
(
|
| 277 |
+
# Feature in X anti-correlated with y - not forcing finite
|
| 278 |
+
np.array([[3, 1], [2, 0], [1, 10], [0, 4]]),
|
| 279 |
+
np.array([0, 1, 2, 3]),
|
| 280 |
+
np.array([np.inf, 0.845433]),
|
| 281 |
+
np.array([0.0, 0.454913]),
|
| 282 |
+
False,
|
| 283 |
+
),
|
| 284 |
+
],
|
| 285 |
+
)
|
| 286 |
+
def test_f_regression_corner_case(
|
| 287 |
+
X, y, expected_f_statistic, expected_p_values, force_finite
|
| 288 |
+
):
|
| 289 |
+
"""Check the behaviour of `force_finite` for some corner cases with `f_regression`.
|
| 290 |
+
|
| 291 |
+
Non-regression test for:
|
| 292 |
+
https://github.com/scikit-learn/scikit-learn/issues/15672
|
| 293 |
+
"""
|
| 294 |
+
with warnings.catch_warnings():
|
| 295 |
+
warnings.simplefilter("error", RuntimeWarning)
|
| 296 |
+
f_statistic, p_values = f_regression(X, y, force_finite=force_finite)
|
| 297 |
+
np.testing.assert_array_almost_equal(f_statistic, expected_f_statistic)
|
| 298 |
+
np.testing.assert_array_almost_equal(p_values, expected_p_values)
|
| 299 |
+
|
| 300 |
+
|
| 301 |
+
def test_f_classif_multi_class():
|
| 302 |
+
# Test whether the F test yields meaningful results
|
| 303 |
+
# on a simple simulated classification problem
|
| 304 |
+
X, y = make_classification(
|
| 305 |
+
n_samples=200,
|
| 306 |
+
n_features=20,
|
| 307 |
+
n_informative=3,
|
| 308 |
+
n_redundant=2,
|
| 309 |
+
n_repeated=0,
|
| 310 |
+
n_classes=8,
|
| 311 |
+
n_clusters_per_class=1,
|
| 312 |
+
flip_y=0.0,
|
| 313 |
+
class_sep=10,
|
| 314 |
+
shuffle=False,
|
| 315 |
+
random_state=0,
|
| 316 |
+
)
|
| 317 |
+
|
| 318 |
+
F, pv = f_classif(X, y)
|
| 319 |
+
assert (F > 0).all()
|
| 320 |
+
assert (pv > 0).all()
|
| 321 |
+
assert (pv < 1).all()
|
| 322 |
+
assert (pv[:5] < 0.05).all()
|
| 323 |
+
assert (pv[5:] > 1.0e-4).all()
|
| 324 |
+
|
| 325 |
+
|
| 326 |
+
def test_select_percentile_classif():
|
| 327 |
+
# Test whether the relative univariate feature selection
|
| 328 |
+
# gets the correct items in a simple classification problem
|
| 329 |
+
# with the percentile heuristic
|
| 330 |
+
X, y = make_classification(
|
| 331 |
+
n_samples=200,
|
| 332 |
+
n_features=20,
|
| 333 |
+
n_informative=3,
|
| 334 |
+
n_redundant=2,
|
| 335 |
+
n_repeated=0,
|
| 336 |
+
n_classes=8,
|
| 337 |
+
n_clusters_per_class=1,
|
| 338 |
+
flip_y=0.0,
|
| 339 |
+
class_sep=10,
|
| 340 |
+
shuffle=False,
|
| 341 |
+
random_state=0,
|
| 342 |
+
)
|
| 343 |
+
|
| 344 |
+
univariate_filter = SelectPercentile(f_classif, percentile=25)
|
| 345 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 346 |
+
X_r2 = (
|
| 347 |
+
GenericUnivariateSelect(f_classif, mode="percentile", param=25)
|
| 348 |
+
.fit(X, y)
|
| 349 |
+
.transform(X)
|
| 350 |
+
)
|
| 351 |
+
assert_array_equal(X_r, X_r2)
|
| 352 |
+
support = univariate_filter.get_support()
|
| 353 |
+
gtruth = np.zeros(20)
|
| 354 |
+
gtruth[:5] = 1
|
| 355 |
+
assert_array_equal(support, gtruth)
|
| 356 |
+
|
| 357 |
+
|
| 358 |
+
def test_select_percentile_classif_sparse():
|
| 359 |
+
# Test whether the relative univariate feature selection
|
| 360 |
+
# gets the correct items in a simple classification problem
|
| 361 |
+
# with the percentile heuristic
|
| 362 |
+
X, y = make_classification(
|
| 363 |
+
n_samples=200,
|
| 364 |
+
n_features=20,
|
| 365 |
+
n_informative=3,
|
| 366 |
+
n_redundant=2,
|
| 367 |
+
n_repeated=0,
|
| 368 |
+
n_classes=8,
|
| 369 |
+
n_clusters_per_class=1,
|
| 370 |
+
flip_y=0.0,
|
| 371 |
+
class_sep=10,
|
| 372 |
+
shuffle=False,
|
| 373 |
+
random_state=0,
|
| 374 |
+
)
|
| 375 |
+
X = sparse.csr_matrix(X)
|
| 376 |
+
univariate_filter = SelectPercentile(f_classif, percentile=25)
|
| 377 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 378 |
+
X_r2 = (
|
| 379 |
+
GenericUnivariateSelect(f_classif, mode="percentile", param=25)
|
| 380 |
+
.fit(X, y)
|
| 381 |
+
.transform(X)
|
| 382 |
+
)
|
| 383 |
+
assert_array_equal(X_r.toarray(), X_r2.toarray())
|
| 384 |
+
support = univariate_filter.get_support()
|
| 385 |
+
gtruth = np.zeros(20)
|
| 386 |
+
gtruth[:5] = 1
|
| 387 |
+
assert_array_equal(support, gtruth)
|
| 388 |
+
|
| 389 |
+
X_r2inv = univariate_filter.inverse_transform(X_r2)
|
| 390 |
+
assert sparse.issparse(X_r2inv)
|
| 391 |
+
support_mask = safe_mask(X_r2inv, support)
|
| 392 |
+
assert X_r2inv.shape == X.shape
|
| 393 |
+
assert_array_equal(X_r2inv[:, support_mask].toarray(), X_r.toarray())
|
| 394 |
+
# Check other columns are empty
|
| 395 |
+
assert X_r2inv.getnnz() == X_r.getnnz()
|
| 396 |
+
|
| 397 |
+
|
| 398 |
+
##############################################################################
|
| 399 |
+
# Test univariate selection in classification settings
|
| 400 |
+
|
| 401 |
+
|
| 402 |
+
def test_select_kbest_classif():
|
| 403 |
+
# Test whether the relative univariate feature selection
|
| 404 |
+
# gets the correct items in a simple classification problem
|
| 405 |
+
# with the k best heuristic
|
| 406 |
+
X, y = make_classification(
|
| 407 |
+
n_samples=200,
|
| 408 |
+
n_features=20,
|
| 409 |
+
n_informative=3,
|
| 410 |
+
n_redundant=2,
|
| 411 |
+
n_repeated=0,
|
| 412 |
+
n_classes=8,
|
| 413 |
+
n_clusters_per_class=1,
|
| 414 |
+
flip_y=0.0,
|
| 415 |
+
class_sep=10,
|
| 416 |
+
shuffle=False,
|
| 417 |
+
random_state=0,
|
| 418 |
+
)
|
| 419 |
+
|
| 420 |
+
univariate_filter = SelectKBest(f_classif, k=5)
|
| 421 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 422 |
+
X_r2 = (
|
| 423 |
+
GenericUnivariateSelect(f_classif, mode="k_best", param=5)
|
| 424 |
+
.fit(X, y)
|
| 425 |
+
.transform(X)
|
| 426 |
+
)
|
| 427 |
+
assert_array_equal(X_r, X_r2)
|
| 428 |
+
support = univariate_filter.get_support()
|
| 429 |
+
gtruth = np.zeros(20)
|
| 430 |
+
gtruth[:5] = 1
|
| 431 |
+
assert_array_equal(support, gtruth)
|
| 432 |
+
|
| 433 |
+
|
| 434 |
+
def test_select_kbest_all():
|
| 435 |
+
# Test whether k="all" correctly returns all features.
|
| 436 |
+
X, y = make_classification(
|
| 437 |
+
n_samples=20, n_features=10, shuffle=False, random_state=0
|
| 438 |
+
)
|
| 439 |
+
|
| 440 |
+
univariate_filter = SelectKBest(f_classif, k="all")
|
| 441 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 442 |
+
assert_array_equal(X, X_r)
|
| 443 |
+
# Non-regression test for:
|
| 444 |
+
# https://github.com/scikit-learn/scikit-learn/issues/24949
|
| 445 |
+
X_r2 = (
|
| 446 |
+
GenericUnivariateSelect(f_classif, mode="k_best", param="all")
|
| 447 |
+
.fit(X, y)
|
| 448 |
+
.transform(X)
|
| 449 |
+
)
|
| 450 |
+
assert_array_equal(X_r, X_r2)
|
| 451 |
+
|
| 452 |
+
|
| 453 |
+
@pytest.mark.parametrize("dtype_in", [np.float32, np.float64])
|
| 454 |
+
def test_select_kbest_zero(dtype_in):
|
| 455 |
+
# Test whether k=0 correctly returns no features.
|
| 456 |
+
X, y = make_classification(
|
| 457 |
+
n_samples=20, n_features=10, shuffle=False, random_state=0
|
| 458 |
+
)
|
| 459 |
+
X = X.astype(dtype_in)
|
| 460 |
+
|
| 461 |
+
univariate_filter = SelectKBest(f_classif, k=0)
|
| 462 |
+
univariate_filter.fit(X, y)
|
| 463 |
+
support = univariate_filter.get_support()
|
| 464 |
+
gtruth = np.zeros(10, dtype=bool)
|
| 465 |
+
assert_array_equal(support, gtruth)
|
| 466 |
+
with pytest.warns(UserWarning, match="No features were selected"):
|
| 467 |
+
X_selected = univariate_filter.transform(X)
|
| 468 |
+
assert X_selected.shape == (20, 0)
|
| 469 |
+
assert X_selected.dtype == dtype_in
|
| 470 |
+
|
| 471 |
+
|
| 472 |
+
def test_select_heuristics_classif():
|
| 473 |
+
# Test whether the relative univariate feature selection
|
| 474 |
+
# gets the correct items in a simple classification problem
|
| 475 |
+
# with the fdr, fwe and fpr heuristics
|
| 476 |
+
X, y = make_classification(
|
| 477 |
+
n_samples=200,
|
| 478 |
+
n_features=20,
|
| 479 |
+
n_informative=3,
|
| 480 |
+
n_redundant=2,
|
| 481 |
+
n_repeated=0,
|
| 482 |
+
n_classes=8,
|
| 483 |
+
n_clusters_per_class=1,
|
| 484 |
+
flip_y=0.0,
|
| 485 |
+
class_sep=10,
|
| 486 |
+
shuffle=False,
|
| 487 |
+
random_state=0,
|
| 488 |
+
)
|
| 489 |
+
|
| 490 |
+
univariate_filter = SelectFwe(f_classif, alpha=0.01)
|
| 491 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 492 |
+
gtruth = np.zeros(20)
|
| 493 |
+
gtruth[:5] = 1
|
| 494 |
+
for mode in ["fdr", "fpr", "fwe"]:
|
| 495 |
+
X_r2 = (
|
| 496 |
+
GenericUnivariateSelect(f_classif, mode=mode, param=0.01)
|
| 497 |
+
.fit(X, y)
|
| 498 |
+
.transform(X)
|
| 499 |
+
)
|
| 500 |
+
assert_array_equal(X_r, X_r2)
|
| 501 |
+
support = univariate_filter.get_support()
|
| 502 |
+
assert_allclose(support, gtruth)
|
| 503 |
+
|
| 504 |
+
|
| 505 |
+
##############################################################################
|
| 506 |
+
# Test univariate selection in regression settings
|
| 507 |
+
|
| 508 |
+
|
| 509 |
+
def assert_best_scores_kept(score_filter):
|
| 510 |
+
scores = score_filter.scores_
|
| 511 |
+
support = score_filter.get_support()
|
| 512 |
+
assert_allclose(np.sort(scores[support]), np.sort(scores)[-support.sum() :])
|
| 513 |
+
|
| 514 |
+
|
| 515 |
+
def test_select_percentile_regression():
|
| 516 |
+
# Test whether the relative univariate feature selection
|
| 517 |
+
# gets the correct items in a simple regression problem
|
| 518 |
+
# with the percentile heuristic
|
| 519 |
+
X, y = make_regression(
|
| 520 |
+
n_samples=200, n_features=20, n_informative=5, shuffle=False, random_state=0
|
| 521 |
+
)
|
| 522 |
+
|
| 523 |
+
univariate_filter = SelectPercentile(f_regression, percentile=25)
|
| 524 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 525 |
+
assert_best_scores_kept(univariate_filter)
|
| 526 |
+
X_r2 = (
|
| 527 |
+
GenericUnivariateSelect(f_regression, mode="percentile", param=25)
|
| 528 |
+
.fit(X, y)
|
| 529 |
+
.transform(X)
|
| 530 |
+
)
|
| 531 |
+
assert_array_equal(X_r, X_r2)
|
| 532 |
+
support = univariate_filter.get_support()
|
| 533 |
+
gtruth = np.zeros(20)
|
| 534 |
+
gtruth[:5] = 1
|
| 535 |
+
assert_array_equal(support, gtruth)
|
| 536 |
+
X_2 = X.copy()
|
| 537 |
+
X_2[:, np.logical_not(support)] = 0
|
| 538 |
+
assert_array_equal(X_2, univariate_filter.inverse_transform(X_r))
|
| 539 |
+
# Check inverse_transform respects dtype
|
| 540 |
+
assert_array_equal(
|
| 541 |
+
X_2.astype(bool), univariate_filter.inverse_transform(X_r.astype(bool))
|
| 542 |
+
)
|
| 543 |
+
|
| 544 |
+
|
| 545 |
+
def test_select_percentile_regression_full():
|
| 546 |
+
# Test whether the relative univariate feature selection
|
| 547 |
+
# selects all features when '100%' is asked.
|
| 548 |
+
X, y = make_regression(
|
| 549 |
+
n_samples=200, n_features=20, n_informative=5, shuffle=False, random_state=0
|
| 550 |
+
)
|
| 551 |
+
|
| 552 |
+
univariate_filter = SelectPercentile(f_regression, percentile=100)
|
| 553 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 554 |
+
assert_best_scores_kept(univariate_filter)
|
| 555 |
+
X_r2 = (
|
| 556 |
+
GenericUnivariateSelect(f_regression, mode="percentile", param=100)
|
| 557 |
+
.fit(X, y)
|
| 558 |
+
.transform(X)
|
| 559 |
+
)
|
| 560 |
+
assert_array_equal(X_r, X_r2)
|
| 561 |
+
support = univariate_filter.get_support()
|
| 562 |
+
gtruth = np.ones(20)
|
| 563 |
+
assert_array_equal(support, gtruth)
|
| 564 |
+
|
| 565 |
+
|
| 566 |
+
def test_select_kbest_regression():
|
| 567 |
+
# Test whether the relative univariate feature selection
|
| 568 |
+
# gets the correct items in a simple regression problem
|
| 569 |
+
# with the k best heuristic
|
| 570 |
+
X, y = make_regression(
|
| 571 |
+
n_samples=200,
|
| 572 |
+
n_features=20,
|
| 573 |
+
n_informative=5,
|
| 574 |
+
shuffle=False,
|
| 575 |
+
random_state=0,
|
| 576 |
+
noise=10,
|
| 577 |
+
)
|
| 578 |
+
|
| 579 |
+
univariate_filter = SelectKBest(f_regression, k=5)
|
| 580 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 581 |
+
assert_best_scores_kept(univariate_filter)
|
| 582 |
+
X_r2 = (
|
| 583 |
+
GenericUnivariateSelect(f_regression, mode="k_best", param=5)
|
| 584 |
+
.fit(X, y)
|
| 585 |
+
.transform(X)
|
| 586 |
+
)
|
| 587 |
+
assert_array_equal(X_r, X_r2)
|
| 588 |
+
support = univariate_filter.get_support()
|
| 589 |
+
gtruth = np.zeros(20)
|
| 590 |
+
gtruth[:5] = 1
|
| 591 |
+
assert_array_equal(support, gtruth)
|
| 592 |
+
|
| 593 |
+
|
| 594 |
+
def test_select_heuristics_regression():
|
| 595 |
+
# Test whether the relative univariate feature selection
|
| 596 |
+
# gets the correct items in a simple regression problem
|
| 597 |
+
# with the fpr, fdr or fwe heuristics
|
| 598 |
+
X, y = make_regression(
|
| 599 |
+
n_samples=200,
|
| 600 |
+
n_features=20,
|
| 601 |
+
n_informative=5,
|
| 602 |
+
shuffle=False,
|
| 603 |
+
random_state=0,
|
| 604 |
+
noise=10,
|
| 605 |
+
)
|
| 606 |
+
|
| 607 |
+
univariate_filter = SelectFpr(f_regression, alpha=0.01)
|
| 608 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 609 |
+
gtruth = np.zeros(20)
|
| 610 |
+
gtruth[:5] = 1
|
| 611 |
+
for mode in ["fdr", "fpr", "fwe"]:
|
| 612 |
+
X_r2 = (
|
| 613 |
+
GenericUnivariateSelect(f_regression, mode=mode, param=0.01)
|
| 614 |
+
.fit(X, y)
|
| 615 |
+
.transform(X)
|
| 616 |
+
)
|
| 617 |
+
assert_array_equal(X_r, X_r2)
|
| 618 |
+
support = univariate_filter.get_support()
|
| 619 |
+
assert_array_equal(support[:5], np.ones((5,), dtype=bool))
|
| 620 |
+
assert np.sum(support[5:] == 1) < 3
|
| 621 |
+
|
| 622 |
+
|
| 623 |
+
def test_boundary_case_ch2():
|
| 624 |
+
# Test boundary case, and always aim to select 1 feature.
|
| 625 |
+
X = np.array([[10, 20], [20, 20], [20, 30]])
|
| 626 |
+
y = np.array([[1], [0], [0]])
|
| 627 |
+
scores, pvalues = chi2(X, y)
|
| 628 |
+
assert_array_almost_equal(scores, np.array([4.0, 0.71428571]))
|
| 629 |
+
assert_array_almost_equal(pvalues, np.array([0.04550026, 0.39802472]))
|
| 630 |
+
|
| 631 |
+
filter_fdr = SelectFdr(chi2, alpha=0.1)
|
| 632 |
+
filter_fdr.fit(X, y)
|
| 633 |
+
support_fdr = filter_fdr.get_support()
|
| 634 |
+
assert_array_equal(support_fdr, np.array([True, False]))
|
| 635 |
+
|
| 636 |
+
filter_kbest = SelectKBest(chi2, k=1)
|
| 637 |
+
filter_kbest.fit(X, y)
|
| 638 |
+
support_kbest = filter_kbest.get_support()
|
| 639 |
+
assert_array_equal(support_kbest, np.array([True, False]))
|
| 640 |
+
|
| 641 |
+
filter_percentile = SelectPercentile(chi2, percentile=50)
|
| 642 |
+
filter_percentile.fit(X, y)
|
| 643 |
+
support_percentile = filter_percentile.get_support()
|
| 644 |
+
assert_array_equal(support_percentile, np.array([True, False]))
|
| 645 |
+
|
| 646 |
+
filter_fpr = SelectFpr(chi2, alpha=0.1)
|
| 647 |
+
filter_fpr.fit(X, y)
|
| 648 |
+
support_fpr = filter_fpr.get_support()
|
| 649 |
+
assert_array_equal(support_fpr, np.array([True, False]))
|
| 650 |
+
|
| 651 |
+
filter_fwe = SelectFwe(chi2, alpha=0.1)
|
| 652 |
+
filter_fwe.fit(X, y)
|
| 653 |
+
support_fwe = filter_fwe.get_support()
|
| 654 |
+
assert_array_equal(support_fwe, np.array([True, False]))
|
| 655 |
+
|
| 656 |
+
|
| 657 |
+
@pytest.mark.parametrize("alpha", [0.001, 0.01, 0.1])
|
| 658 |
+
@pytest.mark.parametrize("n_informative", [1, 5, 10])
|
| 659 |
+
def test_select_fdr_regression(alpha, n_informative):
|
| 660 |
+
# Test that fdr heuristic actually has low FDR.
|
| 661 |
+
def single_fdr(alpha, n_informative, random_state):
|
| 662 |
+
X, y = make_regression(
|
| 663 |
+
n_samples=150,
|
| 664 |
+
n_features=20,
|
| 665 |
+
n_informative=n_informative,
|
| 666 |
+
shuffle=False,
|
| 667 |
+
random_state=random_state,
|
| 668 |
+
noise=10,
|
| 669 |
+
)
|
| 670 |
+
|
| 671 |
+
with warnings.catch_warnings(record=True):
|
| 672 |
+
# Warnings can be raised when no features are selected
|
| 673 |
+
# (low alpha or very noisy data)
|
| 674 |
+
univariate_filter = SelectFdr(f_regression, alpha=alpha)
|
| 675 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 676 |
+
X_r2 = (
|
| 677 |
+
GenericUnivariateSelect(f_regression, mode="fdr", param=alpha)
|
| 678 |
+
.fit(X, y)
|
| 679 |
+
.transform(X)
|
| 680 |
+
)
|
| 681 |
+
|
| 682 |
+
assert_array_equal(X_r, X_r2)
|
| 683 |
+
support = univariate_filter.get_support()
|
| 684 |
+
num_false_positives = np.sum(support[n_informative:] == 1)
|
| 685 |
+
num_true_positives = np.sum(support[:n_informative] == 1)
|
| 686 |
+
|
| 687 |
+
if num_false_positives == 0:
|
| 688 |
+
return 0.0
|
| 689 |
+
false_discovery_rate = num_false_positives / (
|
| 690 |
+
num_true_positives + num_false_positives
|
| 691 |
+
)
|
| 692 |
+
return false_discovery_rate
|
| 693 |
+
|
| 694 |
+
# As per Benjamini-Hochberg, the expected false discovery rate
|
| 695 |
+
# should be lower than alpha:
|
| 696 |
+
# FDR = E(FP / (TP + FP)) <= alpha
|
| 697 |
+
false_discovery_rate = np.mean(
|
| 698 |
+
[single_fdr(alpha, n_informative, random_state) for random_state in range(100)]
|
| 699 |
+
)
|
| 700 |
+
assert alpha >= false_discovery_rate
|
| 701 |
+
|
| 702 |
+
# Make sure that the empirical false discovery rate increases
|
| 703 |
+
# with alpha:
|
| 704 |
+
if false_discovery_rate != 0:
|
| 705 |
+
assert false_discovery_rate > alpha / 10
|
| 706 |
+
|
| 707 |
+
|
| 708 |
+
def test_select_fwe_regression():
|
| 709 |
+
# Test whether the relative univariate feature selection
|
| 710 |
+
# gets the correct items in a simple regression problem
|
| 711 |
+
# with the fwe heuristic
|
| 712 |
+
X, y = make_regression(
|
| 713 |
+
n_samples=200, n_features=20, n_informative=5, shuffle=False, random_state=0
|
| 714 |
+
)
|
| 715 |
+
|
| 716 |
+
univariate_filter = SelectFwe(f_regression, alpha=0.01)
|
| 717 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 718 |
+
X_r2 = (
|
| 719 |
+
GenericUnivariateSelect(f_regression, mode="fwe", param=0.01)
|
| 720 |
+
.fit(X, y)
|
| 721 |
+
.transform(X)
|
| 722 |
+
)
|
| 723 |
+
assert_array_equal(X_r, X_r2)
|
| 724 |
+
support = univariate_filter.get_support()
|
| 725 |
+
gtruth = np.zeros(20)
|
| 726 |
+
gtruth[:5] = 1
|
| 727 |
+
assert_array_equal(support[:5], np.ones((5,), dtype=bool))
|
| 728 |
+
assert np.sum(support[5:] == 1) < 2
|
| 729 |
+
|
| 730 |
+
|
| 731 |
+
def test_selectkbest_tiebreaking():
|
| 732 |
+
# Test whether SelectKBest actually selects k features in case of ties.
|
| 733 |
+
# Prior to 0.11, SelectKBest would return more features than requested.
|
| 734 |
+
Xs = [[0, 1, 1], [0, 0, 1], [1, 0, 0], [1, 1, 0]]
|
| 735 |
+
y = [1]
|
| 736 |
+
dummy_score = lambda X, y: (X[0], X[0])
|
| 737 |
+
for X in Xs:
|
| 738 |
+
sel = SelectKBest(dummy_score, k=1)
|
| 739 |
+
X1 = ignore_warnings(sel.fit_transform)([X], y)
|
| 740 |
+
assert X1.shape[1] == 1
|
| 741 |
+
assert_best_scores_kept(sel)
|
| 742 |
+
|
| 743 |
+
sel = SelectKBest(dummy_score, k=2)
|
| 744 |
+
X2 = ignore_warnings(sel.fit_transform)([X], y)
|
| 745 |
+
assert X2.shape[1] == 2
|
| 746 |
+
assert_best_scores_kept(sel)
|
| 747 |
+
|
| 748 |
+
|
| 749 |
+
def test_selectpercentile_tiebreaking():
|
| 750 |
+
# Test if SelectPercentile selects the right n_features in case of ties.
|
| 751 |
+
Xs = [[0, 1, 1], [0, 0, 1], [1, 0, 0], [1, 1, 0]]
|
| 752 |
+
y = [1]
|
| 753 |
+
dummy_score = lambda X, y: (X[0], X[0])
|
| 754 |
+
for X in Xs:
|
| 755 |
+
sel = SelectPercentile(dummy_score, percentile=34)
|
| 756 |
+
X1 = ignore_warnings(sel.fit_transform)([X], y)
|
| 757 |
+
assert X1.shape[1] == 1
|
| 758 |
+
assert_best_scores_kept(sel)
|
| 759 |
+
|
| 760 |
+
sel = SelectPercentile(dummy_score, percentile=67)
|
| 761 |
+
X2 = ignore_warnings(sel.fit_transform)([X], y)
|
| 762 |
+
assert X2.shape[1] == 2
|
| 763 |
+
assert_best_scores_kept(sel)
|
| 764 |
+
|
| 765 |
+
|
| 766 |
+
def test_tied_pvalues():
|
| 767 |
+
# Test whether k-best and percentiles work with tied pvalues from chi2.
|
| 768 |
+
# chi2 will return the same p-values for the following features, but it
|
| 769 |
+
# will return different scores.
|
| 770 |
+
X0 = np.array([[10000, 9999, 9998], [1, 1, 1]])
|
| 771 |
+
y = [0, 1]
|
| 772 |
+
|
| 773 |
+
for perm in itertools.permutations((0, 1, 2)):
|
| 774 |
+
X = X0[:, perm]
|
| 775 |
+
Xt = SelectKBest(chi2, k=2).fit_transform(X, y)
|
| 776 |
+
assert Xt.shape == (2, 2)
|
| 777 |
+
assert 9998 not in Xt
|
| 778 |
+
|
| 779 |
+
Xt = SelectPercentile(chi2, percentile=67).fit_transform(X, y)
|
| 780 |
+
assert Xt.shape == (2, 2)
|
| 781 |
+
assert 9998 not in Xt
|
| 782 |
+
|
| 783 |
+
|
| 784 |
+
def test_scorefunc_multilabel():
|
| 785 |
+
# Test whether k-best and percentiles works with multilabels with chi2.
|
| 786 |
+
|
| 787 |
+
X = np.array([[10000, 9999, 0], [100, 9999, 0], [1000, 99, 0]])
|
| 788 |
+
y = [[1, 1], [0, 1], [1, 0]]
|
| 789 |
+
|
| 790 |
+
Xt = SelectKBest(chi2, k=2).fit_transform(X, y)
|
| 791 |
+
assert Xt.shape == (3, 2)
|
| 792 |
+
assert 0 not in Xt
|
| 793 |
+
|
| 794 |
+
Xt = SelectPercentile(chi2, percentile=67).fit_transform(X, y)
|
| 795 |
+
assert Xt.shape == (3, 2)
|
| 796 |
+
assert 0 not in Xt
|
| 797 |
+
|
| 798 |
+
|
| 799 |
+
def test_tied_scores():
|
| 800 |
+
# Test for stable sorting in k-best with tied scores.
|
| 801 |
+
X_train = np.array([[0, 0, 0], [1, 1, 1]])
|
| 802 |
+
y_train = [0, 1]
|
| 803 |
+
|
| 804 |
+
for n_features in [1, 2, 3]:
|
| 805 |
+
sel = SelectKBest(chi2, k=n_features).fit(X_train, y_train)
|
| 806 |
+
X_test = sel.transform([[0, 1, 2]])
|
| 807 |
+
assert_array_equal(X_test[0], np.arange(3)[-n_features:])
|
| 808 |
+
|
| 809 |
+
|
| 810 |
+
def test_nans():
|
| 811 |
+
# Assert that SelectKBest and SelectPercentile can handle NaNs.
|
| 812 |
+
# First feature has zero variance to confuse f_classif (ANOVA) and
|
| 813 |
+
# make it return a NaN.
|
| 814 |
+
X = [[0, 1, 0], [0, -1, -1], [0, 0.5, 0.5]]
|
| 815 |
+
y = [1, 0, 1]
|
| 816 |
+
|
| 817 |
+
for select in (
|
| 818 |
+
SelectKBest(f_classif, k=2),
|
| 819 |
+
SelectPercentile(f_classif, percentile=67),
|
| 820 |
+
):
|
| 821 |
+
ignore_warnings(select.fit)(X, y)
|
| 822 |
+
assert_array_equal(select.get_support(indices=True), np.array([1, 2]))
|
| 823 |
+
|
| 824 |
+
|
| 825 |
+
def test_invalid_k():
|
| 826 |
+
X = [[0, 1, 0], [0, -1, -1], [0, 0.5, 0.5]]
|
| 827 |
+
y = [1, 0, 1]
|
| 828 |
+
|
| 829 |
+
with pytest.raises(ValueError):
|
| 830 |
+
SelectKBest(k=4).fit(X, y)
|
| 831 |
+
with pytest.raises(ValueError):
|
| 832 |
+
GenericUnivariateSelect(mode="k_best", param=4).fit(X, y)
|
| 833 |
+
|
| 834 |
+
|
| 835 |
+
def test_f_classif_constant_feature():
|
| 836 |
+
# Test that f_classif warns if a feature is constant throughout.
|
| 837 |
+
|
| 838 |
+
X, y = make_classification(n_samples=10, n_features=5)
|
| 839 |
+
X[:, 0] = 2.0
|
| 840 |
+
with pytest.warns(UserWarning):
|
| 841 |
+
f_classif(X, y)
|
| 842 |
+
|
| 843 |
+
|
| 844 |
+
def test_no_feature_selected():
|
| 845 |
+
rng = np.random.RandomState(0)
|
| 846 |
+
|
| 847 |
+
# Generate random uncorrelated data: a strict univariate test should
|
| 848 |
+
# rejects all the features
|
| 849 |
+
X = rng.rand(40, 10)
|
| 850 |
+
y = rng.randint(0, 4, size=40)
|
| 851 |
+
strict_selectors = [
|
| 852 |
+
SelectFwe(alpha=0.01).fit(X, y),
|
| 853 |
+
SelectFdr(alpha=0.01).fit(X, y),
|
| 854 |
+
SelectFpr(alpha=0.01).fit(X, y),
|
| 855 |
+
SelectPercentile(percentile=0).fit(X, y),
|
| 856 |
+
SelectKBest(k=0).fit(X, y),
|
| 857 |
+
]
|
| 858 |
+
for selector in strict_selectors:
|
| 859 |
+
assert_array_equal(selector.get_support(), np.zeros(10))
|
| 860 |
+
with pytest.warns(UserWarning, match="No features were selected"):
|
| 861 |
+
X_selected = selector.transform(X)
|
| 862 |
+
assert X_selected.shape == (40, 0)
|
| 863 |
+
|
| 864 |
+
|
| 865 |
+
def test_mutual_info_classif():
|
| 866 |
+
X, y = make_classification(
|
| 867 |
+
n_samples=100,
|
| 868 |
+
n_features=5,
|
| 869 |
+
n_informative=1,
|
| 870 |
+
n_redundant=1,
|
| 871 |
+
n_repeated=0,
|
| 872 |
+
n_classes=2,
|
| 873 |
+
n_clusters_per_class=1,
|
| 874 |
+
flip_y=0.0,
|
| 875 |
+
class_sep=10,
|
| 876 |
+
shuffle=False,
|
| 877 |
+
random_state=0,
|
| 878 |
+
)
|
| 879 |
+
|
| 880 |
+
# Test in KBest mode.
|
| 881 |
+
univariate_filter = SelectKBest(mutual_info_classif, k=2)
|
| 882 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 883 |
+
X_r2 = (
|
| 884 |
+
GenericUnivariateSelect(mutual_info_classif, mode="k_best", param=2)
|
| 885 |
+
.fit(X, y)
|
| 886 |
+
.transform(X)
|
| 887 |
+
)
|
| 888 |
+
assert_array_equal(X_r, X_r2)
|
| 889 |
+
support = univariate_filter.get_support()
|
| 890 |
+
gtruth = np.zeros(5)
|
| 891 |
+
gtruth[:2] = 1
|
| 892 |
+
assert_array_equal(support, gtruth)
|
| 893 |
+
|
| 894 |
+
# Test in Percentile mode.
|
| 895 |
+
univariate_filter = SelectPercentile(mutual_info_classif, percentile=40)
|
| 896 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 897 |
+
X_r2 = (
|
| 898 |
+
GenericUnivariateSelect(mutual_info_classif, mode="percentile", param=40)
|
| 899 |
+
.fit(X, y)
|
| 900 |
+
.transform(X)
|
| 901 |
+
)
|
| 902 |
+
assert_array_equal(X_r, X_r2)
|
| 903 |
+
support = univariate_filter.get_support()
|
| 904 |
+
gtruth = np.zeros(5)
|
| 905 |
+
gtruth[:2] = 1
|
| 906 |
+
assert_array_equal(support, gtruth)
|
| 907 |
+
|
| 908 |
+
|
| 909 |
+
def test_mutual_info_regression():
|
| 910 |
+
X, y = make_regression(
|
| 911 |
+
n_samples=100,
|
| 912 |
+
n_features=10,
|
| 913 |
+
n_informative=2,
|
| 914 |
+
shuffle=False,
|
| 915 |
+
random_state=0,
|
| 916 |
+
noise=10,
|
| 917 |
+
)
|
| 918 |
+
|
| 919 |
+
# Test in KBest mode.
|
| 920 |
+
univariate_filter = SelectKBest(mutual_info_regression, k=2)
|
| 921 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 922 |
+
assert_best_scores_kept(univariate_filter)
|
| 923 |
+
X_r2 = (
|
| 924 |
+
GenericUnivariateSelect(mutual_info_regression, mode="k_best", param=2)
|
| 925 |
+
.fit(X, y)
|
| 926 |
+
.transform(X)
|
| 927 |
+
)
|
| 928 |
+
assert_array_equal(X_r, X_r2)
|
| 929 |
+
support = univariate_filter.get_support()
|
| 930 |
+
gtruth = np.zeros(10)
|
| 931 |
+
gtruth[:2] = 1
|
| 932 |
+
assert_array_equal(support, gtruth)
|
| 933 |
+
|
| 934 |
+
# Test in Percentile mode.
|
| 935 |
+
univariate_filter = SelectPercentile(mutual_info_regression, percentile=20)
|
| 936 |
+
X_r = univariate_filter.fit(X, y).transform(X)
|
| 937 |
+
X_r2 = (
|
| 938 |
+
GenericUnivariateSelect(mutual_info_regression, mode="percentile", param=20)
|
| 939 |
+
.fit(X, y)
|
| 940 |
+
.transform(X)
|
| 941 |
+
)
|
| 942 |
+
assert_array_equal(X_r, X_r2)
|
| 943 |
+
support = univariate_filter.get_support()
|
| 944 |
+
gtruth = np.zeros(10)
|
| 945 |
+
gtruth[:2] = 1
|
| 946 |
+
assert_array_equal(support, gtruth)
|
videollama2/lib/python3.10/site-packages/sklearn/tree/_classes.py
ADDED
|
@@ -0,0 +1,1769 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
This module gathers tree-based methods, including decision, regression and
|
| 3 |
+
randomized trees. Single and multi-output problems are both handled.
|
| 4 |
+
"""
|
| 5 |
+
|
| 6 |
+
# Authors: Gilles Louppe <g.louppe@gmail.com>
|
| 7 |
+
# Peter Prettenhofer <peter.prettenhofer@gmail.com>
|
| 8 |
+
# Brian Holt <bdholt1@gmail.com>
|
| 9 |
+
# Noel Dawe <noel@dawe.me>
|
| 10 |
+
# Satrajit Gosh <satrajit.ghosh@gmail.com>
|
| 11 |
+
# Joly Arnaud <arnaud.v.joly@gmail.com>
|
| 12 |
+
# Fares Hedayati <fares.hedayati@gmail.com>
|
| 13 |
+
# Nelson Liu <nelson@nelsonliu.me>
|
| 14 |
+
#
|
| 15 |
+
# License: BSD 3 clause
|
| 16 |
+
|
| 17 |
+
import numbers
|
| 18 |
+
import warnings
|
| 19 |
+
import copy
|
| 20 |
+
from abc import ABCMeta
|
| 21 |
+
from abc import abstractmethod
|
| 22 |
+
from math import ceil
|
| 23 |
+
from numbers import Integral, Real
|
| 24 |
+
|
| 25 |
+
import numpy as np
|
| 26 |
+
from scipy.sparse import issparse
|
| 27 |
+
|
| 28 |
+
from ..base import BaseEstimator
|
| 29 |
+
from ..base import ClassifierMixin
|
| 30 |
+
from ..base import clone
|
| 31 |
+
from ..base import RegressorMixin
|
| 32 |
+
from ..base import is_classifier
|
| 33 |
+
from ..base import MultiOutputMixin
|
| 34 |
+
from ..utils import Bunch
|
| 35 |
+
from ..utils import check_random_state
|
| 36 |
+
from ..utils.validation import _check_sample_weight
|
| 37 |
+
from ..utils import compute_sample_weight
|
| 38 |
+
from ..utils.multiclass import check_classification_targets
|
| 39 |
+
from ..utils.validation import check_is_fitted
|
| 40 |
+
from ..utils._param_validation import Hidden, Interval, StrOptions
|
| 41 |
+
|
| 42 |
+
from ._criterion import Criterion
|
| 43 |
+
from ._splitter import Splitter
|
| 44 |
+
from ._tree import DepthFirstTreeBuilder
|
| 45 |
+
from ._tree import BestFirstTreeBuilder
|
| 46 |
+
from ._tree import Tree
|
| 47 |
+
from ._tree import _build_pruned_tree_ccp
|
| 48 |
+
from ._tree import ccp_pruning_path
|
| 49 |
+
from . import _tree, _splitter, _criterion
|
| 50 |
+
|
| 51 |
+
__all__ = [
|
| 52 |
+
"DecisionTreeClassifier",
|
| 53 |
+
"DecisionTreeRegressor",
|
| 54 |
+
"ExtraTreeClassifier",
|
| 55 |
+
"ExtraTreeRegressor",
|
| 56 |
+
]
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
# =============================================================================
|
| 60 |
+
# Types and constants
|
| 61 |
+
# =============================================================================
|
| 62 |
+
|
| 63 |
+
DTYPE = _tree.DTYPE
|
| 64 |
+
DOUBLE = _tree.DOUBLE
|
| 65 |
+
|
| 66 |
+
CRITERIA_CLF = {
|
| 67 |
+
"gini": _criterion.Gini,
|
| 68 |
+
"log_loss": _criterion.Entropy,
|
| 69 |
+
"entropy": _criterion.Entropy,
|
| 70 |
+
}
|
| 71 |
+
CRITERIA_REG = {
|
| 72 |
+
"squared_error": _criterion.MSE,
|
| 73 |
+
"friedman_mse": _criterion.FriedmanMSE,
|
| 74 |
+
"absolute_error": _criterion.MAE,
|
| 75 |
+
"poisson": _criterion.Poisson,
|
| 76 |
+
}
|
| 77 |
+
|
| 78 |
+
DENSE_SPLITTERS = {"best": _splitter.BestSplitter, "random": _splitter.RandomSplitter}
|
| 79 |
+
|
| 80 |
+
SPARSE_SPLITTERS = {
|
| 81 |
+
"best": _splitter.BestSparseSplitter,
|
| 82 |
+
"random": _splitter.RandomSparseSplitter,
|
| 83 |
+
}
|
| 84 |
+
|
| 85 |
+
# =============================================================================
|
| 86 |
+
# Base decision tree
|
| 87 |
+
# =============================================================================
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
class BaseDecisionTree(MultiOutputMixin, BaseEstimator, metaclass=ABCMeta):
|
| 91 |
+
"""Base class for decision trees.
|
| 92 |
+
|
| 93 |
+
Warning: This class should not be used directly.
|
| 94 |
+
Use derived classes instead.
|
| 95 |
+
"""
|
| 96 |
+
|
| 97 |
+
_parameter_constraints: dict = {
|
| 98 |
+
"splitter": [StrOptions({"best", "random"})],
|
| 99 |
+
"max_depth": [Interval(Integral, 1, None, closed="left"), None],
|
| 100 |
+
"min_samples_split": [
|
| 101 |
+
Interval(Integral, 2, None, closed="left"),
|
| 102 |
+
Interval("real_not_int", 0.0, 1.0, closed="right"),
|
| 103 |
+
],
|
| 104 |
+
"min_samples_leaf": [
|
| 105 |
+
Interval(Integral, 1, None, closed="left"),
|
| 106 |
+
Interval("real_not_int", 0.0, 1.0, closed="neither"),
|
| 107 |
+
],
|
| 108 |
+
"min_weight_fraction_leaf": [Interval(Real, 0.0, 0.5, closed="both")],
|
| 109 |
+
"max_features": [
|
| 110 |
+
Interval(Integral, 1, None, closed="left"),
|
| 111 |
+
Interval("real_not_int", 0.0, 1.0, closed="right"),
|
| 112 |
+
StrOptions({"auto", "sqrt", "log2"}, deprecated={"auto"}),
|
| 113 |
+
None,
|
| 114 |
+
],
|
| 115 |
+
"random_state": ["random_state"],
|
| 116 |
+
"max_leaf_nodes": [Interval(Integral, 2, None, closed="left"), None],
|
| 117 |
+
"min_impurity_decrease": [Interval(Real, 0.0, None, closed="left")],
|
| 118 |
+
"ccp_alpha": [Interval(Real, 0.0, None, closed="left")],
|
| 119 |
+
}
|
| 120 |
+
|
| 121 |
+
@abstractmethod
|
| 122 |
+
def __init__(
|
| 123 |
+
self,
|
| 124 |
+
*,
|
| 125 |
+
criterion,
|
| 126 |
+
splitter,
|
| 127 |
+
max_depth,
|
| 128 |
+
min_samples_split,
|
| 129 |
+
min_samples_leaf,
|
| 130 |
+
min_weight_fraction_leaf,
|
| 131 |
+
max_features,
|
| 132 |
+
max_leaf_nodes,
|
| 133 |
+
random_state,
|
| 134 |
+
min_impurity_decrease,
|
| 135 |
+
class_weight=None,
|
| 136 |
+
ccp_alpha=0.0,
|
| 137 |
+
):
|
| 138 |
+
self.criterion = criterion
|
| 139 |
+
self.splitter = splitter
|
| 140 |
+
self.max_depth = max_depth
|
| 141 |
+
self.min_samples_split = min_samples_split
|
| 142 |
+
self.min_samples_leaf = min_samples_leaf
|
| 143 |
+
self.min_weight_fraction_leaf = min_weight_fraction_leaf
|
| 144 |
+
self.max_features = max_features
|
| 145 |
+
self.max_leaf_nodes = max_leaf_nodes
|
| 146 |
+
self.random_state = random_state
|
| 147 |
+
self.min_impurity_decrease = min_impurity_decrease
|
| 148 |
+
self.class_weight = class_weight
|
| 149 |
+
self.ccp_alpha = ccp_alpha
|
| 150 |
+
|
| 151 |
+
def get_depth(self):
|
| 152 |
+
"""Return the depth of the decision tree.
|
| 153 |
+
|
| 154 |
+
The depth of a tree is the maximum distance between the root
|
| 155 |
+
and any leaf.
|
| 156 |
+
|
| 157 |
+
Returns
|
| 158 |
+
-------
|
| 159 |
+
self.tree_.max_depth : int
|
| 160 |
+
The maximum depth of the tree.
|
| 161 |
+
"""
|
| 162 |
+
check_is_fitted(self)
|
| 163 |
+
return self.tree_.max_depth
|
| 164 |
+
|
| 165 |
+
def get_n_leaves(self):
|
| 166 |
+
"""Return the number of leaves of the decision tree.
|
| 167 |
+
|
| 168 |
+
Returns
|
| 169 |
+
-------
|
| 170 |
+
self.tree_.n_leaves : int
|
| 171 |
+
Number of leaves.
|
| 172 |
+
"""
|
| 173 |
+
check_is_fitted(self)
|
| 174 |
+
return self.tree_.n_leaves
|
| 175 |
+
|
| 176 |
+
def fit(self, X, y, sample_weight=None, check_input=True):
|
| 177 |
+
self._validate_params()
|
| 178 |
+
random_state = check_random_state(self.random_state)
|
| 179 |
+
|
| 180 |
+
if check_input:
|
| 181 |
+
# Need to validate separately here.
|
| 182 |
+
# We can't pass multi_output=True because that would allow y to be
|
| 183 |
+
# csr.
|
| 184 |
+
check_X_params = dict(dtype=DTYPE, accept_sparse="csc")
|
| 185 |
+
check_y_params = dict(ensure_2d=False, dtype=None)
|
| 186 |
+
X, y = self._validate_data(
|
| 187 |
+
X, y, validate_separately=(check_X_params, check_y_params)
|
| 188 |
+
)
|
| 189 |
+
if issparse(X):
|
| 190 |
+
X.sort_indices()
|
| 191 |
+
|
| 192 |
+
if X.indices.dtype != np.intc or X.indptr.dtype != np.intc:
|
| 193 |
+
raise ValueError(
|
| 194 |
+
"No support for np.int64 index based sparse matrices"
|
| 195 |
+
)
|
| 196 |
+
|
| 197 |
+
if self.criterion == "poisson":
|
| 198 |
+
if np.any(y < 0):
|
| 199 |
+
raise ValueError(
|
| 200 |
+
"Some value(s) of y are negative which is"
|
| 201 |
+
" not allowed for Poisson regression."
|
| 202 |
+
)
|
| 203 |
+
if np.sum(y) <= 0:
|
| 204 |
+
raise ValueError(
|
| 205 |
+
"Sum of y is not positive which is "
|
| 206 |
+
"necessary for Poisson regression."
|
| 207 |
+
)
|
| 208 |
+
|
| 209 |
+
# Determine output settings
|
| 210 |
+
n_samples, self.n_features_in_ = X.shape
|
| 211 |
+
is_classification = is_classifier(self)
|
| 212 |
+
|
| 213 |
+
y = np.atleast_1d(y)
|
| 214 |
+
expanded_class_weight = None
|
| 215 |
+
|
| 216 |
+
if y.ndim == 1:
|
| 217 |
+
# reshape is necessary to preserve the data contiguity against vs
|
| 218 |
+
# [:, np.newaxis] that does not.
|
| 219 |
+
y = np.reshape(y, (-1, 1))
|
| 220 |
+
|
| 221 |
+
self.n_outputs_ = y.shape[1]
|
| 222 |
+
|
| 223 |
+
if is_classification:
|
| 224 |
+
check_classification_targets(y)
|
| 225 |
+
y = np.copy(y)
|
| 226 |
+
|
| 227 |
+
self.classes_ = []
|
| 228 |
+
self.n_classes_ = []
|
| 229 |
+
|
| 230 |
+
if self.class_weight is not None:
|
| 231 |
+
y_original = np.copy(y)
|
| 232 |
+
|
| 233 |
+
y_encoded = np.zeros(y.shape, dtype=int)
|
| 234 |
+
for k in range(self.n_outputs_):
|
| 235 |
+
classes_k, y_encoded[:, k] = np.unique(y[:, k], return_inverse=True)
|
| 236 |
+
self.classes_.append(classes_k)
|
| 237 |
+
self.n_classes_.append(classes_k.shape[0])
|
| 238 |
+
y = y_encoded
|
| 239 |
+
|
| 240 |
+
if self.class_weight is not None:
|
| 241 |
+
expanded_class_weight = compute_sample_weight(
|
| 242 |
+
self.class_weight, y_original
|
| 243 |
+
)
|
| 244 |
+
|
| 245 |
+
self.n_classes_ = np.array(self.n_classes_, dtype=np.intp)
|
| 246 |
+
|
| 247 |
+
if getattr(y, "dtype", None) != DOUBLE or not y.flags.contiguous:
|
| 248 |
+
y = np.ascontiguousarray(y, dtype=DOUBLE)
|
| 249 |
+
|
| 250 |
+
max_depth = np.iinfo(np.int32).max if self.max_depth is None else self.max_depth
|
| 251 |
+
|
| 252 |
+
if isinstance(self.min_samples_leaf, numbers.Integral):
|
| 253 |
+
min_samples_leaf = self.min_samples_leaf
|
| 254 |
+
else: # float
|
| 255 |
+
min_samples_leaf = int(ceil(self.min_samples_leaf * n_samples))
|
| 256 |
+
|
| 257 |
+
if isinstance(self.min_samples_split, numbers.Integral):
|
| 258 |
+
min_samples_split = self.min_samples_split
|
| 259 |
+
else: # float
|
| 260 |
+
min_samples_split = int(ceil(self.min_samples_split * n_samples))
|
| 261 |
+
min_samples_split = max(2, min_samples_split)
|
| 262 |
+
|
| 263 |
+
min_samples_split = max(min_samples_split, 2 * min_samples_leaf)
|
| 264 |
+
|
| 265 |
+
if isinstance(self.max_features, str):
|
| 266 |
+
if self.max_features == "auto":
|
| 267 |
+
if is_classification:
|
| 268 |
+
max_features = max(1, int(np.sqrt(self.n_features_in_)))
|
| 269 |
+
warnings.warn(
|
| 270 |
+
"`max_features='auto'` has been deprecated in 1.1 "
|
| 271 |
+
"and will be removed in 1.3. To keep the past behaviour, "
|
| 272 |
+
"explicitly set `max_features='sqrt'`.",
|
| 273 |
+
FutureWarning,
|
| 274 |
+
)
|
| 275 |
+
else:
|
| 276 |
+
max_features = self.n_features_in_
|
| 277 |
+
warnings.warn(
|
| 278 |
+
"`max_features='auto'` has been deprecated in 1.1 "
|
| 279 |
+
"and will be removed in 1.3. To keep the past behaviour, "
|
| 280 |
+
"explicitly set `max_features=1.0'`.",
|
| 281 |
+
FutureWarning,
|
| 282 |
+
)
|
| 283 |
+
elif self.max_features == "sqrt":
|
| 284 |
+
max_features = max(1, int(np.sqrt(self.n_features_in_)))
|
| 285 |
+
elif self.max_features == "log2":
|
| 286 |
+
max_features = max(1, int(np.log2(self.n_features_in_)))
|
| 287 |
+
elif self.max_features is None:
|
| 288 |
+
max_features = self.n_features_in_
|
| 289 |
+
elif isinstance(self.max_features, numbers.Integral):
|
| 290 |
+
max_features = self.max_features
|
| 291 |
+
else: # float
|
| 292 |
+
if self.max_features > 0.0:
|
| 293 |
+
max_features = max(1, int(self.max_features * self.n_features_in_))
|
| 294 |
+
else:
|
| 295 |
+
max_features = 0
|
| 296 |
+
|
| 297 |
+
self.max_features_ = max_features
|
| 298 |
+
|
| 299 |
+
max_leaf_nodes = -1 if self.max_leaf_nodes is None else self.max_leaf_nodes
|
| 300 |
+
|
| 301 |
+
if len(y) != n_samples:
|
| 302 |
+
raise ValueError(
|
| 303 |
+
"Number of labels=%d does not match number of samples=%d"
|
| 304 |
+
% (len(y), n_samples)
|
| 305 |
+
)
|
| 306 |
+
|
| 307 |
+
if sample_weight is not None:
|
| 308 |
+
sample_weight = _check_sample_weight(sample_weight, X, DOUBLE)
|
| 309 |
+
|
| 310 |
+
if expanded_class_weight is not None:
|
| 311 |
+
if sample_weight is not None:
|
| 312 |
+
sample_weight = sample_weight * expanded_class_weight
|
| 313 |
+
else:
|
| 314 |
+
sample_weight = expanded_class_weight
|
| 315 |
+
|
| 316 |
+
# Set min_weight_leaf from min_weight_fraction_leaf
|
| 317 |
+
if sample_weight is None:
|
| 318 |
+
min_weight_leaf = self.min_weight_fraction_leaf * n_samples
|
| 319 |
+
else:
|
| 320 |
+
min_weight_leaf = self.min_weight_fraction_leaf * np.sum(sample_weight)
|
| 321 |
+
|
| 322 |
+
# Build tree
|
| 323 |
+
criterion = self.criterion
|
| 324 |
+
if not isinstance(criterion, Criterion):
|
| 325 |
+
if is_classification:
|
| 326 |
+
criterion = CRITERIA_CLF[self.criterion](
|
| 327 |
+
self.n_outputs_, self.n_classes_
|
| 328 |
+
)
|
| 329 |
+
else:
|
| 330 |
+
criterion = CRITERIA_REG[self.criterion](self.n_outputs_, n_samples)
|
| 331 |
+
else:
|
| 332 |
+
# Make a deepcopy in case the criterion has mutable attributes that
|
| 333 |
+
# might be shared and modified concurrently during parallel fitting
|
| 334 |
+
criterion = copy.deepcopy(criterion)
|
| 335 |
+
|
| 336 |
+
SPLITTERS = SPARSE_SPLITTERS if issparse(X) else DENSE_SPLITTERS
|
| 337 |
+
|
| 338 |
+
splitter = self.splitter
|
| 339 |
+
if not isinstance(self.splitter, Splitter):
|
| 340 |
+
splitter = SPLITTERS[self.splitter](
|
| 341 |
+
criterion,
|
| 342 |
+
self.max_features_,
|
| 343 |
+
min_samples_leaf,
|
| 344 |
+
min_weight_leaf,
|
| 345 |
+
random_state,
|
| 346 |
+
)
|
| 347 |
+
|
| 348 |
+
if is_classifier(self):
|
| 349 |
+
self.tree_ = Tree(self.n_features_in_, self.n_classes_, self.n_outputs_)
|
| 350 |
+
else:
|
| 351 |
+
self.tree_ = Tree(
|
| 352 |
+
self.n_features_in_,
|
| 353 |
+
# TODO: tree shouldn't need this in this case
|
| 354 |
+
np.array([1] * self.n_outputs_, dtype=np.intp),
|
| 355 |
+
self.n_outputs_,
|
| 356 |
+
)
|
| 357 |
+
|
| 358 |
+
# Use BestFirst if max_leaf_nodes given; use DepthFirst otherwise
|
| 359 |
+
if max_leaf_nodes < 0:
|
| 360 |
+
builder = DepthFirstTreeBuilder(
|
| 361 |
+
splitter,
|
| 362 |
+
min_samples_split,
|
| 363 |
+
min_samples_leaf,
|
| 364 |
+
min_weight_leaf,
|
| 365 |
+
max_depth,
|
| 366 |
+
self.min_impurity_decrease,
|
| 367 |
+
)
|
| 368 |
+
else:
|
| 369 |
+
builder = BestFirstTreeBuilder(
|
| 370 |
+
splitter,
|
| 371 |
+
min_samples_split,
|
| 372 |
+
min_samples_leaf,
|
| 373 |
+
min_weight_leaf,
|
| 374 |
+
max_depth,
|
| 375 |
+
max_leaf_nodes,
|
| 376 |
+
self.min_impurity_decrease,
|
| 377 |
+
)
|
| 378 |
+
|
| 379 |
+
builder.build(self.tree_, X, y, sample_weight)
|
| 380 |
+
|
| 381 |
+
if self.n_outputs_ == 1 and is_classifier(self):
|
| 382 |
+
self.n_classes_ = self.n_classes_[0]
|
| 383 |
+
self.classes_ = self.classes_[0]
|
| 384 |
+
|
| 385 |
+
self._prune_tree()
|
| 386 |
+
|
| 387 |
+
return self
|
| 388 |
+
|
| 389 |
+
def _validate_X_predict(self, X, check_input):
|
| 390 |
+
"""Validate the training data on predict (probabilities)."""
|
| 391 |
+
if check_input:
|
| 392 |
+
X = self._validate_data(X, dtype=DTYPE, accept_sparse="csr", reset=False)
|
| 393 |
+
if issparse(X) and (
|
| 394 |
+
X.indices.dtype != np.intc or X.indptr.dtype != np.intc
|
| 395 |
+
):
|
| 396 |
+
raise ValueError("No support for np.int64 index based sparse matrices")
|
| 397 |
+
else:
|
| 398 |
+
# The number of features is checked regardless of `check_input`
|
| 399 |
+
self._check_n_features(X, reset=False)
|
| 400 |
+
return X
|
| 401 |
+
|
| 402 |
+
def predict(self, X, check_input=True):
|
| 403 |
+
"""Predict class or regression value for X.
|
| 404 |
+
|
| 405 |
+
For a classification model, the predicted class for each sample in X is
|
| 406 |
+
returned. For a regression model, the predicted value based on X is
|
| 407 |
+
returned.
|
| 408 |
+
|
| 409 |
+
Parameters
|
| 410 |
+
----------
|
| 411 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 412 |
+
The input samples. Internally, it will be converted to
|
| 413 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 414 |
+
to a sparse ``csr_matrix``.
|
| 415 |
+
|
| 416 |
+
check_input : bool, default=True
|
| 417 |
+
Allow to bypass several input checking.
|
| 418 |
+
Don't use this parameter unless you know what you're doing.
|
| 419 |
+
|
| 420 |
+
Returns
|
| 421 |
+
-------
|
| 422 |
+
y : array-like of shape (n_samples,) or (n_samples, n_outputs)
|
| 423 |
+
The predicted classes, or the predict values.
|
| 424 |
+
"""
|
| 425 |
+
check_is_fitted(self)
|
| 426 |
+
X = self._validate_X_predict(X, check_input)
|
| 427 |
+
proba = self.tree_.predict(X)
|
| 428 |
+
n_samples = X.shape[0]
|
| 429 |
+
|
| 430 |
+
# Classification
|
| 431 |
+
if is_classifier(self):
|
| 432 |
+
if self.n_outputs_ == 1:
|
| 433 |
+
return self.classes_.take(np.argmax(proba, axis=1), axis=0)
|
| 434 |
+
|
| 435 |
+
else:
|
| 436 |
+
class_type = self.classes_[0].dtype
|
| 437 |
+
predictions = np.zeros((n_samples, self.n_outputs_), dtype=class_type)
|
| 438 |
+
for k in range(self.n_outputs_):
|
| 439 |
+
predictions[:, k] = self.classes_[k].take(
|
| 440 |
+
np.argmax(proba[:, k], axis=1), axis=0
|
| 441 |
+
)
|
| 442 |
+
|
| 443 |
+
return predictions
|
| 444 |
+
|
| 445 |
+
# Regression
|
| 446 |
+
else:
|
| 447 |
+
if self.n_outputs_ == 1:
|
| 448 |
+
return proba[:, 0]
|
| 449 |
+
|
| 450 |
+
else:
|
| 451 |
+
return proba[:, :, 0]
|
| 452 |
+
|
| 453 |
+
def apply(self, X, check_input=True):
|
| 454 |
+
"""Return the index of the leaf that each sample is predicted as.
|
| 455 |
+
|
| 456 |
+
.. versionadded:: 0.17
|
| 457 |
+
|
| 458 |
+
Parameters
|
| 459 |
+
----------
|
| 460 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 461 |
+
The input samples. Internally, it will be converted to
|
| 462 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 463 |
+
to a sparse ``csr_matrix``.
|
| 464 |
+
|
| 465 |
+
check_input : bool, default=True
|
| 466 |
+
Allow to bypass several input checking.
|
| 467 |
+
Don't use this parameter unless you know what you're doing.
|
| 468 |
+
|
| 469 |
+
Returns
|
| 470 |
+
-------
|
| 471 |
+
X_leaves : array-like of shape (n_samples,)
|
| 472 |
+
For each datapoint x in X, return the index of the leaf x
|
| 473 |
+
ends up in. Leaves are numbered within
|
| 474 |
+
``[0; self.tree_.node_count)``, possibly with gaps in the
|
| 475 |
+
numbering.
|
| 476 |
+
"""
|
| 477 |
+
check_is_fitted(self)
|
| 478 |
+
X = self._validate_X_predict(X, check_input)
|
| 479 |
+
return self.tree_.apply(X)
|
| 480 |
+
|
| 481 |
+
def decision_path(self, X, check_input=True):
|
| 482 |
+
"""Return the decision path in the tree.
|
| 483 |
+
|
| 484 |
+
.. versionadded:: 0.18
|
| 485 |
+
|
| 486 |
+
Parameters
|
| 487 |
+
----------
|
| 488 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 489 |
+
The input samples. Internally, it will be converted to
|
| 490 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 491 |
+
to a sparse ``csr_matrix``.
|
| 492 |
+
|
| 493 |
+
check_input : bool, default=True
|
| 494 |
+
Allow to bypass several input checking.
|
| 495 |
+
Don't use this parameter unless you know what you're doing.
|
| 496 |
+
|
| 497 |
+
Returns
|
| 498 |
+
-------
|
| 499 |
+
indicator : sparse matrix of shape (n_samples, n_nodes)
|
| 500 |
+
Return a node indicator CSR matrix where non zero elements
|
| 501 |
+
indicates that the samples goes through the nodes.
|
| 502 |
+
"""
|
| 503 |
+
X = self._validate_X_predict(X, check_input)
|
| 504 |
+
return self.tree_.decision_path(X)
|
| 505 |
+
|
| 506 |
+
def _prune_tree(self):
|
| 507 |
+
"""Prune tree using Minimal Cost-Complexity Pruning."""
|
| 508 |
+
check_is_fitted(self)
|
| 509 |
+
|
| 510 |
+
if self.ccp_alpha == 0.0:
|
| 511 |
+
return
|
| 512 |
+
|
| 513 |
+
# build pruned tree
|
| 514 |
+
if is_classifier(self):
|
| 515 |
+
n_classes = np.atleast_1d(self.n_classes_)
|
| 516 |
+
pruned_tree = Tree(self.n_features_in_, n_classes, self.n_outputs_)
|
| 517 |
+
else:
|
| 518 |
+
pruned_tree = Tree(
|
| 519 |
+
self.n_features_in_,
|
| 520 |
+
# TODO: the tree shouldn't need this param
|
| 521 |
+
np.array([1] * self.n_outputs_, dtype=np.intp),
|
| 522 |
+
self.n_outputs_,
|
| 523 |
+
)
|
| 524 |
+
_build_pruned_tree_ccp(pruned_tree, self.tree_, self.ccp_alpha)
|
| 525 |
+
|
| 526 |
+
self.tree_ = pruned_tree
|
| 527 |
+
|
| 528 |
+
def cost_complexity_pruning_path(self, X, y, sample_weight=None):
|
| 529 |
+
"""Compute the pruning path during Minimal Cost-Complexity Pruning.
|
| 530 |
+
|
| 531 |
+
See :ref:`minimal_cost_complexity_pruning` for details on the pruning
|
| 532 |
+
process.
|
| 533 |
+
|
| 534 |
+
Parameters
|
| 535 |
+
----------
|
| 536 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 537 |
+
The training input samples. Internally, it will be converted to
|
| 538 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 539 |
+
to a sparse ``csc_matrix``.
|
| 540 |
+
|
| 541 |
+
y : array-like of shape (n_samples,) or (n_samples, n_outputs)
|
| 542 |
+
The target values (class labels) as integers or strings.
|
| 543 |
+
|
| 544 |
+
sample_weight : array-like of shape (n_samples,), default=None
|
| 545 |
+
Sample weights. If None, then samples are equally weighted. Splits
|
| 546 |
+
that would create child nodes with net zero or negative weight are
|
| 547 |
+
ignored while searching for a split in each node. Splits are also
|
| 548 |
+
ignored if they would result in any single class carrying a
|
| 549 |
+
negative weight in either child node.
|
| 550 |
+
|
| 551 |
+
Returns
|
| 552 |
+
-------
|
| 553 |
+
ccp_path : :class:`~sklearn.utils.Bunch`
|
| 554 |
+
Dictionary-like object, with the following attributes.
|
| 555 |
+
|
| 556 |
+
ccp_alphas : ndarray
|
| 557 |
+
Effective alphas of subtree during pruning.
|
| 558 |
+
|
| 559 |
+
impurities : ndarray
|
| 560 |
+
Sum of the impurities of the subtree leaves for the
|
| 561 |
+
corresponding alpha value in ``ccp_alphas``.
|
| 562 |
+
"""
|
| 563 |
+
est = clone(self).set_params(ccp_alpha=0.0)
|
| 564 |
+
est.fit(X, y, sample_weight=sample_weight)
|
| 565 |
+
return Bunch(**ccp_pruning_path(est.tree_))
|
| 566 |
+
|
| 567 |
+
@property
|
| 568 |
+
def feature_importances_(self):
|
| 569 |
+
"""Return the feature importances.
|
| 570 |
+
|
| 571 |
+
The importance of a feature is computed as the (normalized) total
|
| 572 |
+
reduction of the criterion brought by that feature.
|
| 573 |
+
It is also known as the Gini importance.
|
| 574 |
+
|
| 575 |
+
Warning: impurity-based feature importances can be misleading for
|
| 576 |
+
high cardinality features (many unique values). See
|
| 577 |
+
:func:`sklearn.inspection.permutation_importance` as an alternative.
|
| 578 |
+
|
| 579 |
+
Returns
|
| 580 |
+
-------
|
| 581 |
+
feature_importances_ : ndarray of shape (n_features,)
|
| 582 |
+
Normalized total reduction of criteria by feature
|
| 583 |
+
(Gini importance).
|
| 584 |
+
"""
|
| 585 |
+
check_is_fitted(self)
|
| 586 |
+
|
| 587 |
+
return self.tree_.compute_feature_importances()
|
| 588 |
+
|
| 589 |
+
|
| 590 |
+
# =============================================================================
|
| 591 |
+
# Public estimators
|
| 592 |
+
# =============================================================================
|
| 593 |
+
|
| 594 |
+
|
| 595 |
+
class DecisionTreeClassifier(ClassifierMixin, BaseDecisionTree):
|
| 596 |
+
"""A decision tree classifier.
|
| 597 |
+
|
| 598 |
+
Read more in the :ref:`User Guide <tree>`.
|
| 599 |
+
|
| 600 |
+
Parameters
|
| 601 |
+
----------
|
| 602 |
+
criterion : {"gini", "entropy", "log_loss"}, default="gini"
|
| 603 |
+
The function to measure the quality of a split. Supported criteria are
|
| 604 |
+
"gini" for the Gini impurity and "log_loss" and "entropy" both for the
|
| 605 |
+
Shannon information gain, see :ref:`tree_mathematical_formulation`.
|
| 606 |
+
|
| 607 |
+
splitter : {"best", "random"}, default="best"
|
| 608 |
+
The strategy used to choose the split at each node. Supported
|
| 609 |
+
strategies are "best" to choose the best split and "random" to choose
|
| 610 |
+
the best random split.
|
| 611 |
+
|
| 612 |
+
max_depth : int, default=None
|
| 613 |
+
The maximum depth of the tree. If None, then nodes are expanded until
|
| 614 |
+
all leaves are pure or until all leaves contain less than
|
| 615 |
+
min_samples_split samples.
|
| 616 |
+
|
| 617 |
+
min_samples_split : int or float, default=2
|
| 618 |
+
The minimum number of samples required to split an internal node:
|
| 619 |
+
|
| 620 |
+
- If int, then consider `min_samples_split` as the minimum number.
|
| 621 |
+
- If float, then `min_samples_split` is a fraction and
|
| 622 |
+
`ceil(min_samples_split * n_samples)` are the minimum
|
| 623 |
+
number of samples for each split.
|
| 624 |
+
|
| 625 |
+
.. versionchanged:: 0.18
|
| 626 |
+
Added float values for fractions.
|
| 627 |
+
|
| 628 |
+
min_samples_leaf : int or float, default=1
|
| 629 |
+
The minimum number of samples required to be at a leaf node.
|
| 630 |
+
A split point at any depth will only be considered if it leaves at
|
| 631 |
+
least ``min_samples_leaf`` training samples in each of the left and
|
| 632 |
+
right branches. This may have the effect of smoothing the model,
|
| 633 |
+
especially in regression.
|
| 634 |
+
|
| 635 |
+
- If int, then consider `min_samples_leaf` as the minimum number.
|
| 636 |
+
- If float, then `min_samples_leaf` is a fraction and
|
| 637 |
+
`ceil(min_samples_leaf * n_samples)` are the minimum
|
| 638 |
+
number of samples for each node.
|
| 639 |
+
|
| 640 |
+
.. versionchanged:: 0.18
|
| 641 |
+
Added float values for fractions.
|
| 642 |
+
|
| 643 |
+
min_weight_fraction_leaf : float, default=0.0
|
| 644 |
+
The minimum weighted fraction of the sum total of weights (of all
|
| 645 |
+
the input samples) required to be at a leaf node. Samples have
|
| 646 |
+
equal weight when sample_weight is not provided.
|
| 647 |
+
|
| 648 |
+
max_features : int, float or {"auto", "sqrt", "log2"}, default=None
|
| 649 |
+
The number of features to consider when looking for the best split:
|
| 650 |
+
|
| 651 |
+
- If int, then consider `max_features` features at each split.
|
| 652 |
+
- If float, then `max_features` is a fraction and
|
| 653 |
+
`max(1, int(max_features * n_features_in_))` features are considered at
|
| 654 |
+
each split.
|
| 655 |
+
- If "auto", then `max_features=sqrt(n_features)`.
|
| 656 |
+
- If "sqrt", then `max_features=sqrt(n_features)`.
|
| 657 |
+
- If "log2", then `max_features=log2(n_features)`.
|
| 658 |
+
- If None, then `max_features=n_features`.
|
| 659 |
+
|
| 660 |
+
.. deprecated:: 1.1
|
| 661 |
+
The `"auto"` option was deprecated in 1.1 and will be removed
|
| 662 |
+
in 1.3.
|
| 663 |
+
|
| 664 |
+
Note: the search for a split does not stop until at least one
|
| 665 |
+
valid partition of the node samples is found, even if it requires to
|
| 666 |
+
effectively inspect more than ``max_features`` features.
|
| 667 |
+
|
| 668 |
+
random_state : int, RandomState instance or None, default=None
|
| 669 |
+
Controls the randomness of the estimator. The features are always
|
| 670 |
+
randomly permuted at each split, even if ``splitter`` is set to
|
| 671 |
+
``"best"``. When ``max_features < n_features``, the algorithm will
|
| 672 |
+
select ``max_features`` at random at each split before finding the best
|
| 673 |
+
split among them. But the best found split may vary across different
|
| 674 |
+
runs, even if ``max_features=n_features``. That is the case, if the
|
| 675 |
+
improvement of the criterion is identical for several splits and one
|
| 676 |
+
split has to be selected at random. To obtain a deterministic behaviour
|
| 677 |
+
during fitting, ``random_state`` has to be fixed to an integer.
|
| 678 |
+
See :term:`Glossary <random_state>` for details.
|
| 679 |
+
|
| 680 |
+
max_leaf_nodes : int, default=None
|
| 681 |
+
Grow a tree with ``max_leaf_nodes`` in best-first fashion.
|
| 682 |
+
Best nodes are defined as relative reduction in impurity.
|
| 683 |
+
If None then unlimited number of leaf nodes.
|
| 684 |
+
|
| 685 |
+
min_impurity_decrease : float, default=0.0
|
| 686 |
+
A node will be split if this split induces a decrease of the impurity
|
| 687 |
+
greater than or equal to this value.
|
| 688 |
+
|
| 689 |
+
The weighted impurity decrease equation is the following::
|
| 690 |
+
|
| 691 |
+
N_t / N * (impurity - N_t_R / N_t * right_impurity
|
| 692 |
+
- N_t_L / N_t * left_impurity)
|
| 693 |
+
|
| 694 |
+
where ``N`` is the total number of samples, ``N_t`` is the number of
|
| 695 |
+
samples at the current node, ``N_t_L`` is the number of samples in the
|
| 696 |
+
left child, and ``N_t_R`` is the number of samples in the right child.
|
| 697 |
+
|
| 698 |
+
``N``, ``N_t``, ``N_t_R`` and ``N_t_L`` all refer to the weighted sum,
|
| 699 |
+
if ``sample_weight`` is passed.
|
| 700 |
+
|
| 701 |
+
.. versionadded:: 0.19
|
| 702 |
+
|
| 703 |
+
class_weight : dict, list of dict or "balanced", default=None
|
| 704 |
+
Weights associated with classes in the form ``{class_label: weight}``.
|
| 705 |
+
If None, all classes are supposed to have weight one. For
|
| 706 |
+
multi-output problems, a list of dicts can be provided in the same
|
| 707 |
+
order as the columns of y.
|
| 708 |
+
|
| 709 |
+
Note that for multioutput (including multilabel) weights should be
|
| 710 |
+
defined for each class of every column in its own dict. For example,
|
| 711 |
+
for four-class multilabel classification weights should be
|
| 712 |
+
[{0: 1, 1: 1}, {0: 1, 1: 5}, {0: 1, 1: 1}, {0: 1, 1: 1}] instead of
|
| 713 |
+
[{1:1}, {2:5}, {3:1}, {4:1}].
|
| 714 |
+
|
| 715 |
+
The "balanced" mode uses the values of y to automatically adjust
|
| 716 |
+
weights inversely proportional to class frequencies in the input data
|
| 717 |
+
as ``n_samples / (n_classes * np.bincount(y))``
|
| 718 |
+
|
| 719 |
+
For multi-output, the weights of each column of y will be multiplied.
|
| 720 |
+
|
| 721 |
+
Note that these weights will be multiplied with sample_weight (passed
|
| 722 |
+
through the fit method) if sample_weight is specified.
|
| 723 |
+
|
| 724 |
+
ccp_alpha : non-negative float, default=0.0
|
| 725 |
+
Complexity parameter used for Minimal Cost-Complexity Pruning. The
|
| 726 |
+
subtree with the largest cost complexity that is smaller than
|
| 727 |
+
``ccp_alpha`` will be chosen. By default, no pruning is performed. See
|
| 728 |
+
:ref:`minimal_cost_complexity_pruning` for details.
|
| 729 |
+
|
| 730 |
+
.. versionadded:: 0.22
|
| 731 |
+
|
| 732 |
+
Attributes
|
| 733 |
+
----------
|
| 734 |
+
classes_ : ndarray of shape (n_classes,) or list of ndarray
|
| 735 |
+
The classes labels (single output problem),
|
| 736 |
+
or a list of arrays of class labels (multi-output problem).
|
| 737 |
+
|
| 738 |
+
feature_importances_ : ndarray of shape (n_features,)
|
| 739 |
+
The impurity-based feature importances.
|
| 740 |
+
The higher, the more important the feature.
|
| 741 |
+
The importance of a feature is computed as the (normalized)
|
| 742 |
+
total reduction of the criterion brought by that feature. It is also
|
| 743 |
+
known as the Gini importance [4]_.
|
| 744 |
+
|
| 745 |
+
Warning: impurity-based feature importances can be misleading for
|
| 746 |
+
high cardinality features (many unique values). See
|
| 747 |
+
:func:`sklearn.inspection.permutation_importance` as an alternative.
|
| 748 |
+
|
| 749 |
+
max_features_ : int
|
| 750 |
+
The inferred value of max_features.
|
| 751 |
+
|
| 752 |
+
n_classes_ : int or list of int
|
| 753 |
+
The number of classes (for single output problems),
|
| 754 |
+
or a list containing the number of classes for each
|
| 755 |
+
output (for multi-output problems).
|
| 756 |
+
|
| 757 |
+
n_features_in_ : int
|
| 758 |
+
Number of features seen during :term:`fit`.
|
| 759 |
+
|
| 760 |
+
.. versionadded:: 0.24
|
| 761 |
+
|
| 762 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 763 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 764 |
+
has feature names that are all strings.
|
| 765 |
+
|
| 766 |
+
.. versionadded:: 1.0
|
| 767 |
+
|
| 768 |
+
n_outputs_ : int
|
| 769 |
+
The number of outputs when ``fit`` is performed.
|
| 770 |
+
|
| 771 |
+
tree_ : Tree instance
|
| 772 |
+
The underlying Tree object. Please refer to
|
| 773 |
+
``help(sklearn.tree._tree.Tree)`` for attributes of Tree object and
|
| 774 |
+
:ref:`sphx_glr_auto_examples_tree_plot_unveil_tree_structure.py`
|
| 775 |
+
for basic usage of these attributes.
|
| 776 |
+
|
| 777 |
+
See Also
|
| 778 |
+
--------
|
| 779 |
+
DecisionTreeRegressor : A decision tree regressor.
|
| 780 |
+
|
| 781 |
+
Notes
|
| 782 |
+
-----
|
| 783 |
+
The default values for the parameters controlling the size of the trees
|
| 784 |
+
(e.g. ``max_depth``, ``min_samples_leaf``, etc.) lead to fully grown and
|
| 785 |
+
unpruned trees which can potentially be very large on some data sets. To
|
| 786 |
+
reduce memory consumption, the complexity and size of the trees should be
|
| 787 |
+
controlled by setting those parameter values.
|
| 788 |
+
|
| 789 |
+
The :meth:`predict` method operates using the :func:`numpy.argmax`
|
| 790 |
+
function on the outputs of :meth:`predict_proba`. This means that in
|
| 791 |
+
case the highest predicted probabilities are tied, the classifier will
|
| 792 |
+
predict the tied class with the lowest index in :term:`classes_`.
|
| 793 |
+
|
| 794 |
+
References
|
| 795 |
+
----------
|
| 796 |
+
|
| 797 |
+
.. [1] https://en.wikipedia.org/wiki/Decision_tree_learning
|
| 798 |
+
|
| 799 |
+
.. [2] L. Breiman, J. Friedman, R. Olshen, and C. Stone, "Classification
|
| 800 |
+
and Regression Trees", Wadsworth, Belmont, CA, 1984.
|
| 801 |
+
|
| 802 |
+
.. [3] T. Hastie, R. Tibshirani and J. Friedman. "Elements of Statistical
|
| 803 |
+
Learning", Springer, 2009.
|
| 804 |
+
|
| 805 |
+
.. [4] L. Breiman, and A. Cutler, "Random Forests",
|
| 806 |
+
https://www.stat.berkeley.edu/~breiman/RandomForests/cc_home.htm
|
| 807 |
+
|
| 808 |
+
Examples
|
| 809 |
+
--------
|
| 810 |
+
>>> from sklearn.datasets import load_iris
|
| 811 |
+
>>> from sklearn.model_selection import cross_val_score
|
| 812 |
+
>>> from sklearn.tree import DecisionTreeClassifier
|
| 813 |
+
>>> clf = DecisionTreeClassifier(random_state=0)
|
| 814 |
+
>>> iris = load_iris()
|
| 815 |
+
>>> cross_val_score(clf, iris.data, iris.target, cv=10)
|
| 816 |
+
... # doctest: +SKIP
|
| 817 |
+
...
|
| 818 |
+
array([ 1. , 0.93..., 0.86..., 0.93..., 0.93...,
|
| 819 |
+
0.93..., 0.93..., 1. , 0.93..., 1. ])
|
| 820 |
+
"""
|
| 821 |
+
|
| 822 |
+
_parameter_constraints: dict = {
|
| 823 |
+
**BaseDecisionTree._parameter_constraints,
|
| 824 |
+
"criterion": [StrOptions({"gini", "entropy", "log_loss"}), Hidden(Criterion)],
|
| 825 |
+
"class_weight": [dict, list, StrOptions({"balanced"}), None],
|
| 826 |
+
}
|
| 827 |
+
|
| 828 |
+
def __init__(
|
| 829 |
+
self,
|
| 830 |
+
*,
|
| 831 |
+
criterion="gini",
|
| 832 |
+
splitter="best",
|
| 833 |
+
max_depth=None,
|
| 834 |
+
min_samples_split=2,
|
| 835 |
+
min_samples_leaf=1,
|
| 836 |
+
min_weight_fraction_leaf=0.0,
|
| 837 |
+
max_features=None,
|
| 838 |
+
random_state=None,
|
| 839 |
+
max_leaf_nodes=None,
|
| 840 |
+
min_impurity_decrease=0.0,
|
| 841 |
+
class_weight=None,
|
| 842 |
+
ccp_alpha=0.0,
|
| 843 |
+
):
|
| 844 |
+
super().__init__(
|
| 845 |
+
criterion=criterion,
|
| 846 |
+
splitter=splitter,
|
| 847 |
+
max_depth=max_depth,
|
| 848 |
+
min_samples_split=min_samples_split,
|
| 849 |
+
min_samples_leaf=min_samples_leaf,
|
| 850 |
+
min_weight_fraction_leaf=min_weight_fraction_leaf,
|
| 851 |
+
max_features=max_features,
|
| 852 |
+
max_leaf_nodes=max_leaf_nodes,
|
| 853 |
+
class_weight=class_weight,
|
| 854 |
+
random_state=random_state,
|
| 855 |
+
min_impurity_decrease=min_impurity_decrease,
|
| 856 |
+
ccp_alpha=ccp_alpha,
|
| 857 |
+
)
|
| 858 |
+
|
| 859 |
+
def fit(self, X, y, sample_weight=None, check_input=True):
|
| 860 |
+
"""Build a decision tree classifier from the training set (X, y).
|
| 861 |
+
|
| 862 |
+
Parameters
|
| 863 |
+
----------
|
| 864 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 865 |
+
The training input samples. Internally, it will be converted to
|
| 866 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 867 |
+
to a sparse ``csc_matrix``.
|
| 868 |
+
|
| 869 |
+
y : array-like of shape (n_samples,) or (n_samples, n_outputs)
|
| 870 |
+
The target values (class labels) as integers or strings.
|
| 871 |
+
|
| 872 |
+
sample_weight : array-like of shape (n_samples,), default=None
|
| 873 |
+
Sample weights. If None, then samples are equally weighted. Splits
|
| 874 |
+
that would create child nodes with net zero or negative weight are
|
| 875 |
+
ignored while searching for a split in each node. Splits are also
|
| 876 |
+
ignored if they would result in any single class carrying a
|
| 877 |
+
negative weight in either child node.
|
| 878 |
+
|
| 879 |
+
check_input : bool, default=True
|
| 880 |
+
Allow to bypass several input checking.
|
| 881 |
+
Don't use this parameter unless you know what you're doing.
|
| 882 |
+
|
| 883 |
+
Returns
|
| 884 |
+
-------
|
| 885 |
+
self : DecisionTreeClassifier
|
| 886 |
+
Fitted estimator.
|
| 887 |
+
"""
|
| 888 |
+
|
| 889 |
+
super().fit(
|
| 890 |
+
X,
|
| 891 |
+
y,
|
| 892 |
+
sample_weight=sample_weight,
|
| 893 |
+
check_input=check_input,
|
| 894 |
+
)
|
| 895 |
+
return self
|
| 896 |
+
|
| 897 |
+
def predict_proba(self, X, check_input=True):
|
| 898 |
+
"""Predict class probabilities of the input samples X.
|
| 899 |
+
|
| 900 |
+
The predicted class probability is the fraction of samples of the same
|
| 901 |
+
class in a leaf.
|
| 902 |
+
|
| 903 |
+
Parameters
|
| 904 |
+
----------
|
| 905 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 906 |
+
The input samples. Internally, it will be converted to
|
| 907 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 908 |
+
to a sparse ``csr_matrix``.
|
| 909 |
+
|
| 910 |
+
check_input : bool, default=True
|
| 911 |
+
Allow to bypass several input checking.
|
| 912 |
+
Don't use this parameter unless you know what you're doing.
|
| 913 |
+
|
| 914 |
+
Returns
|
| 915 |
+
-------
|
| 916 |
+
proba : ndarray of shape (n_samples, n_classes) or list of n_outputs \
|
| 917 |
+
such arrays if n_outputs > 1
|
| 918 |
+
The class probabilities of the input samples. The order of the
|
| 919 |
+
classes corresponds to that in the attribute :term:`classes_`.
|
| 920 |
+
"""
|
| 921 |
+
check_is_fitted(self)
|
| 922 |
+
X = self._validate_X_predict(X, check_input)
|
| 923 |
+
proba = self.tree_.predict(X)
|
| 924 |
+
|
| 925 |
+
if self.n_outputs_ == 1:
|
| 926 |
+
proba = proba[:, : self.n_classes_]
|
| 927 |
+
normalizer = proba.sum(axis=1)[:, np.newaxis]
|
| 928 |
+
normalizer[normalizer == 0.0] = 1.0
|
| 929 |
+
proba /= normalizer
|
| 930 |
+
|
| 931 |
+
return proba
|
| 932 |
+
|
| 933 |
+
else:
|
| 934 |
+
all_proba = []
|
| 935 |
+
|
| 936 |
+
for k in range(self.n_outputs_):
|
| 937 |
+
proba_k = proba[:, k, : self.n_classes_[k]]
|
| 938 |
+
normalizer = proba_k.sum(axis=1)[:, np.newaxis]
|
| 939 |
+
normalizer[normalizer == 0.0] = 1.0
|
| 940 |
+
proba_k /= normalizer
|
| 941 |
+
all_proba.append(proba_k)
|
| 942 |
+
|
| 943 |
+
return all_proba
|
| 944 |
+
|
| 945 |
+
def predict_log_proba(self, X):
|
| 946 |
+
"""Predict class log-probabilities of the input samples X.
|
| 947 |
+
|
| 948 |
+
Parameters
|
| 949 |
+
----------
|
| 950 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 951 |
+
The input samples. Internally, it will be converted to
|
| 952 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 953 |
+
to a sparse ``csr_matrix``.
|
| 954 |
+
|
| 955 |
+
Returns
|
| 956 |
+
-------
|
| 957 |
+
proba : ndarray of shape (n_samples, n_classes) or list of n_outputs \
|
| 958 |
+
such arrays if n_outputs > 1
|
| 959 |
+
The class log-probabilities of the input samples. The order of the
|
| 960 |
+
classes corresponds to that in the attribute :term:`classes_`.
|
| 961 |
+
"""
|
| 962 |
+
proba = self.predict_proba(X)
|
| 963 |
+
|
| 964 |
+
if self.n_outputs_ == 1:
|
| 965 |
+
return np.log(proba)
|
| 966 |
+
|
| 967 |
+
else:
|
| 968 |
+
for k in range(self.n_outputs_):
|
| 969 |
+
proba[k] = np.log(proba[k])
|
| 970 |
+
|
| 971 |
+
return proba
|
| 972 |
+
|
| 973 |
+
def _more_tags(self):
|
| 974 |
+
return {"multilabel": True}
|
| 975 |
+
|
| 976 |
+
|
| 977 |
+
class DecisionTreeRegressor(RegressorMixin, BaseDecisionTree):
|
| 978 |
+
"""A decision tree regressor.
|
| 979 |
+
|
| 980 |
+
Read more in the :ref:`User Guide <tree>`.
|
| 981 |
+
|
| 982 |
+
Parameters
|
| 983 |
+
----------
|
| 984 |
+
criterion : {"squared_error", "friedman_mse", "absolute_error", \
|
| 985 |
+
"poisson"}, default="squared_error"
|
| 986 |
+
The function to measure the quality of a split. Supported criteria
|
| 987 |
+
are "squared_error" for the mean squared error, which is equal to
|
| 988 |
+
variance reduction as feature selection criterion and minimizes the L2
|
| 989 |
+
loss using the mean of each terminal node, "friedman_mse", which uses
|
| 990 |
+
mean squared error with Friedman's improvement score for potential
|
| 991 |
+
splits, "absolute_error" for the mean absolute error, which minimizes
|
| 992 |
+
the L1 loss using the median of each terminal node, and "poisson" which
|
| 993 |
+
uses reduction in Poisson deviance to find splits.
|
| 994 |
+
|
| 995 |
+
.. versionadded:: 0.18
|
| 996 |
+
Mean Absolute Error (MAE) criterion.
|
| 997 |
+
|
| 998 |
+
.. versionadded:: 0.24
|
| 999 |
+
Poisson deviance criterion.
|
| 1000 |
+
|
| 1001 |
+
splitter : {"best", "random"}, default="best"
|
| 1002 |
+
The strategy used to choose the split at each node. Supported
|
| 1003 |
+
strategies are "best" to choose the best split and "random" to choose
|
| 1004 |
+
the best random split.
|
| 1005 |
+
|
| 1006 |
+
max_depth : int, default=None
|
| 1007 |
+
The maximum depth of the tree. If None, then nodes are expanded until
|
| 1008 |
+
all leaves are pure or until all leaves contain less than
|
| 1009 |
+
min_samples_split samples.
|
| 1010 |
+
|
| 1011 |
+
min_samples_split : int or float, default=2
|
| 1012 |
+
The minimum number of samples required to split an internal node:
|
| 1013 |
+
|
| 1014 |
+
- If int, then consider `min_samples_split` as the minimum number.
|
| 1015 |
+
- If float, then `min_samples_split` is a fraction and
|
| 1016 |
+
`ceil(min_samples_split * n_samples)` are the minimum
|
| 1017 |
+
number of samples for each split.
|
| 1018 |
+
|
| 1019 |
+
.. versionchanged:: 0.18
|
| 1020 |
+
Added float values for fractions.
|
| 1021 |
+
|
| 1022 |
+
min_samples_leaf : int or float, default=1
|
| 1023 |
+
The minimum number of samples required to be at a leaf node.
|
| 1024 |
+
A split point at any depth will only be considered if it leaves at
|
| 1025 |
+
least ``min_samples_leaf`` training samples in each of the left and
|
| 1026 |
+
right branches. This may have the effect of smoothing the model,
|
| 1027 |
+
especially in regression.
|
| 1028 |
+
|
| 1029 |
+
- If int, then consider `min_samples_leaf` as the minimum number.
|
| 1030 |
+
- If float, then `min_samples_leaf` is a fraction and
|
| 1031 |
+
`ceil(min_samples_leaf * n_samples)` are the minimum
|
| 1032 |
+
number of samples for each node.
|
| 1033 |
+
|
| 1034 |
+
.. versionchanged:: 0.18
|
| 1035 |
+
Added float values for fractions.
|
| 1036 |
+
|
| 1037 |
+
min_weight_fraction_leaf : float, default=0.0
|
| 1038 |
+
The minimum weighted fraction of the sum total of weights (of all
|
| 1039 |
+
the input samples) required to be at a leaf node. Samples have
|
| 1040 |
+
equal weight when sample_weight is not provided.
|
| 1041 |
+
|
| 1042 |
+
max_features : int, float or {"auto", "sqrt", "log2"}, default=None
|
| 1043 |
+
The number of features to consider when looking for the best split:
|
| 1044 |
+
|
| 1045 |
+
- If int, then consider `max_features` features at each split.
|
| 1046 |
+
- If float, then `max_features` is a fraction and
|
| 1047 |
+
`max(1, int(max_features * n_features_in_))` features are considered at each
|
| 1048 |
+
split.
|
| 1049 |
+
- If "auto", then `max_features=n_features`.
|
| 1050 |
+
- If "sqrt", then `max_features=sqrt(n_features)`.
|
| 1051 |
+
- If "log2", then `max_features=log2(n_features)`.
|
| 1052 |
+
- If None, then `max_features=n_features`.
|
| 1053 |
+
|
| 1054 |
+
.. deprecated:: 1.1
|
| 1055 |
+
The `"auto"` option was deprecated in 1.1 and will be removed
|
| 1056 |
+
in 1.3.
|
| 1057 |
+
|
| 1058 |
+
Note: the search for a split does not stop until at least one
|
| 1059 |
+
valid partition of the node samples is found, even if it requires to
|
| 1060 |
+
effectively inspect more than ``max_features`` features.
|
| 1061 |
+
|
| 1062 |
+
random_state : int, RandomState instance or None, default=None
|
| 1063 |
+
Controls the randomness of the estimator. The features are always
|
| 1064 |
+
randomly permuted at each split, even if ``splitter`` is set to
|
| 1065 |
+
``"best"``. When ``max_features < n_features``, the algorithm will
|
| 1066 |
+
select ``max_features`` at random at each split before finding the best
|
| 1067 |
+
split among them. But the best found split may vary across different
|
| 1068 |
+
runs, even if ``max_features=n_features``. That is the case, if the
|
| 1069 |
+
improvement of the criterion is identical for several splits and one
|
| 1070 |
+
split has to be selected at random. To obtain a deterministic behaviour
|
| 1071 |
+
during fitting, ``random_state`` has to be fixed to an integer.
|
| 1072 |
+
See :term:`Glossary <random_state>` for details.
|
| 1073 |
+
|
| 1074 |
+
max_leaf_nodes : int, default=None
|
| 1075 |
+
Grow a tree with ``max_leaf_nodes`` in best-first fashion.
|
| 1076 |
+
Best nodes are defined as relative reduction in impurity.
|
| 1077 |
+
If None then unlimited number of leaf nodes.
|
| 1078 |
+
|
| 1079 |
+
min_impurity_decrease : float, default=0.0
|
| 1080 |
+
A node will be split if this split induces a decrease of the impurity
|
| 1081 |
+
greater than or equal to this value.
|
| 1082 |
+
|
| 1083 |
+
The weighted impurity decrease equation is the following::
|
| 1084 |
+
|
| 1085 |
+
N_t / N * (impurity - N_t_R / N_t * right_impurity
|
| 1086 |
+
- N_t_L / N_t * left_impurity)
|
| 1087 |
+
|
| 1088 |
+
where ``N`` is the total number of samples, ``N_t`` is the number of
|
| 1089 |
+
samples at the current node, ``N_t_L`` is the number of samples in the
|
| 1090 |
+
left child, and ``N_t_R`` is the number of samples in the right child.
|
| 1091 |
+
|
| 1092 |
+
``N``, ``N_t``, ``N_t_R`` and ``N_t_L`` all refer to the weighted sum,
|
| 1093 |
+
if ``sample_weight`` is passed.
|
| 1094 |
+
|
| 1095 |
+
.. versionadded:: 0.19
|
| 1096 |
+
|
| 1097 |
+
ccp_alpha : non-negative float, default=0.0
|
| 1098 |
+
Complexity parameter used for Minimal Cost-Complexity Pruning. The
|
| 1099 |
+
subtree with the largest cost complexity that is smaller than
|
| 1100 |
+
``ccp_alpha`` will be chosen. By default, no pruning is performed. See
|
| 1101 |
+
:ref:`minimal_cost_complexity_pruning` for details.
|
| 1102 |
+
|
| 1103 |
+
.. versionadded:: 0.22
|
| 1104 |
+
|
| 1105 |
+
Attributes
|
| 1106 |
+
----------
|
| 1107 |
+
feature_importances_ : ndarray of shape (n_features,)
|
| 1108 |
+
The feature importances.
|
| 1109 |
+
The higher, the more important the feature.
|
| 1110 |
+
The importance of a feature is computed as the
|
| 1111 |
+
(normalized) total reduction of the criterion brought
|
| 1112 |
+
by that feature. It is also known as the Gini importance [4]_.
|
| 1113 |
+
|
| 1114 |
+
Warning: impurity-based feature importances can be misleading for
|
| 1115 |
+
high cardinality features (many unique values). See
|
| 1116 |
+
:func:`sklearn.inspection.permutation_importance` as an alternative.
|
| 1117 |
+
|
| 1118 |
+
max_features_ : int
|
| 1119 |
+
The inferred value of max_features.
|
| 1120 |
+
|
| 1121 |
+
n_features_in_ : int
|
| 1122 |
+
Number of features seen during :term:`fit`.
|
| 1123 |
+
|
| 1124 |
+
.. versionadded:: 0.24
|
| 1125 |
+
|
| 1126 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 1127 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 1128 |
+
has feature names that are all strings.
|
| 1129 |
+
|
| 1130 |
+
.. versionadded:: 1.0
|
| 1131 |
+
|
| 1132 |
+
n_outputs_ : int
|
| 1133 |
+
The number of outputs when ``fit`` is performed.
|
| 1134 |
+
|
| 1135 |
+
tree_ : Tree instance
|
| 1136 |
+
The underlying Tree object. Please refer to
|
| 1137 |
+
``help(sklearn.tree._tree.Tree)`` for attributes of Tree object and
|
| 1138 |
+
:ref:`sphx_glr_auto_examples_tree_plot_unveil_tree_structure.py`
|
| 1139 |
+
for basic usage of these attributes.
|
| 1140 |
+
|
| 1141 |
+
See Also
|
| 1142 |
+
--------
|
| 1143 |
+
DecisionTreeClassifier : A decision tree classifier.
|
| 1144 |
+
|
| 1145 |
+
Notes
|
| 1146 |
+
-----
|
| 1147 |
+
The default values for the parameters controlling the size of the trees
|
| 1148 |
+
(e.g. ``max_depth``, ``min_samples_leaf``, etc.) lead to fully grown and
|
| 1149 |
+
unpruned trees which can potentially be very large on some data sets. To
|
| 1150 |
+
reduce memory consumption, the complexity and size of the trees should be
|
| 1151 |
+
controlled by setting those parameter values.
|
| 1152 |
+
|
| 1153 |
+
References
|
| 1154 |
+
----------
|
| 1155 |
+
|
| 1156 |
+
.. [1] https://en.wikipedia.org/wiki/Decision_tree_learning
|
| 1157 |
+
|
| 1158 |
+
.. [2] L. Breiman, J. Friedman, R. Olshen, and C. Stone, "Classification
|
| 1159 |
+
and Regression Trees", Wadsworth, Belmont, CA, 1984.
|
| 1160 |
+
|
| 1161 |
+
.. [3] T. Hastie, R. Tibshirani and J. Friedman. "Elements of Statistical
|
| 1162 |
+
Learning", Springer, 2009.
|
| 1163 |
+
|
| 1164 |
+
.. [4] L. Breiman, and A. Cutler, "Random Forests",
|
| 1165 |
+
https://www.stat.berkeley.edu/~breiman/RandomForests/cc_home.htm
|
| 1166 |
+
|
| 1167 |
+
Examples
|
| 1168 |
+
--------
|
| 1169 |
+
>>> from sklearn.datasets import load_diabetes
|
| 1170 |
+
>>> from sklearn.model_selection import cross_val_score
|
| 1171 |
+
>>> from sklearn.tree import DecisionTreeRegressor
|
| 1172 |
+
>>> X, y = load_diabetes(return_X_y=True)
|
| 1173 |
+
>>> regressor = DecisionTreeRegressor(random_state=0)
|
| 1174 |
+
>>> cross_val_score(regressor, X, y, cv=10)
|
| 1175 |
+
... # doctest: +SKIP
|
| 1176 |
+
...
|
| 1177 |
+
array([-0.39..., -0.46..., 0.02..., 0.06..., -0.50...,
|
| 1178 |
+
0.16..., 0.11..., -0.73..., -0.30..., -0.00...])
|
| 1179 |
+
"""
|
| 1180 |
+
|
| 1181 |
+
_parameter_constraints: dict = {
|
| 1182 |
+
**BaseDecisionTree._parameter_constraints,
|
| 1183 |
+
"criterion": [
|
| 1184 |
+
StrOptions({"squared_error", "friedman_mse", "absolute_error", "poisson"}),
|
| 1185 |
+
Hidden(Criterion),
|
| 1186 |
+
],
|
| 1187 |
+
}
|
| 1188 |
+
|
| 1189 |
+
def __init__(
|
| 1190 |
+
self,
|
| 1191 |
+
*,
|
| 1192 |
+
criterion="squared_error",
|
| 1193 |
+
splitter="best",
|
| 1194 |
+
max_depth=None,
|
| 1195 |
+
min_samples_split=2,
|
| 1196 |
+
min_samples_leaf=1,
|
| 1197 |
+
min_weight_fraction_leaf=0.0,
|
| 1198 |
+
max_features=None,
|
| 1199 |
+
random_state=None,
|
| 1200 |
+
max_leaf_nodes=None,
|
| 1201 |
+
min_impurity_decrease=0.0,
|
| 1202 |
+
ccp_alpha=0.0,
|
| 1203 |
+
):
|
| 1204 |
+
super().__init__(
|
| 1205 |
+
criterion=criterion,
|
| 1206 |
+
splitter=splitter,
|
| 1207 |
+
max_depth=max_depth,
|
| 1208 |
+
min_samples_split=min_samples_split,
|
| 1209 |
+
min_samples_leaf=min_samples_leaf,
|
| 1210 |
+
min_weight_fraction_leaf=min_weight_fraction_leaf,
|
| 1211 |
+
max_features=max_features,
|
| 1212 |
+
max_leaf_nodes=max_leaf_nodes,
|
| 1213 |
+
random_state=random_state,
|
| 1214 |
+
min_impurity_decrease=min_impurity_decrease,
|
| 1215 |
+
ccp_alpha=ccp_alpha,
|
| 1216 |
+
)
|
| 1217 |
+
|
| 1218 |
+
def fit(self, X, y, sample_weight=None, check_input=True):
|
| 1219 |
+
"""Build a decision tree regressor from the training set (X, y).
|
| 1220 |
+
|
| 1221 |
+
Parameters
|
| 1222 |
+
----------
|
| 1223 |
+
X : {array-like, sparse matrix} of shape (n_samples, n_features)
|
| 1224 |
+
The training input samples. Internally, it will be converted to
|
| 1225 |
+
``dtype=np.float32`` and if a sparse matrix is provided
|
| 1226 |
+
to a sparse ``csc_matrix``.
|
| 1227 |
+
|
| 1228 |
+
y : array-like of shape (n_samples,) or (n_samples, n_outputs)
|
| 1229 |
+
The target values (real numbers). Use ``dtype=np.float64`` and
|
| 1230 |
+
``order='C'`` for maximum efficiency.
|
| 1231 |
+
|
| 1232 |
+
sample_weight : array-like of shape (n_samples,), default=None
|
| 1233 |
+
Sample weights. If None, then samples are equally weighted. Splits
|
| 1234 |
+
that would create child nodes with net zero or negative weight are
|
| 1235 |
+
ignored while searching for a split in each node.
|
| 1236 |
+
|
| 1237 |
+
check_input : bool, default=True
|
| 1238 |
+
Allow to bypass several input checking.
|
| 1239 |
+
Don't use this parameter unless you know what you're doing.
|
| 1240 |
+
|
| 1241 |
+
Returns
|
| 1242 |
+
-------
|
| 1243 |
+
self : DecisionTreeRegressor
|
| 1244 |
+
Fitted estimator.
|
| 1245 |
+
"""
|
| 1246 |
+
|
| 1247 |
+
super().fit(
|
| 1248 |
+
X,
|
| 1249 |
+
y,
|
| 1250 |
+
sample_weight=sample_weight,
|
| 1251 |
+
check_input=check_input,
|
| 1252 |
+
)
|
| 1253 |
+
return self
|
| 1254 |
+
|
| 1255 |
+
def _compute_partial_dependence_recursion(self, grid, target_features):
|
| 1256 |
+
"""Fast partial dependence computation.
|
| 1257 |
+
|
| 1258 |
+
Parameters
|
| 1259 |
+
----------
|
| 1260 |
+
grid : ndarray of shape (n_samples, n_target_features)
|
| 1261 |
+
The grid points on which the partial dependence should be
|
| 1262 |
+
evaluated.
|
| 1263 |
+
target_features : ndarray of shape (n_target_features)
|
| 1264 |
+
The set of target features for which the partial dependence
|
| 1265 |
+
should be evaluated.
|
| 1266 |
+
|
| 1267 |
+
Returns
|
| 1268 |
+
-------
|
| 1269 |
+
averaged_predictions : ndarray of shape (n_samples,)
|
| 1270 |
+
The value of the partial dependence function on each grid point.
|
| 1271 |
+
"""
|
| 1272 |
+
grid = np.asarray(grid, dtype=DTYPE, order="C")
|
| 1273 |
+
averaged_predictions = np.zeros(
|
| 1274 |
+
shape=grid.shape[0], dtype=np.float64, order="C"
|
| 1275 |
+
)
|
| 1276 |
+
|
| 1277 |
+
self.tree_.compute_partial_dependence(
|
| 1278 |
+
grid, target_features, averaged_predictions
|
| 1279 |
+
)
|
| 1280 |
+
return averaged_predictions
|
| 1281 |
+
|
| 1282 |
+
|
| 1283 |
+
class ExtraTreeClassifier(DecisionTreeClassifier):
|
| 1284 |
+
"""An extremely randomized tree classifier.
|
| 1285 |
+
|
| 1286 |
+
Extra-trees differ from classic decision trees in the way they are built.
|
| 1287 |
+
When looking for the best split to separate the samples of a node into two
|
| 1288 |
+
groups, random splits are drawn for each of the `max_features` randomly
|
| 1289 |
+
selected features and the best split among those is chosen. When
|
| 1290 |
+
`max_features` is set 1, this amounts to building a totally random
|
| 1291 |
+
decision tree.
|
| 1292 |
+
|
| 1293 |
+
Warning: Extra-trees should only be used within ensemble methods.
|
| 1294 |
+
|
| 1295 |
+
Read more in the :ref:`User Guide <tree>`.
|
| 1296 |
+
|
| 1297 |
+
Parameters
|
| 1298 |
+
----------
|
| 1299 |
+
criterion : {"gini", "entropy", "log_loss"}, default="gini"
|
| 1300 |
+
The function to measure the quality of a split. Supported criteria are
|
| 1301 |
+
"gini" for the Gini impurity and "log_loss" and "entropy" both for the
|
| 1302 |
+
Shannon information gain, see :ref:`tree_mathematical_formulation`.
|
| 1303 |
+
|
| 1304 |
+
splitter : {"random", "best"}, default="random"
|
| 1305 |
+
The strategy used to choose the split at each node. Supported
|
| 1306 |
+
strategies are "best" to choose the best split and "random" to choose
|
| 1307 |
+
the best random split.
|
| 1308 |
+
|
| 1309 |
+
max_depth : int, default=None
|
| 1310 |
+
The maximum depth of the tree. If None, then nodes are expanded until
|
| 1311 |
+
all leaves are pure or until all leaves contain less than
|
| 1312 |
+
min_samples_split samples.
|
| 1313 |
+
|
| 1314 |
+
min_samples_split : int or float, default=2
|
| 1315 |
+
The minimum number of samples required to split an internal node:
|
| 1316 |
+
|
| 1317 |
+
- If int, then consider `min_samples_split` as the minimum number.
|
| 1318 |
+
- If float, then `min_samples_split` is a fraction and
|
| 1319 |
+
`ceil(min_samples_split * n_samples)` are the minimum
|
| 1320 |
+
number of samples for each split.
|
| 1321 |
+
|
| 1322 |
+
.. versionchanged:: 0.18
|
| 1323 |
+
Added float values for fractions.
|
| 1324 |
+
|
| 1325 |
+
min_samples_leaf : int or float, default=1
|
| 1326 |
+
The minimum number of samples required to be at a leaf node.
|
| 1327 |
+
A split point at any depth will only be considered if it leaves at
|
| 1328 |
+
least ``min_samples_leaf`` training samples in each of the left and
|
| 1329 |
+
right branches. This may have the effect of smoothing the model,
|
| 1330 |
+
especially in regression.
|
| 1331 |
+
|
| 1332 |
+
- If int, then consider `min_samples_leaf` as the minimum number.
|
| 1333 |
+
- If float, then `min_samples_leaf` is a fraction and
|
| 1334 |
+
`ceil(min_samples_leaf * n_samples)` are the minimum
|
| 1335 |
+
number of samples for each node.
|
| 1336 |
+
|
| 1337 |
+
.. versionchanged:: 0.18
|
| 1338 |
+
Added float values for fractions.
|
| 1339 |
+
|
| 1340 |
+
min_weight_fraction_leaf : float, default=0.0
|
| 1341 |
+
The minimum weighted fraction of the sum total of weights (of all
|
| 1342 |
+
the input samples) required to be at a leaf node. Samples have
|
| 1343 |
+
equal weight when sample_weight is not provided.
|
| 1344 |
+
|
| 1345 |
+
max_features : int, float, {"auto", "sqrt", "log2"} or None, default="sqrt"
|
| 1346 |
+
The number of features to consider when looking for the best split:
|
| 1347 |
+
|
| 1348 |
+
- If int, then consider `max_features` features at each split.
|
| 1349 |
+
- If float, then `max_features` is a fraction and
|
| 1350 |
+
`max(1, int(max_features * n_features_in_))` features are considered at
|
| 1351 |
+
each split.
|
| 1352 |
+
- If "auto", then `max_features=sqrt(n_features)`.
|
| 1353 |
+
- If "sqrt", then `max_features=sqrt(n_features)`.
|
| 1354 |
+
- If "log2", then `max_features=log2(n_features)`.
|
| 1355 |
+
- If None, then `max_features=n_features`.
|
| 1356 |
+
|
| 1357 |
+
.. versionchanged:: 1.1
|
| 1358 |
+
The default of `max_features` changed from `"auto"` to `"sqrt"`.
|
| 1359 |
+
|
| 1360 |
+
.. deprecated:: 1.1
|
| 1361 |
+
The `"auto"` option was deprecated in 1.1 and will be removed
|
| 1362 |
+
in 1.3.
|
| 1363 |
+
|
| 1364 |
+
Note: the search for a split does not stop until at least one
|
| 1365 |
+
valid partition of the node samples is found, even if it requires to
|
| 1366 |
+
effectively inspect more than ``max_features`` features.
|
| 1367 |
+
|
| 1368 |
+
random_state : int, RandomState instance or None, default=None
|
| 1369 |
+
Used to pick randomly the `max_features` used at each split.
|
| 1370 |
+
See :term:`Glossary <random_state>` for details.
|
| 1371 |
+
|
| 1372 |
+
max_leaf_nodes : int, default=None
|
| 1373 |
+
Grow a tree with ``max_leaf_nodes`` in best-first fashion.
|
| 1374 |
+
Best nodes are defined as relative reduction in impurity.
|
| 1375 |
+
If None then unlimited number of leaf nodes.
|
| 1376 |
+
|
| 1377 |
+
min_impurity_decrease : float, default=0.0
|
| 1378 |
+
A node will be split if this split induces a decrease of the impurity
|
| 1379 |
+
greater than or equal to this value.
|
| 1380 |
+
|
| 1381 |
+
The weighted impurity decrease equation is the following::
|
| 1382 |
+
|
| 1383 |
+
N_t / N * (impurity - N_t_R / N_t * right_impurity
|
| 1384 |
+
- N_t_L / N_t * left_impurity)
|
| 1385 |
+
|
| 1386 |
+
where ``N`` is the total number of samples, ``N_t`` is the number of
|
| 1387 |
+
samples at the current node, ``N_t_L`` is the number of samples in the
|
| 1388 |
+
left child, and ``N_t_R`` is the number of samples in the right child.
|
| 1389 |
+
|
| 1390 |
+
``N``, ``N_t``, ``N_t_R`` and ``N_t_L`` all refer to the weighted sum,
|
| 1391 |
+
if ``sample_weight`` is passed.
|
| 1392 |
+
|
| 1393 |
+
.. versionadded:: 0.19
|
| 1394 |
+
|
| 1395 |
+
class_weight : dict, list of dict or "balanced", default=None
|
| 1396 |
+
Weights associated with classes in the form ``{class_label: weight}``.
|
| 1397 |
+
If None, all classes are supposed to have weight one. For
|
| 1398 |
+
multi-output problems, a list of dicts can be provided in the same
|
| 1399 |
+
order as the columns of y.
|
| 1400 |
+
|
| 1401 |
+
Note that for multioutput (including multilabel) weights should be
|
| 1402 |
+
defined for each class of every column in its own dict. For example,
|
| 1403 |
+
for four-class multilabel classification weights should be
|
| 1404 |
+
[{0: 1, 1: 1}, {0: 1, 1: 5}, {0: 1, 1: 1}, {0: 1, 1: 1}] instead of
|
| 1405 |
+
[{1:1}, {2:5}, {3:1}, {4:1}].
|
| 1406 |
+
|
| 1407 |
+
The "balanced" mode uses the values of y to automatically adjust
|
| 1408 |
+
weights inversely proportional to class frequencies in the input data
|
| 1409 |
+
as ``n_samples / (n_classes * np.bincount(y))``
|
| 1410 |
+
|
| 1411 |
+
For multi-output, the weights of each column of y will be multiplied.
|
| 1412 |
+
|
| 1413 |
+
Note that these weights will be multiplied with sample_weight (passed
|
| 1414 |
+
through the fit method) if sample_weight is specified.
|
| 1415 |
+
|
| 1416 |
+
ccp_alpha : non-negative float, default=0.0
|
| 1417 |
+
Complexity parameter used for Minimal Cost-Complexity Pruning. The
|
| 1418 |
+
subtree with the largest cost complexity that is smaller than
|
| 1419 |
+
``ccp_alpha`` will be chosen. By default, no pruning is performed. See
|
| 1420 |
+
:ref:`minimal_cost_complexity_pruning` for details.
|
| 1421 |
+
|
| 1422 |
+
.. versionadded:: 0.22
|
| 1423 |
+
|
| 1424 |
+
Attributes
|
| 1425 |
+
----------
|
| 1426 |
+
classes_ : ndarray of shape (n_classes,) or list of ndarray
|
| 1427 |
+
The classes labels (single output problem),
|
| 1428 |
+
or a list of arrays of class labels (multi-output problem).
|
| 1429 |
+
|
| 1430 |
+
max_features_ : int
|
| 1431 |
+
The inferred value of max_features.
|
| 1432 |
+
|
| 1433 |
+
n_classes_ : int or list of int
|
| 1434 |
+
The number of classes (for single output problems),
|
| 1435 |
+
or a list containing the number of classes for each
|
| 1436 |
+
output (for multi-output problems).
|
| 1437 |
+
|
| 1438 |
+
feature_importances_ : ndarray of shape (n_features,)
|
| 1439 |
+
The impurity-based feature importances.
|
| 1440 |
+
The higher, the more important the feature.
|
| 1441 |
+
The importance of a feature is computed as the (normalized)
|
| 1442 |
+
total reduction of the criterion brought by that feature. It is also
|
| 1443 |
+
known as the Gini importance.
|
| 1444 |
+
|
| 1445 |
+
Warning: impurity-based feature importances can be misleading for
|
| 1446 |
+
high cardinality features (many unique values). See
|
| 1447 |
+
:func:`sklearn.inspection.permutation_importance` as an alternative.
|
| 1448 |
+
|
| 1449 |
+
n_features_in_ : int
|
| 1450 |
+
Number of features seen during :term:`fit`.
|
| 1451 |
+
|
| 1452 |
+
.. versionadded:: 0.24
|
| 1453 |
+
|
| 1454 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 1455 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 1456 |
+
has feature names that are all strings.
|
| 1457 |
+
|
| 1458 |
+
.. versionadded:: 1.0
|
| 1459 |
+
|
| 1460 |
+
n_outputs_ : int
|
| 1461 |
+
The number of outputs when ``fit`` is performed.
|
| 1462 |
+
|
| 1463 |
+
tree_ : Tree instance
|
| 1464 |
+
The underlying Tree object. Please refer to
|
| 1465 |
+
``help(sklearn.tree._tree.Tree)`` for attributes of Tree object and
|
| 1466 |
+
:ref:`sphx_glr_auto_examples_tree_plot_unveil_tree_structure.py`
|
| 1467 |
+
for basic usage of these attributes.
|
| 1468 |
+
|
| 1469 |
+
See Also
|
| 1470 |
+
--------
|
| 1471 |
+
ExtraTreeRegressor : An extremely randomized tree regressor.
|
| 1472 |
+
sklearn.ensemble.ExtraTreesClassifier : An extra-trees classifier.
|
| 1473 |
+
sklearn.ensemble.ExtraTreesRegressor : An extra-trees regressor.
|
| 1474 |
+
sklearn.ensemble.RandomForestClassifier : A random forest classifier.
|
| 1475 |
+
sklearn.ensemble.RandomForestRegressor : A random forest regressor.
|
| 1476 |
+
sklearn.ensemble.RandomTreesEmbedding : An ensemble of
|
| 1477 |
+
totally random trees.
|
| 1478 |
+
|
| 1479 |
+
Notes
|
| 1480 |
+
-----
|
| 1481 |
+
The default values for the parameters controlling the size of the trees
|
| 1482 |
+
(e.g. ``max_depth``, ``min_samples_leaf``, etc.) lead to fully grown and
|
| 1483 |
+
unpruned trees which can potentially be very large on some data sets. To
|
| 1484 |
+
reduce memory consumption, the complexity and size of the trees should be
|
| 1485 |
+
controlled by setting those parameter values.
|
| 1486 |
+
|
| 1487 |
+
References
|
| 1488 |
+
----------
|
| 1489 |
+
|
| 1490 |
+
.. [1] P. Geurts, D. Ernst., and L. Wehenkel, "Extremely randomized trees",
|
| 1491 |
+
Machine Learning, 63(1), 3-42, 2006.
|
| 1492 |
+
|
| 1493 |
+
Examples
|
| 1494 |
+
--------
|
| 1495 |
+
>>> from sklearn.datasets import load_iris
|
| 1496 |
+
>>> from sklearn.model_selection import train_test_split
|
| 1497 |
+
>>> from sklearn.ensemble import BaggingClassifier
|
| 1498 |
+
>>> from sklearn.tree import ExtraTreeClassifier
|
| 1499 |
+
>>> X, y = load_iris(return_X_y=True)
|
| 1500 |
+
>>> X_train, X_test, y_train, y_test = train_test_split(
|
| 1501 |
+
... X, y, random_state=0)
|
| 1502 |
+
>>> extra_tree = ExtraTreeClassifier(random_state=0)
|
| 1503 |
+
>>> cls = BaggingClassifier(extra_tree, random_state=0).fit(
|
| 1504 |
+
... X_train, y_train)
|
| 1505 |
+
>>> cls.score(X_test, y_test)
|
| 1506 |
+
0.8947...
|
| 1507 |
+
"""
|
| 1508 |
+
|
| 1509 |
+
def __init__(
|
| 1510 |
+
self,
|
| 1511 |
+
*,
|
| 1512 |
+
criterion="gini",
|
| 1513 |
+
splitter="random",
|
| 1514 |
+
max_depth=None,
|
| 1515 |
+
min_samples_split=2,
|
| 1516 |
+
min_samples_leaf=1,
|
| 1517 |
+
min_weight_fraction_leaf=0.0,
|
| 1518 |
+
max_features="sqrt",
|
| 1519 |
+
random_state=None,
|
| 1520 |
+
max_leaf_nodes=None,
|
| 1521 |
+
min_impurity_decrease=0.0,
|
| 1522 |
+
class_weight=None,
|
| 1523 |
+
ccp_alpha=0.0,
|
| 1524 |
+
):
|
| 1525 |
+
super().__init__(
|
| 1526 |
+
criterion=criterion,
|
| 1527 |
+
splitter=splitter,
|
| 1528 |
+
max_depth=max_depth,
|
| 1529 |
+
min_samples_split=min_samples_split,
|
| 1530 |
+
min_samples_leaf=min_samples_leaf,
|
| 1531 |
+
min_weight_fraction_leaf=min_weight_fraction_leaf,
|
| 1532 |
+
max_features=max_features,
|
| 1533 |
+
max_leaf_nodes=max_leaf_nodes,
|
| 1534 |
+
class_weight=class_weight,
|
| 1535 |
+
min_impurity_decrease=min_impurity_decrease,
|
| 1536 |
+
random_state=random_state,
|
| 1537 |
+
ccp_alpha=ccp_alpha,
|
| 1538 |
+
)
|
| 1539 |
+
|
| 1540 |
+
|
| 1541 |
+
class ExtraTreeRegressor(DecisionTreeRegressor):
|
| 1542 |
+
"""An extremely randomized tree regressor.
|
| 1543 |
+
|
| 1544 |
+
Extra-trees differ from classic decision trees in the way they are built.
|
| 1545 |
+
When looking for the best split to separate the samples of a node into two
|
| 1546 |
+
groups, random splits are drawn for each of the `max_features` randomly
|
| 1547 |
+
selected features and the best split among those is chosen. When
|
| 1548 |
+
`max_features` is set 1, this amounts to building a totally random
|
| 1549 |
+
decision tree.
|
| 1550 |
+
|
| 1551 |
+
Warning: Extra-trees should only be used within ensemble methods.
|
| 1552 |
+
|
| 1553 |
+
Read more in the :ref:`User Guide <tree>`.
|
| 1554 |
+
|
| 1555 |
+
Parameters
|
| 1556 |
+
----------
|
| 1557 |
+
criterion : {"squared_error", "friedman_mse", "absolute_error", "poisson"}, \
|
| 1558 |
+
default="squared_error"
|
| 1559 |
+
The function to measure the quality of a split. Supported criteria
|
| 1560 |
+
are "squared_error" for the mean squared error, which is equal to
|
| 1561 |
+
variance reduction as feature selection criterion and minimizes the L2
|
| 1562 |
+
loss using the mean of each terminal node, "friedman_mse", which uses
|
| 1563 |
+
mean squared error with Friedman's improvement score for potential
|
| 1564 |
+
splits, "absolute_error" for the mean absolute error, which minimizes
|
| 1565 |
+
the L1 loss using the median of each terminal node, and "poisson" which
|
| 1566 |
+
uses reduction in Poisson deviance to find splits.
|
| 1567 |
+
|
| 1568 |
+
.. versionadded:: 0.18
|
| 1569 |
+
Mean Absolute Error (MAE) criterion.
|
| 1570 |
+
|
| 1571 |
+
.. versionadded:: 0.24
|
| 1572 |
+
Poisson deviance criterion.
|
| 1573 |
+
|
| 1574 |
+
splitter : {"random", "best"}, default="random"
|
| 1575 |
+
The strategy used to choose the split at each node. Supported
|
| 1576 |
+
strategies are "best" to choose the best split and "random" to choose
|
| 1577 |
+
the best random split.
|
| 1578 |
+
|
| 1579 |
+
max_depth : int, default=None
|
| 1580 |
+
The maximum depth of the tree. If None, then nodes are expanded until
|
| 1581 |
+
all leaves are pure or until all leaves contain less than
|
| 1582 |
+
min_samples_split samples.
|
| 1583 |
+
|
| 1584 |
+
min_samples_split : int or float, default=2
|
| 1585 |
+
The minimum number of samples required to split an internal node:
|
| 1586 |
+
|
| 1587 |
+
- If int, then consider `min_samples_split` as the minimum number.
|
| 1588 |
+
- If float, then `min_samples_split` is a fraction and
|
| 1589 |
+
`ceil(min_samples_split * n_samples)` are the minimum
|
| 1590 |
+
number of samples for each split.
|
| 1591 |
+
|
| 1592 |
+
.. versionchanged:: 0.18
|
| 1593 |
+
Added float values for fractions.
|
| 1594 |
+
|
| 1595 |
+
min_samples_leaf : int or float, default=1
|
| 1596 |
+
The minimum number of samples required to be at a leaf node.
|
| 1597 |
+
A split point at any depth will only be considered if it leaves at
|
| 1598 |
+
least ``min_samples_leaf`` training samples in each of the left and
|
| 1599 |
+
right branches. This may have the effect of smoothing the model,
|
| 1600 |
+
especially in regression.
|
| 1601 |
+
|
| 1602 |
+
- If int, then consider `min_samples_leaf` as the minimum number.
|
| 1603 |
+
- If float, then `min_samples_leaf` is a fraction and
|
| 1604 |
+
`ceil(min_samples_leaf * n_samples)` are the minimum
|
| 1605 |
+
number of samples for each node.
|
| 1606 |
+
|
| 1607 |
+
.. versionchanged:: 0.18
|
| 1608 |
+
Added float values for fractions.
|
| 1609 |
+
|
| 1610 |
+
min_weight_fraction_leaf : float, default=0.0
|
| 1611 |
+
The minimum weighted fraction of the sum total of weights (of all
|
| 1612 |
+
the input samples) required to be at a leaf node. Samples have
|
| 1613 |
+
equal weight when sample_weight is not provided.
|
| 1614 |
+
|
| 1615 |
+
max_features : int, float, {"auto", "sqrt", "log2"} or None, default=1.0
|
| 1616 |
+
The number of features to consider when looking for the best split:
|
| 1617 |
+
|
| 1618 |
+
- If int, then consider `max_features` features at each split.
|
| 1619 |
+
- If float, then `max_features` is a fraction and
|
| 1620 |
+
`max(1, int(max_features * n_features_in_))` features are considered at each
|
| 1621 |
+
split.
|
| 1622 |
+
- If "auto", then `max_features=n_features`.
|
| 1623 |
+
- If "sqrt", then `max_features=sqrt(n_features)`.
|
| 1624 |
+
- If "log2", then `max_features=log2(n_features)`.
|
| 1625 |
+
- If None, then `max_features=n_features`.
|
| 1626 |
+
|
| 1627 |
+
.. versionchanged:: 1.1
|
| 1628 |
+
The default of `max_features` changed from `"auto"` to `1.0`.
|
| 1629 |
+
|
| 1630 |
+
.. deprecated:: 1.1
|
| 1631 |
+
The `"auto"` option was deprecated in 1.1 and will be removed
|
| 1632 |
+
in 1.3.
|
| 1633 |
+
|
| 1634 |
+
Note: the search for a split does not stop until at least one
|
| 1635 |
+
valid partition of the node samples is found, even if it requires to
|
| 1636 |
+
effectively inspect more than ``max_features`` features.
|
| 1637 |
+
|
| 1638 |
+
random_state : int, RandomState instance or None, default=None
|
| 1639 |
+
Used to pick randomly the `max_features` used at each split.
|
| 1640 |
+
See :term:`Glossary <random_state>` for details.
|
| 1641 |
+
|
| 1642 |
+
min_impurity_decrease : float, default=0.0
|
| 1643 |
+
A node will be split if this split induces a decrease of the impurity
|
| 1644 |
+
greater than or equal to this value.
|
| 1645 |
+
|
| 1646 |
+
The weighted impurity decrease equation is the following::
|
| 1647 |
+
|
| 1648 |
+
N_t / N * (impurity - N_t_R / N_t * right_impurity
|
| 1649 |
+
- N_t_L / N_t * left_impurity)
|
| 1650 |
+
|
| 1651 |
+
where ``N`` is the total number of samples, ``N_t`` is the number of
|
| 1652 |
+
samples at the current node, ``N_t_L`` is the number of samples in the
|
| 1653 |
+
left child, and ``N_t_R`` is the number of samples in the right child.
|
| 1654 |
+
|
| 1655 |
+
``N``, ``N_t``, ``N_t_R`` and ``N_t_L`` all refer to the weighted sum,
|
| 1656 |
+
if ``sample_weight`` is passed.
|
| 1657 |
+
|
| 1658 |
+
.. versionadded:: 0.19
|
| 1659 |
+
|
| 1660 |
+
max_leaf_nodes : int, default=None
|
| 1661 |
+
Grow a tree with ``max_leaf_nodes`` in best-first fashion.
|
| 1662 |
+
Best nodes are defined as relative reduction in impurity.
|
| 1663 |
+
If None then unlimited number of leaf nodes.
|
| 1664 |
+
|
| 1665 |
+
ccp_alpha : non-negative float, default=0.0
|
| 1666 |
+
Complexity parameter used for Minimal Cost-Complexity Pruning. The
|
| 1667 |
+
subtree with the largest cost complexity that is smaller than
|
| 1668 |
+
``ccp_alpha`` will be chosen. By default, no pruning is performed. See
|
| 1669 |
+
:ref:`minimal_cost_complexity_pruning` for details.
|
| 1670 |
+
|
| 1671 |
+
.. versionadded:: 0.22
|
| 1672 |
+
|
| 1673 |
+
Attributes
|
| 1674 |
+
----------
|
| 1675 |
+
max_features_ : int
|
| 1676 |
+
The inferred value of max_features.
|
| 1677 |
+
|
| 1678 |
+
n_features_in_ : int
|
| 1679 |
+
Number of features seen during :term:`fit`.
|
| 1680 |
+
|
| 1681 |
+
.. versionadded:: 0.24
|
| 1682 |
+
|
| 1683 |
+
feature_names_in_ : ndarray of shape (`n_features_in_`,)
|
| 1684 |
+
Names of features seen during :term:`fit`. Defined only when `X`
|
| 1685 |
+
has feature names that are all strings.
|
| 1686 |
+
|
| 1687 |
+
.. versionadded:: 1.0
|
| 1688 |
+
|
| 1689 |
+
feature_importances_ : ndarray of shape (n_features,)
|
| 1690 |
+
Return impurity-based feature importances (the higher, the more
|
| 1691 |
+
important the feature).
|
| 1692 |
+
|
| 1693 |
+
Warning: impurity-based feature importances can be misleading for
|
| 1694 |
+
high cardinality features (many unique values). See
|
| 1695 |
+
:func:`sklearn.inspection.permutation_importance` as an alternative.
|
| 1696 |
+
|
| 1697 |
+
n_outputs_ : int
|
| 1698 |
+
The number of outputs when ``fit`` is performed.
|
| 1699 |
+
|
| 1700 |
+
tree_ : Tree instance
|
| 1701 |
+
The underlying Tree object. Please refer to
|
| 1702 |
+
``help(sklearn.tree._tree.Tree)`` for attributes of Tree object and
|
| 1703 |
+
:ref:`sphx_glr_auto_examples_tree_plot_unveil_tree_structure.py`
|
| 1704 |
+
for basic usage of these attributes.
|
| 1705 |
+
|
| 1706 |
+
See Also
|
| 1707 |
+
--------
|
| 1708 |
+
ExtraTreeClassifier : An extremely randomized tree classifier.
|
| 1709 |
+
sklearn.ensemble.ExtraTreesClassifier : An extra-trees classifier.
|
| 1710 |
+
sklearn.ensemble.ExtraTreesRegressor : An extra-trees regressor.
|
| 1711 |
+
|
| 1712 |
+
Notes
|
| 1713 |
+
-----
|
| 1714 |
+
The default values for the parameters controlling the size of the trees
|
| 1715 |
+
(e.g. ``max_depth``, ``min_samples_leaf``, etc.) lead to fully grown and
|
| 1716 |
+
unpruned trees which can potentially be very large on some data sets. To
|
| 1717 |
+
reduce memory consumption, the complexity and size of the trees should be
|
| 1718 |
+
controlled by setting those parameter values.
|
| 1719 |
+
|
| 1720 |
+
References
|
| 1721 |
+
----------
|
| 1722 |
+
|
| 1723 |
+
.. [1] P. Geurts, D. Ernst., and L. Wehenkel, "Extremely randomized trees",
|
| 1724 |
+
Machine Learning, 63(1), 3-42, 2006.
|
| 1725 |
+
|
| 1726 |
+
Examples
|
| 1727 |
+
--------
|
| 1728 |
+
>>> from sklearn.datasets import load_diabetes
|
| 1729 |
+
>>> from sklearn.model_selection import train_test_split
|
| 1730 |
+
>>> from sklearn.ensemble import BaggingRegressor
|
| 1731 |
+
>>> from sklearn.tree import ExtraTreeRegressor
|
| 1732 |
+
>>> X, y = load_diabetes(return_X_y=True)
|
| 1733 |
+
>>> X_train, X_test, y_train, y_test = train_test_split(
|
| 1734 |
+
... X, y, random_state=0)
|
| 1735 |
+
>>> extra_tree = ExtraTreeRegressor(random_state=0)
|
| 1736 |
+
>>> reg = BaggingRegressor(extra_tree, random_state=0).fit(
|
| 1737 |
+
... X_train, y_train)
|
| 1738 |
+
>>> reg.score(X_test, y_test)
|
| 1739 |
+
0.33...
|
| 1740 |
+
"""
|
| 1741 |
+
|
| 1742 |
+
def __init__(
|
| 1743 |
+
self,
|
| 1744 |
+
*,
|
| 1745 |
+
criterion="squared_error",
|
| 1746 |
+
splitter="random",
|
| 1747 |
+
max_depth=None,
|
| 1748 |
+
min_samples_split=2,
|
| 1749 |
+
min_samples_leaf=1,
|
| 1750 |
+
min_weight_fraction_leaf=0.0,
|
| 1751 |
+
max_features=1.0,
|
| 1752 |
+
random_state=None,
|
| 1753 |
+
min_impurity_decrease=0.0,
|
| 1754 |
+
max_leaf_nodes=None,
|
| 1755 |
+
ccp_alpha=0.0,
|
| 1756 |
+
):
|
| 1757 |
+
super().__init__(
|
| 1758 |
+
criterion=criterion,
|
| 1759 |
+
splitter=splitter,
|
| 1760 |
+
max_depth=max_depth,
|
| 1761 |
+
min_samples_split=min_samples_split,
|
| 1762 |
+
min_samples_leaf=min_samples_leaf,
|
| 1763 |
+
min_weight_fraction_leaf=min_weight_fraction_leaf,
|
| 1764 |
+
max_features=max_features,
|
| 1765 |
+
max_leaf_nodes=max_leaf_nodes,
|
| 1766 |
+
min_impurity_decrease=min_impurity_decrease,
|
| 1767 |
+
random_state=random_state,
|
| 1768 |
+
ccp_alpha=ccp_alpha,
|
| 1769 |
+
)
|
videollama2/lib/python3.10/site-packages/sklearn/tree/_splitter.pxd
ADDED
|
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Authors: Gilles Louppe <g.louppe@gmail.com>
|
| 2 |
+
# Peter Prettenhofer <peter.prettenhofer@gmail.com>
|
| 3 |
+
# Brian Holt <bdholt1@gmail.com>
|
| 4 |
+
# Joel Nothman <joel.nothman@gmail.com>
|
| 5 |
+
# Arnaud Joly <arnaud.v.joly@gmail.com>
|
| 6 |
+
# Jacob Schreiber <jmschreiber91@gmail.com>
|
| 7 |
+
#
|
| 8 |
+
# License: BSD 3 clause
|
| 9 |
+
|
| 10 |
+
# See _splitter.pyx for details.
|
| 11 |
+
|
| 12 |
+
from ._criterion cimport Criterion
|
| 13 |
+
|
| 14 |
+
from ._tree cimport DTYPE_t # Type of X
|
| 15 |
+
from ._tree cimport DOUBLE_t # Type of y, sample_weight
|
| 16 |
+
from ._tree cimport SIZE_t # Type for indices and counters
|
| 17 |
+
from ._tree cimport INT32_t # Signed 32 bit integer
|
| 18 |
+
from ._tree cimport UINT32_t # Unsigned 32 bit integer
|
| 19 |
+
|
| 20 |
+
cdef struct SplitRecord:
|
| 21 |
+
# Data to track sample split
|
| 22 |
+
SIZE_t feature # Which feature to split on.
|
| 23 |
+
SIZE_t pos # Split samples array at the given position,
|
| 24 |
+
# i.e. count of samples below threshold for feature.
|
| 25 |
+
# pos is >= end if the node is a leaf.
|
| 26 |
+
double threshold # Threshold to split at.
|
| 27 |
+
double improvement # Impurity improvement given parent node.
|
| 28 |
+
double impurity_left # Impurity of the left split.
|
| 29 |
+
double impurity_right # Impurity of the right split.
|
| 30 |
+
|
| 31 |
+
cdef class Splitter:
|
| 32 |
+
# The splitter searches in the input space for a feature and a threshold
|
| 33 |
+
# to split the samples samples[start:end].
|
| 34 |
+
#
|
| 35 |
+
# The impurity computations are delegated to a criterion object.
|
| 36 |
+
|
| 37 |
+
# Internal structures
|
| 38 |
+
cdef public Criterion criterion # Impurity criterion
|
| 39 |
+
cdef public SIZE_t max_features # Number of features to test
|
| 40 |
+
cdef public SIZE_t min_samples_leaf # Min samples in a leaf
|
| 41 |
+
cdef public double min_weight_leaf # Minimum weight in a leaf
|
| 42 |
+
|
| 43 |
+
cdef object random_state # Random state
|
| 44 |
+
cdef UINT32_t rand_r_state # sklearn_rand_r random number state
|
| 45 |
+
|
| 46 |
+
cdef SIZE_t[::1] samples # Sample indices in X, y
|
| 47 |
+
cdef SIZE_t n_samples # X.shape[0]
|
| 48 |
+
cdef double weighted_n_samples # Weighted number of samples
|
| 49 |
+
cdef SIZE_t[::1] features # Feature indices in X
|
| 50 |
+
cdef SIZE_t[::1] constant_features # Constant features indices
|
| 51 |
+
cdef SIZE_t n_features # X.shape[1]
|
| 52 |
+
cdef DTYPE_t[::1] feature_values # temp. array holding feature values
|
| 53 |
+
|
| 54 |
+
cdef SIZE_t start # Start position for the current node
|
| 55 |
+
cdef SIZE_t end # End position for the current node
|
| 56 |
+
|
| 57 |
+
cdef const DOUBLE_t[:, ::1] y
|
| 58 |
+
cdef const DOUBLE_t[:] sample_weight
|
| 59 |
+
|
| 60 |
+
# The samples vector `samples` is maintained by the Splitter object such
|
| 61 |
+
# that the samples contained in a node are contiguous. With this setting,
|
| 62 |
+
# `node_split` reorganizes the node samples `samples[start:end]` in two
|
| 63 |
+
# subsets `samples[start:pos]` and `samples[pos:end]`.
|
| 64 |
+
|
| 65 |
+
# The 1-d `features` array of size n_features contains the features
|
| 66 |
+
# indices and allows fast sampling without replacement of features.
|
| 67 |
+
|
| 68 |
+
# The 1-d `constant_features` array of size n_features holds in
|
| 69 |
+
# `constant_features[:n_constant_features]` the feature ids with
|
| 70 |
+
# constant values for all the samples that reached a specific node.
|
| 71 |
+
# The value `n_constant_features` is given by the parent node to its
|
| 72 |
+
# child nodes. The content of the range `[n_constant_features:]` is left
|
| 73 |
+
# undefined, but preallocated for performance reasons
|
| 74 |
+
# This allows optimization with depth-based tree building.
|
| 75 |
+
|
| 76 |
+
# Methods
|
| 77 |
+
cdef int init(
|
| 78 |
+
self,
|
| 79 |
+
object X,
|
| 80 |
+
const DOUBLE_t[:, ::1] y,
|
| 81 |
+
const DOUBLE_t[:] sample_weight
|
| 82 |
+
) except -1
|
| 83 |
+
|
| 84 |
+
cdef int node_reset(
|
| 85 |
+
self,
|
| 86 |
+
SIZE_t start,
|
| 87 |
+
SIZE_t end,
|
| 88 |
+
double* weighted_n_node_samples
|
| 89 |
+
) nogil except -1
|
| 90 |
+
|
| 91 |
+
cdef int node_split(
|
| 92 |
+
self,
|
| 93 |
+
double impurity, # Impurity of the node
|
| 94 |
+
SplitRecord* split,
|
| 95 |
+
SIZE_t* n_constant_features
|
| 96 |
+
) nogil except -1
|
| 97 |
+
|
| 98 |
+
cdef void node_value(self, double* dest) nogil
|
| 99 |
+
|
| 100 |
+
cdef double node_impurity(self) nogil
|
videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__init__.py
ADDED
|
File without changes
|
videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__pycache__/__init__.cpython-310.pyc
ADDED
|
Binary file (175 Bytes). View file
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__pycache__/test_export.cpython-310.pyc
ADDED
|
Binary file (11.9 kB). View file
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__pycache__/test_reingold_tilford.cpython-310.pyc
ADDED
|
Binary file (1.36 kB). View file
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/tree/tests/__pycache__/test_tree.cpython-310.pyc
ADDED
|
Binary file (58.6 kB). View file
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/tree/tests/test_export.py
ADDED
|
@@ -0,0 +1,512 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Testing for export functions of decision trees (sklearn.tree.export).
|
| 3 |
+
"""
|
| 4 |
+
from re import finditer, search
|
| 5 |
+
from textwrap import dedent
|
| 6 |
+
|
| 7 |
+
from numpy.random import RandomState
|
| 8 |
+
import pytest
|
| 9 |
+
|
| 10 |
+
from sklearn.base import is_classifier
|
| 11 |
+
from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor
|
| 12 |
+
from sklearn.ensemble import GradientBoostingClassifier
|
| 13 |
+
from sklearn.tree import export_graphviz, plot_tree, export_text
|
| 14 |
+
from io import StringIO
|
| 15 |
+
from sklearn.exceptions import NotFittedError
|
| 16 |
+
|
| 17 |
+
# toy sample
|
| 18 |
+
X = [[-2, -1], [-1, -1], [-1, -2], [1, 1], [1, 2], [2, 1]]
|
| 19 |
+
y = [-1, -1, -1, 1, 1, 1]
|
| 20 |
+
y2 = [[-1, 1], [-1, 1], [-1, 1], [1, 2], [1, 2], [1, 3]]
|
| 21 |
+
w = [1, 1, 1, 0.5, 0.5, 0.5]
|
| 22 |
+
y_degraded = [1, 1, 1, 1, 1, 1]
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
def test_graphviz_toy():
|
| 26 |
+
# Check correctness of export_graphviz
|
| 27 |
+
clf = DecisionTreeClassifier(
|
| 28 |
+
max_depth=3, min_samples_split=2, criterion="gini", random_state=2
|
| 29 |
+
)
|
| 30 |
+
clf.fit(X, y)
|
| 31 |
+
|
| 32 |
+
# Test export code
|
| 33 |
+
contents1 = export_graphviz(clf, out_file=None)
|
| 34 |
+
contents2 = (
|
| 35 |
+
"digraph Tree {\n"
|
| 36 |
+
'node [shape=box, fontname="helvetica"] ;\n'
|
| 37 |
+
'edge [fontname="helvetica"] ;\n'
|
| 38 |
+
'0 [label="x[0] <= 0.0\\ngini = 0.5\\nsamples = 6\\n'
|
| 39 |
+
'value = [3, 3]"] ;\n'
|
| 40 |
+
'1 [label="gini = 0.0\\nsamples = 3\\nvalue = [3, 0]"] ;\n'
|
| 41 |
+
"0 -> 1 [labeldistance=2.5, labelangle=45, "
|
| 42 |
+
'headlabel="True"] ;\n'
|
| 43 |
+
'2 [label="gini = 0.0\\nsamples = 3\\nvalue = [0, 3]"] ;\n'
|
| 44 |
+
"0 -> 2 [labeldistance=2.5, labelangle=-45, "
|
| 45 |
+
'headlabel="False"] ;\n'
|
| 46 |
+
"}"
|
| 47 |
+
)
|
| 48 |
+
|
| 49 |
+
assert contents1 == contents2
|
| 50 |
+
|
| 51 |
+
# Test with feature_names
|
| 52 |
+
contents1 = export_graphviz(
|
| 53 |
+
clf, feature_names=["feature0", "feature1"], out_file=None
|
| 54 |
+
)
|
| 55 |
+
contents2 = (
|
| 56 |
+
"digraph Tree {\n"
|
| 57 |
+
'node [shape=box, fontname="helvetica"] ;\n'
|
| 58 |
+
'edge [fontname="helvetica"] ;\n'
|
| 59 |
+
'0 [label="feature0 <= 0.0\\ngini = 0.5\\nsamples = 6\\n'
|
| 60 |
+
'value = [3, 3]"] ;\n'
|
| 61 |
+
'1 [label="gini = 0.0\\nsamples = 3\\nvalue = [3, 0]"] ;\n'
|
| 62 |
+
"0 -> 1 [labeldistance=2.5, labelangle=45, "
|
| 63 |
+
'headlabel="True"] ;\n'
|
| 64 |
+
'2 [label="gini = 0.0\\nsamples = 3\\nvalue = [0, 3]"] ;\n'
|
| 65 |
+
"0 -> 2 [labeldistance=2.5, labelangle=-45, "
|
| 66 |
+
'headlabel="False"] ;\n'
|
| 67 |
+
"}"
|
| 68 |
+
)
|
| 69 |
+
|
| 70 |
+
assert contents1 == contents2
|
| 71 |
+
|
| 72 |
+
# Test with class_names
|
| 73 |
+
contents1 = export_graphviz(clf, class_names=["yes", "no"], out_file=None)
|
| 74 |
+
contents2 = (
|
| 75 |
+
"digraph Tree {\n"
|
| 76 |
+
'node [shape=box, fontname="helvetica"] ;\n'
|
| 77 |
+
'edge [fontname="helvetica"] ;\n'
|
| 78 |
+
'0 [label="x[0] <= 0.0\\ngini = 0.5\\nsamples = 6\\n'
|
| 79 |
+
'value = [3, 3]\\nclass = yes"] ;\n'
|
| 80 |
+
'1 [label="gini = 0.0\\nsamples = 3\\nvalue = [3, 0]\\n'
|
| 81 |
+
'class = yes"] ;\n'
|
| 82 |
+
"0 -> 1 [labeldistance=2.5, labelangle=45, "
|
| 83 |
+
'headlabel="True"] ;\n'
|
| 84 |
+
'2 [label="gini = 0.0\\nsamples = 3\\nvalue = [0, 3]\\n'
|
| 85 |
+
'class = no"] ;\n'
|
| 86 |
+
"0 -> 2 [labeldistance=2.5, labelangle=-45, "
|
| 87 |
+
'headlabel="False"] ;\n'
|
| 88 |
+
"}"
|
| 89 |
+
)
|
| 90 |
+
|
| 91 |
+
assert contents1 == contents2
|
| 92 |
+
|
| 93 |
+
# Test plot_options
|
| 94 |
+
contents1 = export_graphviz(
|
| 95 |
+
clf,
|
| 96 |
+
filled=True,
|
| 97 |
+
impurity=False,
|
| 98 |
+
proportion=True,
|
| 99 |
+
special_characters=True,
|
| 100 |
+
rounded=True,
|
| 101 |
+
out_file=None,
|
| 102 |
+
fontname="sans",
|
| 103 |
+
)
|
| 104 |
+
contents2 = (
|
| 105 |
+
"digraph Tree {\n"
|
| 106 |
+
'node [shape=box, style="filled, rounded", color="black", '
|
| 107 |
+
'fontname="sans"] ;\n'
|
| 108 |
+
'edge [fontname="sans"] ;\n'
|
| 109 |
+
"0 [label=<x<SUB>0</SUB> ≤ 0.0<br/>samples = 100.0%<br/>"
|
| 110 |
+
'value = [0.5, 0.5]>, fillcolor="#ffffff"] ;\n'
|
| 111 |
+
"1 [label=<samples = 50.0%<br/>value = [1.0, 0.0]>, "
|
| 112 |
+
'fillcolor="#e58139"] ;\n'
|
| 113 |
+
"0 -> 1 [labeldistance=2.5, labelangle=45, "
|
| 114 |
+
'headlabel="True"] ;\n'
|
| 115 |
+
"2 [label=<samples = 50.0%<br/>value = [0.0, 1.0]>, "
|
| 116 |
+
'fillcolor="#399de5"] ;\n'
|
| 117 |
+
"0 -> 2 [labeldistance=2.5, labelangle=-45, "
|
| 118 |
+
'headlabel="False"] ;\n'
|
| 119 |
+
"}"
|
| 120 |
+
)
|
| 121 |
+
|
| 122 |
+
assert contents1 == contents2
|
| 123 |
+
|
| 124 |
+
# Test max_depth
|
| 125 |
+
contents1 = export_graphviz(clf, max_depth=0, class_names=True, out_file=None)
|
| 126 |
+
contents2 = (
|
| 127 |
+
"digraph Tree {\n"
|
| 128 |
+
'node [shape=box, fontname="helvetica"] ;\n'
|
| 129 |
+
'edge [fontname="helvetica"] ;\n'
|
| 130 |
+
'0 [label="x[0] <= 0.0\\ngini = 0.5\\nsamples = 6\\n'
|
| 131 |
+
'value = [3, 3]\\nclass = y[0]"] ;\n'
|
| 132 |
+
'1 [label="(...)"] ;\n'
|
| 133 |
+
"0 -> 1 ;\n"
|
| 134 |
+
'2 [label="(...)"] ;\n'
|
| 135 |
+
"0 -> 2 ;\n"
|
| 136 |
+
"}"
|
| 137 |
+
)
|
| 138 |
+
|
| 139 |
+
assert contents1 == contents2
|
| 140 |
+
|
| 141 |
+
# Test max_depth with plot_options
|
| 142 |
+
contents1 = export_graphviz(
|
| 143 |
+
clf, max_depth=0, filled=True, out_file=None, node_ids=True
|
| 144 |
+
)
|
| 145 |
+
contents2 = (
|
| 146 |
+
"digraph Tree {\n"
|
| 147 |
+
'node [shape=box, style="filled", color="black", '
|
| 148 |
+
'fontname="helvetica"] ;\n'
|
| 149 |
+
'edge [fontname="helvetica"] ;\n'
|
| 150 |
+
'0 [label="node #0\\nx[0] <= 0.0\\ngini = 0.5\\n'
|
| 151 |
+
'samples = 6\\nvalue = [3, 3]", fillcolor="#ffffff"] ;\n'
|
| 152 |
+
'1 [label="(...)", fillcolor="#C0C0C0"] ;\n'
|
| 153 |
+
"0 -> 1 ;\n"
|
| 154 |
+
'2 [label="(...)", fillcolor="#C0C0C0"] ;\n'
|
| 155 |
+
"0 -> 2 ;\n"
|
| 156 |
+
"}"
|
| 157 |
+
)
|
| 158 |
+
|
| 159 |
+
assert contents1 == contents2
|
| 160 |
+
|
| 161 |
+
# Test multi-output with weighted samples
|
| 162 |
+
clf = DecisionTreeClassifier(
|
| 163 |
+
max_depth=2, min_samples_split=2, criterion="gini", random_state=2
|
| 164 |
+
)
|
| 165 |
+
clf = clf.fit(X, y2, sample_weight=w)
|
| 166 |
+
|
| 167 |
+
contents1 = export_graphviz(clf, filled=True, impurity=False, out_file=None)
|
| 168 |
+
contents2 = (
|
| 169 |
+
"digraph Tree {\n"
|
| 170 |
+
'node [shape=box, style="filled", color="black", '
|
| 171 |
+
'fontname="helvetica"] ;\n'
|
| 172 |
+
'edge [fontname="helvetica"] ;\n'
|
| 173 |
+
'0 [label="x[0] <= 0.0\\nsamples = 6\\n'
|
| 174 |
+
"value = [[3.0, 1.5, 0.0]\\n"
|
| 175 |
+
'[3.0, 1.0, 0.5]]", fillcolor="#ffffff"] ;\n'
|
| 176 |
+
'1 [label="samples = 3\\nvalue = [[3, 0, 0]\\n'
|
| 177 |
+
'[3, 0, 0]]", fillcolor="#e58139"] ;\n'
|
| 178 |
+
"0 -> 1 [labeldistance=2.5, labelangle=45, "
|
| 179 |
+
'headlabel="True"] ;\n'
|
| 180 |
+
'2 [label="x[0] <= 1.5\\nsamples = 3\\n'
|
| 181 |
+
"value = [[0.0, 1.5, 0.0]\\n"
|
| 182 |
+
'[0.0, 1.0, 0.5]]", fillcolor="#f1bd97"] ;\n'
|
| 183 |
+
"0 -> 2 [labeldistance=2.5, labelangle=-45, "
|
| 184 |
+
'headlabel="False"] ;\n'
|
| 185 |
+
'3 [label="samples = 2\\nvalue = [[0, 1, 0]\\n'
|
| 186 |
+
'[0, 1, 0]]", fillcolor="#e58139"] ;\n'
|
| 187 |
+
"2 -> 3 ;\n"
|
| 188 |
+
'4 [label="samples = 1\\nvalue = [[0.0, 0.5, 0.0]\\n'
|
| 189 |
+
'[0.0, 0.0, 0.5]]", fillcolor="#e58139"] ;\n'
|
| 190 |
+
"2 -> 4 ;\n"
|
| 191 |
+
"}"
|
| 192 |
+
)
|
| 193 |
+
|
| 194 |
+
assert contents1 == contents2
|
| 195 |
+
|
| 196 |
+
# Test regression output with plot_options
|
| 197 |
+
clf = DecisionTreeRegressor(
|
| 198 |
+
max_depth=3, min_samples_split=2, criterion="squared_error", random_state=2
|
| 199 |
+
)
|
| 200 |
+
clf.fit(X, y)
|
| 201 |
+
|
| 202 |
+
contents1 = export_graphviz(
|
| 203 |
+
clf,
|
| 204 |
+
filled=True,
|
| 205 |
+
leaves_parallel=True,
|
| 206 |
+
out_file=None,
|
| 207 |
+
rotate=True,
|
| 208 |
+
rounded=True,
|
| 209 |
+
fontname="sans",
|
| 210 |
+
)
|
| 211 |
+
contents2 = (
|
| 212 |
+
"digraph Tree {\n"
|
| 213 |
+
'node [shape=box, style="filled, rounded", color="black", '
|
| 214 |
+
'fontname="sans"] ;\n'
|
| 215 |
+
"graph [ranksep=equally, splines=polyline] ;\n"
|
| 216 |
+
'edge [fontname="sans"] ;\n'
|
| 217 |
+
"rankdir=LR ;\n"
|
| 218 |
+
'0 [label="x[0] <= 0.0\\nsquared_error = 1.0\\nsamples = 6\\n'
|
| 219 |
+
'value = 0.0", fillcolor="#f2c09c"] ;\n'
|
| 220 |
+
'1 [label="squared_error = 0.0\\nsamples = 3\\'
|
| 221 |
+
'nvalue = -1.0", '
|
| 222 |
+
'fillcolor="#ffffff"] ;\n'
|
| 223 |
+
"0 -> 1 [labeldistance=2.5, labelangle=-45, "
|
| 224 |
+
'headlabel="True"] ;\n'
|
| 225 |
+
'2 [label="squared_error = 0.0\\nsamples = 3\\nvalue = 1.0", '
|
| 226 |
+
'fillcolor="#e58139"] ;\n'
|
| 227 |
+
"0 -> 2 [labeldistance=2.5, labelangle=45, "
|
| 228 |
+
'headlabel="False"] ;\n'
|
| 229 |
+
"{rank=same ; 0} ;\n"
|
| 230 |
+
"{rank=same ; 1; 2} ;\n"
|
| 231 |
+
"}"
|
| 232 |
+
)
|
| 233 |
+
|
| 234 |
+
assert contents1 == contents2
|
| 235 |
+
|
| 236 |
+
# Test classifier with degraded learning set
|
| 237 |
+
clf = DecisionTreeClassifier(max_depth=3)
|
| 238 |
+
clf.fit(X, y_degraded)
|
| 239 |
+
|
| 240 |
+
contents1 = export_graphviz(clf, filled=True, out_file=None)
|
| 241 |
+
contents2 = (
|
| 242 |
+
"digraph Tree {\n"
|
| 243 |
+
'node [shape=box, style="filled", color="black", '
|
| 244 |
+
'fontname="helvetica"] ;\n'
|
| 245 |
+
'edge [fontname="helvetica"] ;\n'
|
| 246 |
+
'0 [label="gini = 0.0\\nsamples = 6\\nvalue = 6.0", '
|
| 247 |
+
'fillcolor="#ffffff"] ;\n'
|
| 248 |
+
"}"
|
| 249 |
+
)
|
| 250 |
+
|
| 251 |
+
|
| 252 |
+
def test_graphviz_errors():
|
| 253 |
+
# Check for errors of export_graphviz
|
| 254 |
+
clf = DecisionTreeClassifier(max_depth=3, min_samples_split=2)
|
| 255 |
+
|
| 256 |
+
# Check not-fitted decision tree error
|
| 257 |
+
out = StringIO()
|
| 258 |
+
with pytest.raises(NotFittedError):
|
| 259 |
+
export_graphviz(clf, out)
|
| 260 |
+
|
| 261 |
+
clf.fit(X, y)
|
| 262 |
+
|
| 263 |
+
# Check if it errors when length of feature_names
|
| 264 |
+
# mismatches with number of features
|
| 265 |
+
message = "Length of feature_names, 1 does not match number of features, 2"
|
| 266 |
+
with pytest.raises(ValueError, match=message):
|
| 267 |
+
export_graphviz(clf, None, feature_names=["a"])
|
| 268 |
+
|
| 269 |
+
message = "Length of feature_names, 3 does not match number of features, 2"
|
| 270 |
+
with pytest.raises(ValueError, match=message):
|
| 271 |
+
export_graphviz(clf, None, feature_names=["a", "b", "c"])
|
| 272 |
+
|
| 273 |
+
# Check error when argument is not an estimator
|
| 274 |
+
message = "is not an estimator instance"
|
| 275 |
+
with pytest.raises(TypeError, match=message):
|
| 276 |
+
export_graphviz(clf.fit(X, y).tree_)
|
| 277 |
+
|
| 278 |
+
# Check class_names error
|
| 279 |
+
out = StringIO()
|
| 280 |
+
with pytest.raises(IndexError):
|
| 281 |
+
export_graphviz(clf, out, class_names=[])
|
| 282 |
+
|
| 283 |
+
# Check precision error
|
| 284 |
+
out = StringIO()
|
| 285 |
+
with pytest.raises(ValueError, match="should be greater or equal"):
|
| 286 |
+
export_graphviz(clf, out, precision=-1)
|
| 287 |
+
with pytest.raises(ValueError, match="should be an integer"):
|
| 288 |
+
export_graphviz(clf, out, precision="1")
|
| 289 |
+
|
| 290 |
+
|
| 291 |
+
def test_friedman_mse_in_graphviz():
|
| 292 |
+
clf = DecisionTreeRegressor(criterion="friedman_mse", random_state=0)
|
| 293 |
+
clf.fit(X, y)
|
| 294 |
+
dot_data = StringIO()
|
| 295 |
+
export_graphviz(clf, out_file=dot_data)
|
| 296 |
+
|
| 297 |
+
clf = GradientBoostingClassifier(n_estimators=2, random_state=0)
|
| 298 |
+
clf.fit(X, y)
|
| 299 |
+
for estimator in clf.estimators_:
|
| 300 |
+
export_graphviz(estimator[0], out_file=dot_data)
|
| 301 |
+
|
| 302 |
+
for finding in finditer(r"\[.*?samples.*?\]", dot_data.getvalue()):
|
| 303 |
+
assert "friedman_mse" in finding.group()
|
| 304 |
+
|
| 305 |
+
|
| 306 |
+
def test_precision():
|
| 307 |
+
|
| 308 |
+
rng_reg = RandomState(2)
|
| 309 |
+
rng_clf = RandomState(8)
|
| 310 |
+
for X, y, clf in zip(
|
| 311 |
+
(rng_reg.random_sample((5, 2)), rng_clf.random_sample((1000, 4))),
|
| 312 |
+
(rng_reg.random_sample((5,)), rng_clf.randint(2, size=(1000,))),
|
| 313 |
+
(
|
| 314 |
+
DecisionTreeRegressor(
|
| 315 |
+
criterion="friedman_mse", random_state=0, max_depth=1
|
| 316 |
+
),
|
| 317 |
+
DecisionTreeClassifier(max_depth=1, random_state=0),
|
| 318 |
+
),
|
| 319 |
+
):
|
| 320 |
+
|
| 321 |
+
clf.fit(X, y)
|
| 322 |
+
for precision in (4, 3):
|
| 323 |
+
dot_data = export_graphviz(
|
| 324 |
+
clf, out_file=None, precision=precision, proportion=True
|
| 325 |
+
)
|
| 326 |
+
|
| 327 |
+
# With the current random state, the impurity and the threshold
|
| 328 |
+
# will have the number of precision set in the export_graphviz
|
| 329 |
+
# function. We will check the number of precision with a strict
|
| 330 |
+
# equality. The value reported will have only 2 precision and
|
| 331 |
+
# therefore, only a less equal comparison will be done.
|
| 332 |
+
|
| 333 |
+
# check value
|
| 334 |
+
for finding in finditer(r"value = \d+\.\d+", dot_data):
|
| 335 |
+
assert len(search(r"\.\d+", finding.group()).group()) <= precision + 1
|
| 336 |
+
# check impurity
|
| 337 |
+
if is_classifier(clf):
|
| 338 |
+
pattern = r"gini = \d+\.\d+"
|
| 339 |
+
else:
|
| 340 |
+
pattern = r"friedman_mse = \d+\.\d+"
|
| 341 |
+
|
| 342 |
+
# check impurity
|
| 343 |
+
for finding in finditer(pattern, dot_data):
|
| 344 |
+
assert len(search(r"\.\d+", finding.group()).group()) == precision + 1
|
| 345 |
+
# check threshold
|
| 346 |
+
for finding in finditer(r"<= \d+\.\d+", dot_data):
|
| 347 |
+
assert len(search(r"\.\d+", finding.group()).group()) == precision + 1
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
def test_export_text_errors():
|
| 351 |
+
clf = DecisionTreeClassifier(max_depth=2, random_state=0)
|
| 352 |
+
clf.fit(X, y)
|
| 353 |
+
|
| 354 |
+
err_msg = "max_depth bust be >= 0, given -1"
|
| 355 |
+
with pytest.raises(ValueError, match=err_msg):
|
| 356 |
+
export_text(clf, max_depth=-1)
|
| 357 |
+
err_msg = "feature_names must contain 2 elements, got 1"
|
| 358 |
+
with pytest.raises(ValueError, match=err_msg):
|
| 359 |
+
export_text(clf, feature_names=["a"])
|
| 360 |
+
err_msg = "decimals must be >= 0, given -1"
|
| 361 |
+
with pytest.raises(ValueError, match=err_msg):
|
| 362 |
+
export_text(clf, decimals=-1)
|
| 363 |
+
err_msg = "spacing must be > 0, given 0"
|
| 364 |
+
with pytest.raises(ValueError, match=err_msg):
|
| 365 |
+
export_text(clf, spacing=0)
|
| 366 |
+
|
| 367 |
+
|
| 368 |
+
def test_export_text():
|
| 369 |
+
clf = DecisionTreeClassifier(max_depth=2, random_state=0)
|
| 370 |
+
clf.fit(X, y)
|
| 371 |
+
|
| 372 |
+
expected_report = dedent(
|
| 373 |
+
"""
|
| 374 |
+
|--- feature_1 <= 0.00
|
| 375 |
+
| |--- class: -1
|
| 376 |
+
|--- feature_1 > 0.00
|
| 377 |
+
| |--- class: 1
|
| 378 |
+
"""
|
| 379 |
+
).lstrip()
|
| 380 |
+
|
| 381 |
+
assert export_text(clf) == expected_report
|
| 382 |
+
# testing that leaves at level 1 are not truncated
|
| 383 |
+
assert export_text(clf, max_depth=0) == expected_report
|
| 384 |
+
# testing that the rest of the tree is truncated
|
| 385 |
+
assert export_text(clf, max_depth=10) == expected_report
|
| 386 |
+
|
| 387 |
+
expected_report = dedent(
|
| 388 |
+
"""
|
| 389 |
+
|--- b <= 0.00
|
| 390 |
+
| |--- class: -1
|
| 391 |
+
|--- b > 0.00
|
| 392 |
+
| |--- class: 1
|
| 393 |
+
"""
|
| 394 |
+
).lstrip()
|
| 395 |
+
assert export_text(clf, feature_names=["a", "b"]) == expected_report
|
| 396 |
+
|
| 397 |
+
expected_report = dedent(
|
| 398 |
+
"""
|
| 399 |
+
|--- feature_1 <= 0.00
|
| 400 |
+
| |--- weights: [3.00, 0.00] class: -1
|
| 401 |
+
|--- feature_1 > 0.00
|
| 402 |
+
| |--- weights: [0.00, 3.00] class: 1
|
| 403 |
+
"""
|
| 404 |
+
).lstrip()
|
| 405 |
+
assert export_text(clf, show_weights=True) == expected_report
|
| 406 |
+
|
| 407 |
+
expected_report = dedent(
|
| 408 |
+
"""
|
| 409 |
+
|- feature_1 <= 0.00
|
| 410 |
+
| |- class: -1
|
| 411 |
+
|- feature_1 > 0.00
|
| 412 |
+
| |- class: 1
|
| 413 |
+
"""
|
| 414 |
+
).lstrip()
|
| 415 |
+
assert export_text(clf, spacing=1) == expected_report
|
| 416 |
+
|
| 417 |
+
X_l = [[-2, -1], [-1, -1], [-1, -2], [1, 1], [1, 2], [2, 1], [-1, 1]]
|
| 418 |
+
y_l = [-1, -1, -1, 1, 1, 1, 2]
|
| 419 |
+
clf = DecisionTreeClassifier(max_depth=4, random_state=0)
|
| 420 |
+
clf.fit(X_l, y_l)
|
| 421 |
+
expected_report = dedent(
|
| 422 |
+
"""
|
| 423 |
+
|--- feature_1 <= 0.00
|
| 424 |
+
| |--- class: -1
|
| 425 |
+
|--- feature_1 > 0.00
|
| 426 |
+
| |--- truncated branch of depth 2
|
| 427 |
+
"""
|
| 428 |
+
).lstrip()
|
| 429 |
+
assert export_text(clf, max_depth=0) == expected_report
|
| 430 |
+
|
| 431 |
+
X_mo = [[-2, -1], [-1, -1], [-1, -2], [1, 1], [1, 2], [2, 1]]
|
| 432 |
+
y_mo = [[-1, -1], [-1, -1], [-1, -1], [1, 1], [1, 1], [1, 1]]
|
| 433 |
+
|
| 434 |
+
reg = DecisionTreeRegressor(max_depth=2, random_state=0)
|
| 435 |
+
reg.fit(X_mo, y_mo)
|
| 436 |
+
|
| 437 |
+
expected_report = dedent(
|
| 438 |
+
"""
|
| 439 |
+
|--- feature_1 <= 0.0
|
| 440 |
+
| |--- value: [-1.0, -1.0]
|
| 441 |
+
|--- feature_1 > 0.0
|
| 442 |
+
| |--- value: [1.0, 1.0]
|
| 443 |
+
"""
|
| 444 |
+
).lstrip()
|
| 445 |
+
assert export_text(reg, decimals=1) == expected_report
|
| 446 |
+
assert export_text(reg, decimals=1, show_weights=True) == expected_report
|
| 447 |
+
|
| 448 |
+
X_single = [[-2], [-1], [-1], [1], [1], [2]]
|
| 449 |
+
reg = DecisionTreeRegressor(max_depth=2, random_state=0)
|
| 450 |
+
reg.fit(X_single, y_mo)
|
| 451 |
+
|
| 452 |
+
expected_report = dedent(
|
| 453 |
+
"""
|
| 454 |
+
|--- first <= 0.0
|
| 455 |
+
| |--- value: [-1.0, -1.0]
|
| 456 |
+
|--- first > 0.0
|
| 457 |
+
| |--- value: [1.0, 1.0]
|
| 458 |
+
"""
|
| 459 |
+
).lstrip()
|
| 460 |
+
assert export_text(reg, decimals=1, feature_names=["first"]) == expected_report
|
| 461 |
+
assert (
|
| 462 |
+
export_text(reg, decimals=1, show_weights=True, feature_names=["first"])
|
| 463 |
+
== expected_report
|
| 464 |
+
)
|
| 465 |
+
|
| 466 |
+
|
| 467 |
+
def test_plot_tree_entropy(pyplot):
|
| 468 |
+
# mostly smoke tests
|
| 469 |
+
# Check correctness of export_graphviz for criterion = entropy
|
| 470 |
+
clf = DecisionTreeClassifier(
|
| 471 |
+
max_depth=3, min_samples_split=2, criterion="entropy", random_state=2
|
| 472 |
+
)
|
| 473 |
+
clf.fit(X, y)
|
| 474 |
+
|
| 475 |
+
# Test export code
|
| 476 |
+
feature_names = ["first feat", "sepal_width"]
|
| 477 |
+
nodes = plot_tree(clf, feature_names=feature_names)
|
| 478 |
+
assert len(nodes) == 3
|
| 479 |
+
assert (
|
| 480 |
+
nodes[0].get_text()
|
| 481 |
+
== "first feat <= 0.0\nentropy = 1.0\nsamples = 6\nvalue = [3, 3]"
|
| 482 |
+
)
|
| 483 |
+
assert nodes[1].get_text() == "entropy = 0.0\nsamples = 3\nvalue = [3, 0]"
|
| 484 |
+
assert nodes[2].get_text() == "entropy = 0.0\nsamples = 3\nvalue = [0, 3]"
|
| 485 |
+
|
| 486 |
+
|
| 487 |
+
def test_plot_tree_gini(pyplot):
|
| 488 |
+
# mostly smoke tests
|
| 489 |
+
# Check correctness of export_graphviz for criterion = gini
|
| 490 |
+
clf = DecisionTreeClassifier(
|
| 491 |
+
max_depth=3, min_samples_split=2, criterion="gini", random_state=2
|
| 492 |
+
)
|
| 493 |
+
clf.fit(X, y)
|
| 494 |
+
|
| 495 |
+
# Test export code
|
| 496 |
+
feature_names = ["first feat", "sepal_width"]
|
| 497 |
+
nodes = plot_tree(clf, feature_names=feature_names)
|
| 498 |
+
assert len(nodes) == 3
|
| 499 |
+
assert (
|
| 500 |
+
nodes[0].get_text()
|
| 501 |
+
== "first feat <= 0.0\ngini = 0.5\nsamples = 6\nvalue = [3, 3]"
|
| 502 |
+
)
|
| 503 |
+
assert nodes[1].get_text() == "gini = 0.0\nsamples = 3\nvalue = [3, 0]"
|
| 504 |
+
assert nodes[2].get_text() == "gini = 0.0\nsamples = 3\nvalue = [0, 3]"
|
| 505 |
+
|
| 506 |
+
|
| 507 |
+
def test_not_fitted_tree(pyplot):
|
| 508 |
+
|
| 509 |
+
# Testing if not fitted tree throws the correct error
|
| 510 |
+
clf = DecisionTreeRegressor()
|
| 511 |
+
with pytest.raises(NotFittedError):
|
| 512 |
+
plot_tree(clf)
|
videollama2/lib/python3.10/site-packages/sklearn/tree/tests/test_tree.py
ADDED
|
@@ -0,0 +1,2430 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Testing for the tree module (sklearn.tree).
|
| 3 |
+
"""
|
| 4 |
+
import copy
|
| 5 |
+
import pickle
|
| 6 |
+
from itertools import product
|
| 7 |
+
import struct
|
| 8 |
+
import io
|
| 9 |
+
import copyreg
|
| 10 |
+
|
| 11 |
+
import pytest
|
| 12 |
+
import numpy as np
|
| 13 |
+
from numpy.testing import assert_allclose
|
| 14 |
+
from scipy.sparse import csc_matrix
|
| 15 |
+
from scipy.sparse import csr_matrix
|
| 16 |
+
from scipy.sparse import coo_matrix
|
| 17 |
+
|
| 18 |
+
import joblib
|
| 19 |
+
from joblib.numpy_pickle import NumpyPickler
|
| 20 |
+
|
| 21 |
+
from sklearn.random_projection import _sparse_random_matrix
|
| 22 |
+
|
| 23 |
+
from sklearn.dummy import DummyRegressor
|
| 24 |
+
|
| 25 |
+
from sklearn.metrics import accuracy_score
|
| 26 |
+
from sklearn.metrics import mean_squared_error
|
| 27 |
+
from sklearn.metrics import mean_poisson_deviance
|
| 28 |
+
|
| 29 |
+
from sklearn.model_selection import train_test_split
|
| 30 |
+
|
| 31 |
+
from sklearn.utils._testing import assert_array_equal
|
| 32 |
+
from sklearn.utils._testing import assert_array_almost_equal
|
| 33 |
+
from sklearn.utils._testing import assert_almost_equal
|
| 34 |
+
from sklearn.utils._testing import create_memmap_backed_data
|
| 35 |
+
from sklearn.utils._testing import ignore_warnings
|
| 36 |
+
from sklearn.utils._testing import skip_if_32bit
|
| 37 |
+
|
| 38 |
+
from sklearn.utils.estimator_checks import check_sample_weights_invariance
|
| 39 |
+
from sklearn.utils.validation import check_random_state
|
| 40 |
+
from sklearn.utils import _IS_32BIT
|
| 41 |
+
|
| 42 |
+
from sklearn.exceptions import NotFittedError
|
| 43 |
+
|
| 44 |
+
from sklearn.tree import DecisionTreeClassifier
|
| 45 |
+
from sklearn.tree import DecisionTreeRegressor
|
| 46 |
+
from sklearn.tree import ExtraTreeClassifier
|
| 47 |
+
from sklearn.tree import ExtraTreeRegressor
|
| 48 |
+
|
| 49 |
+
from sklearn import tree
|
| 50 |
+
from sklearn.tree._tree import TREE_LEAF, TREE_UNDEFINED
|
| 51 |
+
from sklearn.tree._tree import Tree as CythonTree
|
| 52 |
+
from sklearn.tree._tree import _check_n_classes
|
| 53 |
+
from sklearn.tree._tree import _check_value_ndarray
|
| 54 |
+
from sklearn.tree._tree import _check_node_ndarray
|
| 55 |
+
from sklearn.tree._tree import NODE_DTYPE
|
| 56 |
+
|
| 57 |
+
from sklearn.tree._classes import CRITERIA_CLF
|
| 58 |
+
from sklearn.tree._classes import CRITERIA_REG
|
| 59 |
+
from sklearn import datasets
|
| 60 |
+
|
| 61 |
+
from sklearn.utils import compute_sample_weight
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
CLF_CRITERIONS = ("gini", "log_loss")
|
| 65 |
+
REG_CRITERIONS = ("squared_error", "absolute_error", "friedman_mse", "poisson")
|
| 66 |
+
|
| 67 |
+
CLF_TREES = {
|
| 68 |
+
"DecisionTreeClassifier": DecisionTreeClassifier,
|
| 69 |
+
"ExtraTreeClassifier": ExtraTreeClassifier,
|
| 70 |
+
}
|
| 71 |
+
|
| 72 |
+
REG_TREES = {
|
| 73 |
+
"DecisionTreeRegressor": DecisionTreeRegressor,
|
| 74 |
+
"ExtraTreeRegressor": ExtraTreeRegressor,
|
| 75 |
+
}
|
| 76 |
+
|
| 77 |
+
ALL_TREES: dict = dict()
|
| 78 |
+
ALL_TREES.update(CLF_TREES)
|
| 79 |
+
ALL_TREES.update(REG_TREES)
|
| 80 |
+
|
| 81 |
+
SPARSE_TREES = [
|
| 82 |
+
"DecisionTreeClassifier",
|
| 83 |
+
"DecisionTreeRegressor",
|
| 84 |
+
"ExtraTreeClassifier",
|
| 85 |
+
"ExtraTreeRegressor",
|
| 86 |
+
]
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
X_small = np.array(
|
| 90 |
+
[
|
| 91 |
+
[0, 0, 4, 0, 0, 0, 1, -14, 0, -4, 0, 0, 0, 0],
|
| 92 |
+
[0, 0, 5, 3, 0, -4, 0, 0, 1, -5, 0.2, 0, 4, 1],
|
| 93 |
+
[-1, -1, 0, 0, -4.5, 0, 0, 2.1, 1, 0, 0, -4.5, 0, 1],
|
| 94 |
+
[-1, -1, 0, -1.2, 0, 0, 0, 0, 0, 0, 0.2, 0, 0, 1],
|
| 95 |
+
[-1, -1, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 1],
|
| 96 |
+
[-1, -2, 0, 4, -3, 10, 4, 0, -3.2, 0, 4, 3, -4, 1],
|
| 97 |
+
[2.11, 0, -6, -0.5, 0, 11, 0, 0, -3.2, 6, 0.5, 0, -3, 1],
|
| 98 |
+
[2.11, 0, -6, -0.5, 0, 11, 0, 0, -3.2, 6, 0, 0, -2, 1],
|
| 99 |
+
[2.11, 8, -6, -0.5, 0, 11, 0, 0, -3.2, 6, 0, 0, -2, 1],
|
| 100 |
+
[2.11, 8, -6, -0.5, 0, 11, 0, 0, -3.2, 6, 0.5, 0, -1, 0],
|
| 101 |
+
[2, 8, 5, 1, 0.5, -4, 10, 0, 1, -5, 3, 0, 2, 0],
|
| 102 |
+
[2, 0, 1, 1, 1, -1, 1, 0, 0, -2, 3, 0, 1, 0],
|
| 103 |
+
[2, 0, 1, 2, 3, -1, 10, 2, 0, -1, 1, 2, 2, 0],
|
| 104 |
+
[1, 1, 0, 2, 2, -1, 1, 2, 0, -5, 1, 2, 3, 0],
|
| 105 |
+
[3, 1, 0, 3, 0, -4, 10, 0, 1, -5, 3, 0, 3, 1],
|
| 106 |
+
[2.11, 8, -6, -0.5, 0, 1, 0, 0, -3.2, 6, 0.5, 0, -3, 1],
|
| 107 |
+
[2.11, 8, -6, -0.5, 0, 1, 0, 0, -3.2, 6, 1.5, 1, -1, -1],
|
| 108 |
+
[2.11, 8, -6, -0.5, 0, 10, 0, 0, -3.2, 6, 0.5, 0, -1, -1],
|
| 109 |
+
[2, 0, 5, 1, 0.5, -2, 10, 0, 1, -5, 3, 1, 0, -1],
|
| 110 |
+
[2, 0, 1, 1, 1, -2, 1, 0, 0, -2, 0, 0, 0, 1],
|
| 111 |
+
[2, 1, 1, 1, 2, -1, 10, 2, 0, -1, 0, 2, 1, 1],
|
| 112 |
+
[1, 1, 0, 0, 1, -3, 1, 2, 0, -5, 1, 2, 1, 1],
|
| 113 |
+
[3, 1, 0, 1, 0, -4, 1, 0, 1, -2, 0, 0, 1, 0],
|
| 114 |
+
]
|
| 115 |
+
)
|
| 116 |
+
|
| 117 |
+
y_small = [1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0]
|
| 118 |
+
y_small_reg = [
|
| 119 |
+
1.0,
|
| 120 |
+
2.1,
|
| 121 |
+
1.2,
|
| 122 |
+
0.05,
|
| 123 |
+
10,
|
| 124 |
+
2.4,
|
| 125 |
+
3.1,
|
| 126 |
+
1.01,
|
| 127 |
+
0.01,
|
| 128 |
+
2.98,
|
| 129 |
+
3.1,
|
| 130 |
+
1.1,
|
| 131 |
+
0.0,
|
| 132 |
+
1.2,
|
| 133 |
+
2,
|
| 134 |
+
11,
|
| 135 |
+
0,
|
| 136 |
+
0,
|
| 137 |
+
4.5,
|
| 138 |
+
0.201,
|
| 139 |
+
1.06,
|
| 140 |
+
0.9,
|
| 141 |
+
0,
|
| 142 |
+
]
|
| 143 |
+
|
| 144 |
+
# toy sample
|
| 145 |
+
X = [[-2, -1], [-1, -1], [-1, -2], [1, 1], [1, 2], [2, 1]]
|
| 146 |
+
y = [-1, -1, -1, 1, 1, 1]
|
| 147 |
+
T = [[-1, -1], [2, 2], [3, 2]]
|
| 148 |
+
true_result = [-1, 1, 1]
|
| 149 |
+
|
| 150 |
+
# also load the iris dataset
|
| 151 |
+
# and randomly permute it
|
| 152 |
+
iris = datasets.load_iris()
|
| 153 |
+
rng = np.random.RandomState(1)
|
| 154 |
+
perm = rng.permutation(iris.target.size)
|
| 155 |
+
iris.data = iris.data[perm]
|
| 156 |
+
iris.target = iris.target[perm]
|
| 157 |
+
|
| 158 |
+
# also load the diabetes dataset
|
| 159 |
+
# and randomly permute it
|
| 160 |
+
diabetes = datasets.load_diabetes()
|
| 161 |
+
perm = rng.permutation(diabetes.target.size)
|
| 162 |
+
diabetes.data = diabetes.data[perm]
|
| 163 |
+
diabetes.target = diabetes.target[perm]
|
| 164 |
+
|
| 165 |
+
digits = datasets.load_digits()
|
| 166 |
+
perm = rng.permutation(digits.target.size)
|
| 167 |
+
digits.data = digits.data[perm]
|
| 168 |
+
digits.target = digits.target[perm]
|
| 169 |
+
|
| 170 |
+
random_state = check_random_state(0)
|
| 171 |
+
X_multilabel, y_multilabel = datasets.make_multilabel_classification(
|
| 172 |
+
random_state=0, n_samples=30, n_features=10
|
| 173 |
+
)
|
| 174 |
+
|
| 175 |
+
# NB: despite their names X_sparse_* are numpy arrays (and not sparse matrices)
|
| 176 |
+
X_sparse_pos = random_state.uniform(size=(20, 5))
|
| 177 |
+
X_sparse_pos[X_sparse_pos <= 0.8] = 0.0
|
| 178 |
+
y_random = random_state.randint(0, 4, size=(20,))
|
| 179 |
+
X_sparse_mix = _sparse_random_matrix(20, 10, density=0.25, random_state=0).toarray()
|
| 180 |
+
|
| 181 |
+
|
| 182 |
+
DATASETS = {
|
| 183 |
+
"iris": {"X": iris.data, "y": iris.target},
|
| 184 |
+
"diabetes": {"X": diabetes.data, "y": diabetes.target},
|
| 185 |
+
"digits": {"X": digits.data, "y": digits.target},
|
| 186 |
+
"toy": {"X": X, "y": y},
|
| 187 |
+
"clf_small": {"X": X_small, "y": y_small},
|
| 188 |
+
"reg_small": {"X": X_small, "y": y_small_reg},
|
| 189 |
+
"multilabel": {"X": X_multilabel, "y": y_multilabel},
|
| 190 |
+
"sparse-pos": {"X": X_sparse_pos, "y": y_random},
|
| 191 |
+
"sparse-neg": {"X": -X_sparse_pos, "y": y_random},
|
| 192 |
+
"sparse-mix": {"X": X_sparse_mix, "y": y_random},
|
| 193 |
+
"zeros": {"X": np.zeros((20, 3)), "y": y_random},
|
| 194 |
+
}
|
| 195 |
+
|
| 196 |
+
for name in DATASETS:
|
| 197 |
+
DATASETS[name]["X_sparse"] = csc_matrix(DATASETS[name]["X"])
|
| 198 |
+
|
| 199 |
+
|
| 200 |
+
def assert_tree_equal(d, s, message):
|
| 201 |
+
assert (
|
| 202 |
+
s.node_count == d.node_count
|
| 203 |
+
), "{0}: inequal number of node ({1} != {2})".format(
|
| 204 |
+
message, s.node_count, d.node_count
|
| 205 |
+
)
|
| 206 |
+
|
| 207 |
+
assert_array_equal(
|
| 208 |
+
d.children_right, s.children_right, message + ": inequal children_right"
|
| 209 |
+
)
|
| 210 |
+
assert_array_equal(
|
| 211 |
+
d.children_left, s.children_left, message + ": inequal children_left"
|
| 212 |
+
)
|
| 213 |
+
|
| 214 |
+
external = d.children_right == TREE_LEAF
|
| 215 |
+
internal = np.logical_not(external)
|
| 216 |
+
|
| 217 |
+
assert_array_equal(
|
| 218 |
+
d.feature[internal], s.feature[internal], message + ": inequal features"
|
| 219 |
+
)
|
| 220 |
+
assert_array_equal(
|
| 221 |
+
d.threshold[internal], s.threshold[internal], message + ": inequal threshold"
|
| 222 |
+
)
|
| 223 |
+
assert_array_equal(
|
| 224 |
+
d.n_node_samples.sum(),
|
| 225 |
+
s.n_node_samples.sum(),
|
| 226 |
+
message + ": inequal sum(n_node_samples)",
|
| 227 |
+
)
|
| 228 |
+
assert_array_equal(
|
| 229 |
+
d.n_node_samples, s.n_node_samples, message + ": inequal n_node_samples"
|
| 230 |
+
)
|
| 231 |
+
|
| 232 |
+
assert_almost_equal(d.impurity, s.impurity, err_msg=message + ": inequal impurity")
|
| 233 |
+
|
| 234 |
+
assert_array_almost_equal(
|
| 235 |
+
d.value[external], s.value[external], err_msg=message + ": inequal value"
|
| 236 |
+
)
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
def test_classification_toy():
|
| 240 |
+
# Check classification on a toy dataset.
|
| 241 |
+
for name, Tree in CLF_TREES.items():
|
| 242 |
+
clf = Tree(random_state=0)
|
| 243 |
+
clf.fit(X, y)
|
| 244 |
+
assert_array_equal(clf.predict(T), true_result, "Failed with {0}".format(name))
|
| 245 |
+
|
| 246 |
+
clf = Tree(max_features=1, random_state=1)
|
| 247 |
+
clf.fit(X, y)
|
| 248 |
+
assert_array_equal(clf.predict(T), true_result, "Failed with {0}".format(name))
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def test_weighted_classification_toy():
|
| 252 |
+
# Check classification on a weighted toy dataset.
|
| 253 |
+
for name, Tree in CLF_TREES.items():
|
| 254 |
+
clf = Tree(random_state=0)
|
| 255 |
+
|
| 256 |
+
clf.fit(X, y, sample_weight=np.ones(len(X)))
|
| 257 |
+
assert_array_equal(clf.predict(T), true_result, "Failed with {0}".format(name))
|
| 258 |
+
|
| 259 |
+
clf.fit(X, y, sample_weight=np.full(len(X), 0.5))
|
| 260 |
+
assert_array_equal(clf.predict(T), true_result, "Failed with {0}".format(name))
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
@pytest.mark.parametrize("Tree", REG_TREES.values())
|
| 264 |
+
@pytest.mark.parametrize("criterion", REG_CRITERIONS)
|
| 265 |
+
def test_regression_toy(Tree, criterion):
|
| 266 |
+
# Check regression on a toy dataset.
|
| 267 |
+
if criterion == "poisson":
|
| 268 |
+
# make target positive while not touching the original y and
|
| 269 |
+
# true_result
|
| 270 |
+
a = np.abs(np.min(y)) + 1
|
| 271 |
+
y_train = np.array(y) + a
|
| 272 |
+
y_test = np.array(true_result) + a
|
| 273 |
+
else:
|
| 274 |
+
y_train = y
|
| 275 |
+
y_test = true_result
|
| 276 |
+
|
| 277 |
+
reg = Tree(criterion=criterion, random_state=1)
|
| 278 |
+
reg.fit(X, y_train)
|
| 279 |
+
assert_allclose(reg.predict(T), y_test)
|
| 280 |
+
|
| 281 |
+
clf = Tree(criterion=criterion, max_features=1, random_state=1)
|
| 282 |
+
clf.fit(X, y_train)
|
| 283 |
+
assert_allclose(reg.predict(T), y_test)
|
| 284 |
+
|
| 285 |
+
|
| 286 |
+
def test_xor():
|
| 287 |
+
# Check on a XOR problem
|
| 288 |
+
y = np.zeros((10, 10))
|
| 289 |
+
y[:5, :5] = 1
|
| 290 |
+
y[5:, 5:] = 1
|
| 291 |
+
|
| 292 |
+
gridx, gridy = np.indices(y.shape)
|
| 293 |
+
|
| 294 |
+
X = np.vstack([gridx.ravel(), gridy.ravel()]).T
|
| 295 |
+
y = y.ravel()
|
| 296 |
+
|
| 297 |
+
for name, Tree in CLF_TREES.items():
|
| 298 |
+
clf = Tree(random_state=0)
|
| 299 |
+
clf.fit(X, y)
|
| 300 |
+
assert clf.score(X, y) == 1.0, "Failed with {0}".format(name)
|
| 301 |
+
|
| 302 |
+
clf = Tree(random_state=0, max_features=1)
|
| 303 |
+
clf.fit(X, y)
|
| 304 |
+
assert clf.score(X, y) == 1.0, "Failed with {0}".format(name)
|
| 305 |
+
|
| 306 |
+
|
| 307 |
+
def test_iris():
|
| 308 |
+
# Check consistency on dataset iris.
|
| 309 |
+
for (name, Tree), criterion in product(CLF_TREES.items(), CLF_CRITERIONS):
|
| 310 |
+
clf = Tree(criterion=criterion, random_state=0)
|
| 311 |
+
clf.fit(iris.data, iris.target)
|
| 312 |
+
score = accuracy_score(clf.predict(iris.data), iris.target)
|
| 313 |
+
assert score > 0.9, "Failed with {0}, criterion = {1} and score = {2}".format(
|
| 314 |
+
name, criterion, score
|
| 315 |
+
)
|
| 316 |
+
|
| 317 |
+
clf = Tree(criterion=criterion, max_features=2, random_state=0)
|
| 318 |
+
clf.fit(iris.data, iris.target)
|
| 319 |
+
score = accuracy_score(clf.predict(iris.data), iris.target)
|
| 320 |
+
assert score > 0.5, "Failed with {0}, criterion = {1} and score = {2}".format(
|
| 321 |
+
name, criterion, score
|
| 322 |
+
)
|
| 323 |
+
|
| 324 |
+
|
| 325 |
+
@pytest.mark.parametrize("name, Tree", REG_TREES.items())
|
| 326 |
+
@pytest.mark.parametrize("criterion", REG_CRITERIONS)
|
| 327 |
+
def test_diabetes_overfit(name, Tree, criterion):
|
| 328 |
+
# check consistency of overfitted trees on the diabetes dataset
|
| 329 |
+
# since the trees will overfit, we expect an MSE of 0
|
| 330 |
+
reg = Tree(criterion=criterion, random_state=0)
|
| 331 |
+
reg.fit(diabetes.data, diabetes.target)
|
| 332 |
+
score = mean_squared_error(diabetes.target, reg.predict(diabetes.data))
|
| 333 |
+
assert score == pytest.approx(
|
| 334 |
+
0
|
| 335 |
+
), f"Failed with {name}, criterion = {criterion} and score = {score}"
|
| 336 |
+
|
| 337 |
+
|
| 338 |
+
@skip_if_32bit
|
| 339 |
+
@pytest.mark.parametrize("name, Tree", REG_TREES.items())
|
| 340 |
+
@pytest.mark.parametrize(
|
| 341 |
+
"criterion, max_depth, metric, max_loss",
|
| 342 |
+
[
|
| 343 |
+
("squared_error", 15, mean_squared_error, 60),
|
| 344 |
+
("absolute_error", 20, mean_squared_error, 60),
|
| 345 |
+
("friedman_mse", 15, mean_squared_error, 60),
|
| 346 |
+
("poisson", 15, mean_poisson_deviance, 30),
|
| 347 |
+
],
|
| 348 |
+
)
|
| 349 |
+
def test_diabetes_underfit(name, Tree, criterion, max_depth, metric, max_loss):
|
| 350 |
+
# check consistency of trees when the depth and the number of features are
|
| 351 |
+
# limited
|
| 352 |
+
|
| 353 |
+
reg = Tree(criterion=criterion, max_depth=max_depth, max_features=6, random_state=0)
|
| 354 |
+
reg.fit(diabetes.data, diabetes.target)
|
| 355 |
+
loss = metric(diabetes.target, reg.predict(diabetes.data))
|
| 356 |
+
assert 0 < loss < max_loss
|
| 357 |
+
|
| 358 |
+
|
| 359 |
+
def test_probability():
|
| 360 |
+
# Predict probabilities using DecisionTreeClassifier.
|
| 361 |
+
|
| 362 |
+
for name, Tree in CLF_TREES.items():
|
| 363 |
+
clf = Tree(max_depth=1, max_features=1, random_state=42)
|
| 364 |
+
clf.fit(iris.data, iris.target)
|
| 365 |
+
|
| 366 |
+
prob_predict = clf.predict_proba(iris.data)
|
| 367 |
+
assert_array_almost_equal(
|
| 368 |
+
np.sum(prob_predict, 1),
|
| 369 |
+
np.ones(iris.data.shape[0]),
|
| 370 |
+
err_msg="Failed with {0}".format(name),
|
| 371 |
+
)
|
| 372 |
+
assert_array_equal(
|
| 373 |
+
np.argmax(prob_predict, 1),
|
| 374 |
+
clf.predict(iris.data),
|
| 375 |
+
err_msg="Failed with {0}".format(name),
|
| 376 |
+
)
|
| 377 |
+
assert_almost_equal(
|
| 378 |
+
clf.predict_proba(iris.data),
|
| 379 |
+
np.exp(clf.predict_log_proba(iris.data)),
|
| 380 |
+
8,
|
| 381 |
+
err_msg="Failed with {0}".format(name),
|
| 382 |
+
)
|
| 383 |
+
|
| 384 |
+
|
| 385 |
+
def test_arrayrepr():
|
| 386 |
+
# Check the array representation.
|
| 387 |
+
# Check resize
|
| 388 |
+
X = np.arange(10000)[:, np.newaxis]
|
| 389 |
+
y = np.arange(10000)
|
| 390 |
+
|
| 391 |
+
for name, Tree in REG_TREES.items():
|
| 392 |
+
reg = Tree(max_depth=None, random_state=0)
|
| 393 |
+
reg.fit(X, y)
|
| 394 |
+
|
| 395 |
+
|
| 396 |
+
def test_pure_set():
|
| 397 |
+
# Check when y is pure.
|
| 398 |
+
X = [[-2, -1], [-1, -1], [-1, -2], [1, 1], [1, 2], [2, 1]]
|
| 399 |
+
y = [1, 1, 1, 1, 1, 1]
|
| 400 |
+
|
| 401 |
+
for name, TreeClassifier in CLF_TREES.items():
|
| 402 |
+
clf = TreeClassifier(random_state=0)
|
| 403 |
+
clf.fit(X, y)
|
| 404 |
+
assert_array_equal(clf.predict(X), y, err_msg="Failed with {0}".format(name))
|
| 405 |
+
|
| 406 |
+
for name, TreeRegressor in REG_TREES.items():
|
| 407 |
+
reg = TreeRegressor(random_state=0)
|
| 408 |
+
reg.fit(X, y)
|
| 409 |
+
assert_almost_equal(reg.predict(X), y, err_msg="Failed with {0}".format(name))
|
| 410 |
+
|
| 411 |
+
|
| 412 |
+
def test_numerical_stability():
|
| 413 |
+
# Check numerical stability.
|
| 414 |
+
X = np.array(
|
| 415 |
+
[
|
| 416 |
+
[152.08097839, 140.40744019, 129.75102234, 159.90493774],
|
| 417 |
+
[142.50700378, 135.81935120, 117.82884979, 162.75781250],
|
| 418 |
+
[127.28772736, 140.40744019, 129.75102234, 159.90493774],
|
| 419 |
+
[132.37025452, 143.71923828, 138.35694885, 157.84558105],
|
| 420 |
+
[103.10237122, 143.71928406, 138.35696411, 157.84559631],
|
| 421 |
+
[127.71276855, 143.71923828, 138.35694885, 157.84558105],
|
| 422 |
+
[120.91514587, 140.40744019, 129.75102234, 159.90493774],
|
| 423 |
+
]
|
| 424 |
+
)
|
| 425 |
+
|
| 426 |
+
y = np.array([1.0, 0.70209277, 0.53896582, 0.0, 0.90914464, 0.48026916, 0.49622521])
|
| 427 |
+
|
| 428 |
+
with np.errstate(all="raise"):
|
| 429 |
+
for name, Tree in REG_TREES.items():
|
| 430 |
+
reg = Tree(random_state=0)
|
| 431 |
+
reg.fit(X, y)
|
| 432 |
+
reg.fit(X, -y)
|
| 433 |
+
reg.fit(-X, y)
|
| 434 |
+
reg.fit(-X, -y)
|
| 435 |
+
|
| 436 |
+
|
| 437 |
+
def test_importances():
|
| 438 |
+
# Check variable importances.
|
| 439 |
+
X, y = datasets.make_classification(
|
| 440 |
+
n_samples=5000,
|
| 441 |
+
n_features=10,
|
| 442 |
+
n_informative=3,
|
| 443 |
+
n_redundant=0,
|
| 444 |
+
n_repeated=0,
|
| 445 |
+
shuffle=False,
|
| 446 |
+
random_state=0,
|
| 447 |
+
)
|
| 448 |
+
|
| 449 |
+
for name, Tree in CLF_TREES.items():
|
| 450 |
+
clf = Tree(random_state=0)
|
| 451 |
+
|
| 452 |
+
clf.fit(X, y)
|
| 453 |
+
importances = clf.feature_importances_
|
| 454 |
+
n_important = np.sum(importances > 0.1)
|
| 455 |
+
|
| 456 |
+
assert importances.shape[0] == 10, "Failed with {0}".format(name)
|
| 457 |
+
assert n_important == 3, "Failed with {0}".format(name)
|
| 458 |
+
|
| 459 |
+
# Check on iris that importances are the same for all builders
|
| 460 |
+
clf = DecisionTreeClassifier(random_state=0)
|
| 461 |
+
clf.fit(iris.data, iris.target)
|
| 462 |
+
clf2 = DecisionTreeClassifier(random_state=0, max_leaf_nodes=len(iris.data))
|
| 463 |
+
clf2.fit(iris.data, iris.target)
|
| 464 |
+
|
| 465 |
+
assert_array_equal(clf.feature_importances_, clf2.feature_importances_)
|
| 466 |
+
|
| 467 |
+
|
| 468 |
+
def test_importances_raises():
|
| 469 |
+
# Check if variable importance before fit raises ValueError.
|
| 470 |
+
clf = DecisionTreeClassifier()
|
| 471 |
+
with pytest.raises(ValueError):
|
| 472 |
+
getattr(clf, "feature_importances_")
|
| 473 |
+
|
| 474 |
+
|
| 475 |
+
def test_importances_gini_equal_squared_error():
|
| 476 |
+
# Check that gini is equivalent to squared_error for binary output variable
|
| 477 |
+
|
| 478 |
+
X, y = datasets.make_classification(
|
| 479 |
+
n_samples=2000,
|
| 480 |
+
n_features=10,
|
| 481 |
+
n_informative=3,
|
| 482 |
+
n_redundant=0,
|
| 483 |
+
n_repeated=0,
|
| 484 |
+
shuffle=False,
|
| 485 |
+
random_state=0,
|
| 486 |
+
)
|
| 487 |
+
|
| 488 |
+
# The gini index and the mean square error (variance) might differ due
|
| 489 |
+
# to numerical instability. Since those instabilities mainly occurs at
|
| 490 |
+
# high tree depth, we restrict this maximal depth.
|
| 491 |
+
clf = DecisionTreeClassifier(criterion="gini", max_depth=5, random_state=0).fit(
|
| 492 |
+
X, y
|
| 493 |
+
)
|
| 494 |
+
reg = DecisionTreeRegressor(
|
| 495 |
+
criterion="squared_error", max_depth=5, random_state=0
|
| 496 |
+
).fit(X, y)
|
| 497 |
+
|
| 498 |
+
assert_almost_equal(clf.feature_importances_, reg.feature_importances_)
|
| 499 |
+
assert_array_equal(clf.tree_.feature, reg.tree_.feature)
|
| 500 |
+
assert_array_equal(clf.tree_.children_left, reg.tree_.children_left)
|
| 501 |
+
assert_array_equal(clf.tree_.children_right, reg.tree_.children_right)
|
| 502 |
+
assert_array_equal(clf.tree_.n_node_samples, reg.tree_.n_node_samples)
|
| 503 |
+
|
| 504 |
+
|
| 505 |
+
# TODO(1.3): Remove warning filter
|
| 506 |
+
@pytest.mark.filterwarnings("ignore:`max_features='auto'` has been deprecated in 1.1")
|
| 507 |
+
def test_max_features():
|
| 508 |
+
# Check max_features.
|
| 509 |
+
for name, TreeRegressor in REG_TREES.items():
|
| 510 |
+
reg = TreeRegressor(max_features="auto")
|
| 511 |
+
reg.fit(diabetes.data, diabetes.target)
|
| 512 |
+
assert reg.max_features_ == diabetes.data.shape[1]
|
| 513 |
+
|
| 514 |
+
for name, TreeClassifier in CLF_TREES.items():
|
| 515 |
+
clf = TreeClassifier(max_features="auto")
|
| 516 |
+
clf.fit(iris.data, iris.target)
|
| 517 |
+
assert clf.max_features_ == 2
|
| 518 |
+
|
| 519 |
+
for name, TreeEstimator in ALL_TREES.items():
|
| 520 |
+
est = TreeEstimator(max_features="sqrt")
|
| 521 |
+
est.fit(iris.data, iris.target)
|
| 522 |
+
assert est.max_features_ == int(np.sqrt(iris.data.shape[1]))
|
| 523 |
+
|
| 524 |
+
est = TreeEstimator(max_features="log2")
|
| 525 |
+
est.fit(iris.data, iris.target)
|
| 526 |
+
assert est.max_features_ == int(np.log2(iris.data.shape[1]))
|
| 527 |
+
|
| 528 |
+
est = TreeEstimator(max_features=1)
|
| 529 |
+
est.fit(iris.data, iris.target)
|
| 530 |
+
assert est.max_features_ == 1
|
| 531 |
+
|
| 532 |
+
est = TreeEstimator(max_features=3)
|
| 533 |
+
est.fit(iris.data, iris.target)
|
| 534 |
+
assert est.max_features_ == 3
|
| 535 |
+
|
| 536 |
+
est = TreeEstimator(max_features=0.01)
|
| 537 |
+
est.fit(iris.data, iris.target)
|
| 538 |
+
assert est.max_features_ == 1
|
| 539 |
+
|
| 540 |
+
est = TreeEstimator(max_features=0.5)
|
| 541 |
+
est.fit(iris.data, iris.target)
|
| 542 |
+
assert est.max_features_ == int(0.5 * iris.data.shape[1])
|
| 543 |
+
|
| 544 |
+
est = TreeEstimator(max_features=1.0)
|
| 545 |
+
est.fit(iris.data, iris.target)
|
| 546 |
+
assert est.max_features_ == iris.data.shape[1]
|
| 547 |
+
|
| 548 |
+
est = TreeEstimator(max_features=None)
|
| 549 |
+
est.fit(iris.data, iris.target)
|
| 550 |
+
assert est.max_features_ == iris.data.shape[1]
|
| 551 |
+
|
| 552 |
+
|
| 553 |
+
def test_error():
|
| 554 |
+
# Test that it gives proper exception on deficient input.
|
| 555 |
+
for name, TreeEstimator in CLF_TREES.items():
|
| 556 |
+
# predict before fit
|
| 557 |
+
est = TreeEstimator()
|
| 558 |
+
with pytest.raises(NotFittedError):
|
| 559 |
+
est.predict_proba(X)
|
| 560 |
+
|
| 561 |
+
est.fit(X, y)
|
| 562 |
+
X2 = [[-2, -1, 1]] # wrong feature shape for sample
|
| 563 |
+
with pytest.raises(ValueError):
|
| 564 |
+
est.predict_proba(X2)
|
| 565 |
+
|
| 566 |
+
# Wrong dimensions
|
| 567 |
+
est = TreeEstimator()
|
| 568 |
+
y2 = y[:-1]
|
| 569 |
+
with pytest.raises(ValueError):
|
| 570 |
+
est.fit(X, y2)
|
| 571 |
+
|
| 572 |
+
# Test with arrays that are non-contiguous.
|
| 573 |
+
Xf = np.asfortranarray(X)
|
| 574 |
+
est = TreeEstimator()
|
| 575 |
+
est.fit(Xf, y)
|
| 576 |
+
assert_almost_equal(est.predict(T), true_result)
|
| 577 |
+
|
| 578 |
+
# predict before fitting
|
| 579 |
+
est = TreeEstimator()
|
| 580 |
+
with pytest.raises(NotFittedError):
|
| 581 |
+
est.predict(T)
|
| 582 |
+
|
| 583 |
+
# predict on vector with different dims
|
| 584 |
+
est.fit(X, y)
|
| 585 |
+
t = np.asarray(T)
|
| 586 |
+
with pytest.raises(ValueError):
|
| 587 |
+
est.predict(t[:, 1:])
|
| 588 |
+
|
| 589 |
+
# wrong sample shape
|
| 590 |
+
Xt = np.array(X).T
|
| 591 |
+
|
| 592 |
+
est = TreeEstimator()
|
| 593 |
+
est.fit(np.dot(X, Xt), y)
|
| 594 |
+
with pytest.raises(ValueError):
|
| 595 |
+
est.predict(X)
|
| 596 |
+
with pytest.raises(ValueError):
|
| 597 |
+
est.apply(X)
|
| 598 |
+
|
| 599 |
+
clf = TreeEstimator()
|
| 600 |
+
clf.fit(X, y)
|
| 601 |
+
with pytest.raises(ValueError):
|
| 602 |
+
clf.predict(Xt)
|
| 603 |
+
with pytest.raises(ValueError):
|
| 604 |
+
clf.apply(Xt)
|
| 605 |
+
|
| 606 |
+
# apply before fitting
|
| 607 |
+
est = TreeEstimator()
|
| 608 |
+
with pytest.raises(NotFittedError):
|
| 609 |
+
est.apply(T)
|
| 610 |
+
|
| 611 |
+
# non positive target for Poisson splitting Criterion
|
| 612 |
+
est = DecisionTreeRegressor(criterion="poisson")
|
| 613 |
+
with pytest.raises(ValueError, match="y is not positive.*Poisson"):
|
| 614 |
+
est.fit([[0, 1, 2]], [0, 0, 0])
|
| 615 |
+
with pytest.raises(ValueError, match="Some.*y are negative.*Poisson"):
|
| 616 |
+
est.fit([[0, 1, 2]], [5, -0.1, 2])
|
| 617 |
+
|
| 618 |
+
|
| 619 |
+
def test_min_samples_split():
|
| 620 |
+
"""Test min_samples_split parameter"""
|
| 621 |
+
X = np.asfortranarray(iris.data, dtype=tree._tree.DTYPE)
|
| 622 |
+
y = iris.target
|
| 623 |
+
|
| 624 |
+
# test both DepthFirstTreeBuilder and BestFirstTreeBuilder
|
| 625 |
+
# by setting max_leaf_nodes
|
| 626 |
+
for max_leaf_nodes, name in product((None, 1000), ALL_TREES.keys()):
|
| 627 |
+
TreeEstimator = ALL_TREES[name]
|
| 628 |
+
|
| 629 |
+
# test for integer parameter
|
| 630 |
+
est = TreeEstimator(
|
| 631 |
+
min_samples_split=10, max_leaf_nodes=max_leaf_nodes, random_state=0
|
| 632 |
+
)
|
| 633 |
+
est.fit(X, y)
|
| 634 |
+
# count samples on nodes, -1 means it is a leaf
|
| 635 |
+
node_samples = est.tree_.n_node_samples[est.tree_.children_left != -1]
|
| 636 |
+
|
| 637 |
+
assert np.min(node_samples) > 9, "Failed with {0}".format(name)
|
| 638 |
+
|
| 639 |
+
# test for float parameter
|
| 640 |
+
est = TreeEstimator(
|
| 641 |
+
min_samples_split=0.2, max_leaf_nodes=max_leaf_nodes, random_state=0
|
| 642 |
+
)
|
| 643 |
+
est.fit(X, y)
|
| 644 |
+
# count samples on nodes, -1 means it is a leaf
|
| 645 |
+
node_samples = est.tree_.n_node_samples[est.tree_.children_left != -1]
|
| 646 |
+
|
| 647 |
+
assert np.min(node_samples) > 9, "Failed with {0}".format(name)
|
| 648 |
+
|
| 649 |
+
|
| 650 |
+
def test_min_samples_leaf():
|
| 651 |
+
# Test if leaves contain more than leaf_count training examples
|
| 652 |
+
X = np.asfortranarray(iris.data, dtype=tree._tree.DTYPE)
|
| 653 |
+
y = iris.target
|
| 654 |
+
|
| 655 |
+
# test both DepthFirstTreeBuilder and BestFirstTreeBuilder
|
| 656 |
+
# by setting max_leaf_nodes
|
| 657 |
+
for max_leaf_nodes, name in product((None, 1000), ALL_TREES.keys()):
|
| 658 |
+
TreeEstimator = ALL_TREES[name]
|
| 659 |
+
|
| 660 |
+
# test integer parameter
|
| 661 |
+
est = TreeEstimator(
|
| 662 |
+
min_samples_leaf=5, max_leaf_nodes=max_leaf_nodes, random_state=0
|
| 663 |
+
)
|
| 664 |
+
est.fit(X, y)
|
| 665 |
+
out = est.tree_.apply(X)
|
| 666 |
+
node_counts = np.bincount(out)
|
| 667 |
+
# drop inner nodes
|
| 668 |
+
leaf_count = node_counts[node_counts != 0]
|
| 669 |
+
assert np.min(leaf_count) > 4, "Failed with {0}".format(name)
|
| 670 |
+
|
| 671 |
+
# test float parameter
|
| 672 |
+
est = TreeEstimator(
|
| 673 |
+
min_samples_leaf=0.1, max_leaf_nodes=max_leaf_nodes, random_state=0
|
| 674 |
+
)
|
| 675 |
+
est.fit(X, y)
|
| 676 |
+
out = est.tree_.apply(X)
|
| 677 |
+
node_counts = np.bincount(out)
|
| 678 |
+
# drop inner nodes
|
| 679 |
+
leaf_count = node_counts[node_counts != 0]
|
| 680 |
+
assert np.min(leaf_count) > 4, "Failed with {0}".format(name)
|
| 681 |
+
|
| 682 |
+
|
| 683 |
+
def check_min_weight_fraction_leaf(name, datasets, sparse=False):
|
| 684 |
+
"""Test if leaves contain at least min_weight_fraction_leaf of the
|
| 685 |
+
training set"""
|
| 686 |
+
if sparse:
|
| 687 |
+
X = DATASETS[datasets]["X_sparse"].astype(np.float32)
|
| 688 |
+
else:
|
| 689 |
+
X = DATASETS[datasets]["X"].astype(np.float32)
|
| 690 |
+
y = DATASETS[datasets]["y"]
|
| 691 |
+
|
| 692 |
+
weights = rng.rand(X.shape[0])
|
| 693 |
+
total_weight = np.sum(weights)
|
| 694 |
+
|
| 695 |
+
TreeEstimator = ALL_TREES[name]
|
| 696 |
+
|
| 697 |
+
# test both DepthFirstTreeBuilder and BestFirstTreeBuilder
|
| 698 |
+
# by setting max_leaf_nodes
|
| 699 |
+
for max_leaf_nodes, frac in product((None, 1000), np.linspace(0, 0.5, 6)):
|
| 700 |
+
est = TreeEstimator(
|
| 701 |
+
min_weight_fraction_leaf=frac, max_leaf_nodes=max_leaf_nodes, random_state=0
|
| 702 |
+
)
|
| 703 |
+
est.fit(X, y, sample_weight=weights)
|
| 704 |
+
|
| 705 |
+
if sparse:
|
| 706 |
+
out = est.tree_.apply(X.tocsr())
|
| 707 |
+
|
| 708 |
+
else:
|
| 709 |
+
out = est.tree_.apply(X)
|
| 710 |
+
|
| 711 |
+
node_weights = np.bincount(out, weights=weights)
|
| 712 |
+
# drop inner nodes
|
| 713 |
+
leaf_weights = node_weights[node_weights != 0]
|
| 714 |
+
assert (
|
| 715 |
+
np.min(leaf_weights) >= total_weight * est.min_weight_fraction_leaf
|
| 716 |
+
), "Failed with {0} min_weight_fraction_leaf={1}".format(
|
| 717 |
+
name, est.min_weight_fraction_leaf
|
| 718 |
+
)
|
| 719 |
+
|
| 720 |
+
# test case with no weights passed in
|
| 721 |
+
total_weight = X.shape[0]
|
| 722 |
+
|
| 723 |
+
for max_leaf_nodes, frac in product((None, 1000), np.linspace(0, 0.5, 6)):
|
| 724 |
+
est = TreeEstimator(
|
| 725 |
+
min_weight_fraction_leaf=frac, max_leaf_nodes=max_leaf_nodes, random_state=0
|
| 726 |
+
)
|
| 727 |
+
est.fit(X, y)
|
| 728 |
+
|
| 729 |
+
if sparse:
|
| 730 |
+
out = est.tree_.apply(X.tocsr())
|
| 731 |
+
else:
|
| 732 |
+
out = est.tree_.apply(X)
|
| 733 |
+
|
| 734 |
+
node_weights = np.bincount(out)
|
| 735 |
+
# drop inner nodes
|
| 736 |
+
leaf_weights = node_weights[node_weights != 0]
|
| 737 |
+
assert (
|
| 738 |
+
np.min(leaf_weights) >= total_weight * est.min_weight_fraction_leaf
|
| 739 |
+
), "Failed with {0} min_weight_fraction_leaf={1}".format(
|
| 740 |
+
name, est.min_weight_fraction_leaf
|
| 741 |
+
)
|
| 742 |
+
|
| 743 |
+
|
| 744 |
+
@pytest.mark.parametrize("name", ALL_TREES)
|
| 745 |
+
def test_min_weight_fraction_leaf_on_dense_input(name):
|
| 746 |
+
check_min_weight_fraction_leaf(name, "iris")
|
| 747 |
+
|
| 748 |
+
|
| 749 |
+
@pytest.mark.parametrize("name", SPARSE_TREES)
|
| 750 |
+
def test_min_weight_fraction_leaf_on_sparse_input(name):
|
| 751 |
+
check_min_weight_fraction_leaf(name, "multilabel", True)
|
| 752 |
+
|
| 753 |
+
|
| 754 |
+
def check_min_weight_fraction_leaf_with_min_samples_leaf(name, datasets, sparse=False):
|
| 755 |
+
"""Test the interaction between min_weight_fraction_leaf and
|
| 756 |
+
min_samples_leaf when sample_weights is not provided in fit."""
|
| 757 |
+
if sparse:
|
| 758 |
+
X = DATASETS[datasets]["X_sparse"].astype(np.float32)
|
| 759 |
+
else:
|
| 760 |
+
X = DATASETS[datasets]["X"].astype(np.float32)
|
| 761 |
+
y = DATASETS[datasets]["y"]
|
| 762 |
+
|
| 763 |
+
total_weight = X.shape[0]
|
| 764 |
+
TreeEstimator = ALL_TREES[name]
|
| 765 |
+
for max_leaf_nodes, frac in product((None, 1000), np.linspace(0, 0.5, 3)):
|
| 766 |
+
# test integer min_samples_leaf
|
| 767 |
+
est = TreeEstimator(
|
| 768 |
+
min_weight_fraction_leaf=frac,
|
| 769 |
+
max_leaf_nodes=max_leaf_nodes,
|
| 770 |
+
min_samples_leaf=5,
|
| 771 |
+
random_state=0,
|
| 772 |
+
)
|
| 773 |
+
est.fit(X, y)
|
| 774 |
+
|
| 775 |
+
if sparse:
|
| 776 |
+
out = est.tree_.apply(X.tocsr())
|
| 777 |
+
else:
|
| 778 |
+
out = est.tree_.apply(X)
|
| 779 |
+
|
| 780 |
+
node_weights = np.bincount(out)
|
| 781 |
+
# drop inner nodes
|
| 782 |
+
leaf_weights = node_weights[node_weights != 0]
|
| 783 |
+
assert np.min(leaf_weights) >= max(
|
| 784 |
+
(total_weight * est.min_weight_fraction_leaf), 5
|
| 785 |
+
), "Failed with {0} min_weight_fraction_leaf={1}, min_samples_leaf={2}".format(
|
| 786 |
+
name, est.min_weight_fraction_leaf, est.min_samples_leaf
|
| 787 |
+
)
|
| 788 |
+
for max_leaf_nodes, frac in product((None, 1000), np.linspace(0, 0.5, 3)):
|
| 789 |
+
# test float min_samples_leaf
|
| 790 |
+
est = TreeEstimator(
|
| 791 |
+
min_weight_fraction_leaf=frac,
|
| 792 |
+
max_leaf_nodes=max_leaf_nodes,
|
| 793 |
+
min_samples_leaf=0.1,
|
| 794 |
+
random_state=0,
|
| 795 |
+
)
|
| 796 |
+
est.fit(X, y)
|
| 797 |
+
|
| 798 |
+
if sparse:
|
| 799 |
+
out = est.tree_.apply(X.tocsr())
|
| 800 |
+
else:
|
| 801 |
+
out = est.tree_.apply(X)
|
| 802 |
+
|
| 803 |
+
node_weights = np.bincount(out)
|
| 804 |
+
# drop inner nodes
|
| 805 |
+
leaf_weights = node_weights[node_weights != 0]
|
| 806 |
+
assert np.min(leaf_weights) >= max(
|
| 807 |
+
(total_weight * est.min_weight_fraction_leaf),
|
| 808 |
+
(total_weight * est.min_samples_leaf),
|
| 809 |
+
), "Failed with {0} min_weight_fraction_leaf={1}, min_samples_leaf={2}".format(
|
| 810 |
+
name, est.min_weight_fraction_leaf, est.min_samples_leaf
|
| 811 |
+
)
|
| 812 |
+
|
| 813 |
+
|
| 814 |
+
@pytest.mark.parametrize("name", ALL_TREES)
|
| 815 |
+
def test_min_weight_fraction_leaf_with_min_samples_leaf_on_dense_input(name):
|
| 816 |
+
check_min_weight_fraction_leaf_with_min_samples_leaf(name, "iris")
|
| 817 |
+
|
| 818 |
+
|
| 819 |
+
@pytest.mark.parametrize("name", SPARSE_TREES)
|
| 820 |
+
def test_min_weight_fraction_leaf_with_min_samples_leaf_on_sparse_input(name):
|
| 821 |
+
check_min_weight_fraction_leaf_with_min_samples_leaf(name, "multilabel", True)
|
| 822 |
+
|
| 823 |
+
|
| 824 |
+
def test_min_impurity_decrease():
|
| 825 |
+
# test if min_impurity_decrease ensure that a split is made only if
|
| 826 |
+
# if the impurity decrease is at least that value
|
| 827 |
+
X, y = datasets.make_classification(n_samples=10000, random_state=42)
|
| 828 |
+
|
| 829 |
+
# test both DepthFirstTreeBuilder and BestFirstTreeBuilder
|
| 830 |
+
# by setting max_leaf_nodes
|
| 831 |
+
for max_leaf_nodes, name in product((None, 1000), ALL_TREES.keys()):
|
| 832 |
+
TreeEstimator = ALL_TREES[name]
|
| 833 |
+
|
| 834 |
+
# Check default value of min_impurity_decrease, 1e-7
|
| 835 |
+
est1 = TreeEstimator(max_leaf_nodes=max_leaf_nodes, random_state=0)
|
| 836 |
+
# Check with explicit value of 0.05
|
| 837 |
+
est2 = TreeEstimator(
|
| 838 |
+
max_leaf_nodes=max_leaf_nodes, min_impurity_decrease=0.05, random_state=0
|
| 839 |
+
)
|
| 840 |
+
# Check with a much lower value of 0.0001
|
| 841 |
+
est3 = TreeEstimator(
|
| 842 |
+
max_leaf_nodes=max_leaf_nodes, min_impurity_decrease=0.0001, random_state=0
|
| 843 |
+
)
|
| 844 |
+
# Check with a much lower value of 0.1
|
| 845 |
+
est4 = TreeEstimator(
|
| 846 |
+
max_leaf_nodes=max_leaf_nodes, min_impurity_decrease=0.1, random_state=0
|
| 847 |
+
)
|
| 848 |
+
|
| 849 |
+
for est, expected_decrease in (
|
| 850 |
+
(est1, 1e-7),
|
| 851 |
+
(est2, 0.05),
|
| 852 |
+
(est3, 0.0001),
|
| 853 |
+
(est4, 0.1),
|
| 854 |
+
):
|
| 855 |
+
assert (
|
| 856 |
+
est.min_impurity_decrease <= expected_decrease
|
| 857 |
+
), "Failed, min_impurity_decrease = {0} > {1}".format(
|
| 858 |
+
est.min_impurity_decrease, expected_decrease
|
| 859 |
+
)
|
| 860 |
+
est.fit(X, y)
|
| 861 |
+
for node in range(est.tree_.node_count):
|
| 862 |
+
# If current node is a not leaf node, check if the split was
|
| 863 |
+
# justified w.r.t the min_impurity_decrease
|
| 864 |
+
if est.tree_.children_left[node] != TREE_LEAF:
|
| 865 |
+
imp_parent = est.tree_.impurity[node]
|
| 866 |
+
wtd_n_node = est.tree_.weighted_n_node_samples[node]
|
| 867 |
+
|
| 868 |
+
left = est.tree_.children_left[node]
|
| 869 |
+
wtd_n_left = est.tree_.weighted_n_node_samples[left]
|
| 870 |
+
imp_left = est.tree_.impurity[left]
|
| 871 |
+
wtd_imp_left = wtd_n_left * imp_left
|
| 872 |
+
|
| 873 |
+
right = est.tree_.children_right[node]
|
| 874 |
+
wtd_n_right = est.tree_.weighted_n_node_samples[right]
|
| 875 |
+
imp_right = est.tree_.impurity[right]
|
| 876 |
+
wtd_imp_right = wtd_n_right * imp_right
|
| 877 |
+
|
| 878 |
+
wtd_avg_left_right_imp = wtd_imp_right + wtd_imp_left
|
| 879 |
+
wtd_avg_left_right_imp /= wtd_n_node
|
| 880 |
+
|
| 881 |
+
fractional_node_weight = (
|
| 882 |
+
est.tree_.weighted_n_node_samples[node] / X.shape[0]
|
| 883 |
+
)
|
| 884 |
+
|
| 885 |
+
actual_decrease = fractional_node_weight * (
|
| 886 |
+
imp_parent - wtd_avg_left_right_imp
|
| 887 |
+
)
|
| 888 |
+
|
| 889 |
+
assert (
|
| 890 |
+
actual_decrease >= expected_decrease
|
| 891 |
+
), "Failed with {0} expected min_impurity_decrease={1}".format(
|
| 892 |
+
actual_decrease, expected_decrease
|
| 893 |
+
)
|
| 894 |
+
|
| 895 |
+
|
| 896 |
+
def test_pickle():
|
| 897 |
+
"""Test pickling preserves Tree properties and performance."""
|
| 898 |
+
for name, TreeEstimator in ALL_TREES.items():
|
| 899 |
+
if "Classifier" in name:
|
| 900 |
+
X, y = iris.data, iris.target
|
| 901 |
+
else:
|
| 902 |
+
X, y = diabetes.data, diabetes.target
|
| 903 |
+
|
| 904 |
+
est = TreeEstimator(random_state=0)
|
| 905 |
+
est.fit(X, y)
|
| 906 |
+
score = est.score(X, y)
|
| 907 |
+
|
| 908 |
+
# test that all class properties are maintained
|
| 909 |
+
attributes = [
|
| 910 |
+
"max_depth",
|
| 911 |
+
"node_count",
|
| 912 |
+
"capacity",
|
| 913 |
+
"n_classes",
|
| 914 |
+
"children_left",
|
| 915 |
+
"children_right",
|
| 916 |
+
"n_leaves",
|
| 917 |
+
"feature",
|
| 918 |
+
"threshold",
|
| 919 |
+
"impurity",
|
| 920 |
+
"n_node_samples",
|
| 921 |
+
"weighted_n_node_samples",
|
| 922 |
+
"value",
|
| 923 |
+
]
|
| 924 |
+
fitted_attribute = {
|
| 925 |
+
attribute: getattr(est.tree_, attribute) for attribute in attributes
|
| 926 |
+
}
|
| 927 |
+
|
| 928 |
+
serialized_object = pickle.dumps(est)
|
| 929 |
+
est2 = pickle.loads(serialized_object)
|
| 930 |
+
assert type(est2) == est.__class__
|
| 931 |
+
|
| 932 |
+
score2 = est2.score(X, y)
|
| 933 |
+
assert (
|
| 934 |
+
score == score2
|
| 935 |
+
), "Failed to generate same score after pickling with {0}".format(name)
|
| 936 |
+
for attribute in fitted_attribute:
|
| 937 |
+
assert_array_equal(
|
| 938 |
+
getattr(est2.tree_, attribute),
|
| 939 |
+
fitted_attribute[attribute],
|
| 940 |
+
err_msg=(
|
| 941 |
+
f"Failed to generate same attribute {attribute} after pickling with"
|
| 942 |
+
f" {name}"
|
| 943 |
+
),
|
| 944 |
+
)
|
| 945 |
+
|
| 946 |
+
|
| 947 |
+
def test_multioutput():
|
| 948 |
+
# Check estimators on multi-output problems.
|
| 949 |
+
X = [
|
| 950 |
+
[-2, -1],
|
| 951 |
+
[-1, -1],
|
| 952 |
+
[-1, -2],
|
| 953 |
+
[1, 1],
|
| 954 |
+
[1, 2],
|
| 955 |
+
[2, 1],
|
| 956 |
+
[-2, 1],
|
| 957 |
+
[-1, 1],
|
| 958 |
+
[-1, 2],
|
| 959 |
+
[2, -1],
|
| 960 |
+
[1, -1],
|
| 961 |
+
[1, -2],
|
| 962 |
+
]
|
| 963 |
+
|
| 964 |
+
y = [
|
| 965 |
+
[-1, 0],
|
| 966 |
+
[-1, 0],
|
| 967 |
+
[-1, 0],
|
| 968 |
+
[1, 1],
|
| 969 |
+
[1, 1],
|
| 970 |
+
[1, 1],
|
| 971 |
+
[-1, 2],
|
| 972 |
+
[-1, 2],
|
| 973 |
+
[-1, 2],
|
| 974 |
+
[1, 3],
|
| 975 |
+
[1, 3],
|
| 976 |
+
[1, 3],
|
| 977 |
+
]
|
| 978 |
+
|
| 979 |
+
T = [[-1, -1], [1, 1], [-1, 1], [1, -1]]
|
| 980 |
+
y_true = [[-1, 0], [1, 1], [-1, 2], [1, 3]]
|
| 981 |
+
|
| 982 |
+
# toy classification problem
|
| 983 |
+
for name, TreeClassifier in CLF_TREES.items():
|
| 984 |
+
clf = TreeClassifier(random_state=0)
|
| 985 |
+
y_hat = clf.fit(X, y).predict(T)
|
| 986 |
+
assert_array_equal(y_hat, y_true)
|
| 987 |
+
assert y_hat.shape == (4, 2)
|
| 988 |
+
|
| 989 |
+
proba = clf.predict_proba(T)
|
| 990 |
+
assert len(proba) == 2
|
| 991 |
+
assert proba[0].shape == (4, 2)
|
| 992 |
+
assert proba[1].shape == (4, 4)
|
| 993 |
+
|
| 994 |
+
log_proba = clf.predict_log_proba(T)
|
| 995 |
+
assert len(log_proba) == 2
|
| 996 |
+
assert log_proba[0].shape == (4, 2)
|
| 997 |
+
assert log_proba[1].shape == (4, 4)
|
| 998 |
+
|
| 999 |
+
# toy regression problem
|
| 1000 |
+
for name, TreeRegressor in REG_TREES.items():
|
| 1001 |
+
reg = TreeRegressor(random_state=0)
|
| 1002 |
+
y_hat = reg.fit(X, y).predict(T)
|
| 1003 |
+
assert_almost_equal(y_hat, y_true)
|
| 1004 |
+
assert y_hat.shape == (4, 2)
|
| 1005 |
+
|
| 1006 |
+
|
| 1007 |
+
def test_classes_shape():
|
| 1008 |
+
# Test that n_classes_ and classes_ have proper shape.
|
| 1009 |
+
for name, TreeClassifier in CLF_TREES.items():
|
| 1010 |
+
# Classification, single output
|
| 1011 |
+
clf = TreeClassifier(random_state=0)
|
| 1012 |
+
clf.fit(X, y)
|
| 1013 |
+
|
| 1014 |
+
assert clf.n_classes_ == 2
|
| 1015 |
+
assert_array_equal(clf.classes_, [-1, 1])
|
| 1016 |
+
|
| 1017 |
+
# Classification, multi-output
|
| 1018 |
+
_y = np.vstack((y, np.array(y) * 2)).T
|
| 1019 |
+
clf = TreeClassifier(random_state=0)
|
| 1020 |
+
clf.fit(X, _y)
|
| 1021 |
+
assert len(clf.n_classes_) == 2
|
| 1022 |
+
assert len(clf.classes_) == 2
|
| 1023 |
+
assert_array_equal(clf.n_classes_, [2, 2])
|
| 1024 |
+
assert_array_equal(clf.classes_, [[-1, 1], [-2, 2]])
|
| 1025 |
+
|
| 1026 |
+
|
| 1027 |
+
def test_unbalanced_iris():
|
| 1028 |
+
# Check class rebalancing.
|
| 1029 |
+
unbalanced_X = iris.data[:125]
|
| 1030 |
+
unbalanced_y = iris.target[:125]
|
| 1031 |
+
sample_weight = compute_sample_weight("balanced", unbalanced_y)
|
| 1032 |
+
|
| 1033 |
+
for name, TreeClassifier in CLF_TREES.items():
|
| 1034 |
+
clf = TreeClassifier(random_state=0)
|
| 1035 |
+
clf.fit(unbalanced_X, unbalanced_y, sample_weight=sample_weight)
|
| 1036 |
+
assert_almost_equal(clf.predict(unbalanced_X), unbalanced_y)
|
| 1037 |
+
|
| 1038 |
+
|
| 1039 |
+
def test_memory_layout():
|
| 1040 |
+
# Check that it works no matter the memory layout
|
| 1041 |
+
for (name, TreeEstimator), dtype in product(
|
| 1042 |
+
ALL_TREES.items(), [np.float64, np.float32]
|
| 1043 |
+
):
|
| 1044 |
+
est = TreeEstimator(random_state=0)
|
| 1045 |
+
|
| 1046 |
+
# Nothing
|
| 1047 |
+
X = np.asarray(iris.data, dtype=dtype)
|
| 1048 |
+
y = iris.target
|
| 1049 |
+
assert_array_equal(est.fit(X, y).predict(X), y)
|
| 1050 |
+
|
| 1051 |
+
# C-order
|
| 1052 |
+
X = np.asarray(iris.data, order="C", dtype=dtype)
|
| 1053 |
+
y = iris.target
|
| 1054 |
+
assert_array_equal(est.fit(X, y).predict(X), y)
|
| 1055 |
+
|
| 1056 |
+
# F-order
|
| 1057 |
+
X = np.asarray(iris.data, order="F", dtype=dtype)
|
| 1058 |
+
y = iris.target
|
| 1059 |
+
assert_array_equal(est.fit(X, y).predict(X), y)
|
| 1060 |
+
|
| 1061 |
+
# Contiguous
|
| 1062 |
+
X = np.ascontiguousarray(iris.data, dtype=dtype)
|
| 1063 |
+
y = iris.target
|
| 1064 |
+
assert_array_equal(est.fit(X, y).predict(X), y)
|
| 1065 |
+
|
| 1066 |
+
# csr matrix
|
| 1067 |
+
X = csr_matrix(iris.data, dtype=dtype)
|
| 1068 |
+
y = iris.target
|
| 1069 |
+
assert_array_equal(est.fit(X, y).predict(X), y)
|
| 1070 |
+
|
| 1071 |
+
# csc_matrix
|
| 1072 |
+
X = csc_matrix(iris.data, dtype=dtype)
|
| 1073 |
+
y = iris.target
|
| 1074 |
+
assert_array_equal(est.fit(X, y).predict(X), y)
|
| 1075 |
+
|
| 1076 |
+
# Strided
|
| 1077 |
+
X = np.asarray(iris.data[::3], dtype=dtype)
|
| 1078 |
+
y = iris.target[::3]
|
| 1079 |
+
assert_array_equal(est.fit(X, y).predict(X), y)
|
| 1080 |
+
|
| 1081 |
+
|
| 1082 |
+
def test_sample_weight():
|
| 1083 |
+
# Check sample weighting.
|
| 1084 |
+
# Test that zero-weighted samples are not taken into account
|
| 1085 |
+
X = np.arange(100)[:, np.newaxis]
|
| 1086 |
+
y = np.ones(100)
|
| 1087 |
+
y[:50] = 0.0
|
| 1088 |
+
|
| 1089 |
+
sample_weight = np.ones(100)
|
| 1090 |
+
sample_weight[y == 0] = 0.0
|
| 1091 |
+
|
| 1092 |
+
clf = DecisionTreeClassifier(random_state=0)
|
| 1093 |
+
clf.fit(X, y, sample_weight=sample_weight)
|
| 1094 |
+
assert_array_equal(clf.predict(X), np.ones(100))
|
| 1095 |
+
|
| 1096 |
+
# Test that low weighted samples are not taken into account at low depth
|
| 1097 |
+
X = np.arange(200)[:, np.newaxis]
|
| 1098 |
+
y = np.zeros(200)
|
| 1099 |
+
y[50:100] = 1
|
| 1100 |
+
y[100:200] = 2
|
| 1101 |
+
X[100:200, 0] = 200
|
| 1102 |
+
|
| 1103 |
+
sample_weight = np.ones(200)
|
| 1104 |
+
|
| 1105 |
+
sample_weight[y == 2] = 0.51 # Samples of class '2' are still weightier
|
| 1106 |
+
clf = DecisionTreeClassifier(max_depth=1, random_state=0)
|
| 1107 |
+
clf.fit(X, y, sample_weight=sample_weight)
|
| 1108 |
+
assert clf.tree_.threshold[0] == 149.5
|
| 1109 |
+
|
| 1110 |
+
sample_weight[y == 2] = 0.5 # Samples of class '2' are no longer weightier
|
| 1111 |
+
clf = DecisionTreeClassifier(max_depth=1, random_state=0)
|
| 1112 |
+
clf.fit(X, y, sample_weight=sample_weight)
|
| 1113 |
+
assert clf.tree_.threshold[0] == 49.5 # Threshold should have moved
|
| 1114 |
+
|
| 1115 |
+
# Test that sample weighting is the same as having duplicates
|
| 1116 |
+
X = iris.data
|
| 1117 |
+
y = iris.target
|
| 1118 |
+
|
| 1119 |
+
duplicates = rng.randint(0, X.shape[0], 100)
|
| 1120 |
+
|
| 1121 |
+
clf = DecisionTreeClassifier(random_state=1)
|
| 1122 |
+
clf.fit(X[duplicates], y[duplicates])
|
| 1123 |
+
|
| 1124 |
+
sample_weight = np.bincount(duplicates, minlength=X.shape[0])
|
| 1125 |
+
clf2 = DecisionTreeClassifier(random_state=1)
|
| 1126 |
+
clf2.fit(X, y, sample_weight=sample_weight)
|
| 1127 |
+
|
| 1128 |
+
internal = clf.tree_.children_left != tree._tree.TREE_LEAF
|
| 1129 |
+
assert_array_almost_equal(
|
| 1130 |
+
clf.tree_.threshold[internal], clf2.tree_.threshold[internal]
|
| 1131 |
+
)
|
| 1132 |
+
|
| 1133 |
+
|
| 1134 |
+
def test_sample_weight_invalid():
|
| 1135 |
+
# Check sample weighting raises errors.
|
| 1136 |
+
X = np.arange(100)[:, np.newaxis]
|
| 1137 |
+
y = np.ones(100)
|
| 1138 |
+
y[:50] = 0.0
|
| 1139 |
+
|
| 1140 |
+
clf = DecisionTreeClassifier(random_state=0)
|
| 1141 |
+
|
| 1142 |
+
sample_weight = np.random.rand(100, 1)
|
| 1143 |
+
with pytest.raises(ValueError):
|
| 1144 |
+
clf.fit(X, y, sample_weight=sample_weight)
|
| 1145 |
+
|
| 1146 |
+
sample_weight = np.array(0)
|
| 1147 |
+
expected_err = r"Singleton.* cannot be considered a valid collection"
|
| 1148 |
+
with pytest.raises(TypeError, match=expected_err):
|
| 1149 |
+
clf.fit(X, y, sample_weight=sample_weight)
|
| 1150 |
+
|
| 1151 |
+
|
| 1152 |
+
def check_class_weights(name):
|
| 1153 |
+
"""Check class_weights resemble sample_weights behavior."""
|
| 1154 |
+
TreeClassifier = CLF_TREES[name]
|
| 1155 |
+
|
| 1156 |
+
# Iris is balanced, so no effect expected for using 'balanced' weights
|
| 1157 |
+
clf1 = TreeClassifier(random_state=0)
|
| 1158 |
+
clf1.fit(iris.data, iris.target)
|
| 1159 |
+
clf2 = TreeClassifier(class_weight="balanced", random_state=0)
|
| 1160 |
+
clf2.fit(iris.data, iris.target)
|
| 1161 |
+
assert_almost_equal(clf1.feature_importances_, clf2.feature_importances_)
|
| 1162 |
+
|
| 1163 |
+
# Make a multi-output problem with three copies of Iris
|
| 1164 |
+
iris_multi = np.vstack((iris.target, iris.target, iris.target)).T
|
| 1165 |
+
# Create user-defined weights that should balance over the outputs
|
| 1166 |
+
clf3 = TreeClassifier(
|
| 1167 |
+
class_weight=[
|
| 1168 |
+
{0: 2.0, 1: 2.0, 2: 1.0},
|
| 1169 |
+
{0: 2.0, 1: 1.0, 2: 2.0},
|
| 1170 |
+
{0: 1.0, 1: 2.0, 2: 2.0},
|
| 1171 |
+
],
|
| 1172 |
+
random_state=0,
|
| 1173 |
+
)
|
| 1174 |
+
clf3.fit(iris.data, iris_multi)
|
| 1175 |
+
assert_almost_equal(clf2.feature_importances_, clf3.feature_importances_)
|
| 1176 |
+
# Check against multi-output "auto" which should also have no effect
|
| 1177 |
+
clf4 = TreeClassifier(class_weight="balanced", random_state=0)
|
| 1178 |
+
clf4.fit(iris.data, iris_multi)
|
| 1179 |
+
assert_almost_equal(clf3.feature_importances_, clf4.feature_importances_)
|
| 1180 |
+
|
| 1181 |
+
# Inflate importance of class 1, check against user-defined weights
|
| 1182 |
+
sample_weight = np.ones(iris.target.shape)
|
| 1183 |
+
sample_weight[iris.target == 1] *= 100
|
| 1184 |
+
class_weight = {0: 1.0, 1: 100.0, 2: 1.0}
|
| 1185 |
+
clf1 = TreeClassifier(random_state=0)
|
| 1186 |
+
clf1.fit(iris.data, iris.target, sample_weight)
|
| 1187 |
+
clf2 = TreeClassifier(class_weight=class_weight, random_state=0)
|
| 1188 |
+
clf2.fit(iris.data, iris.target)
|
| 1189 |
+
assert_almost_equal(clf1.feature_importances_, clf2.feature_importances_)
|
| 1190 |
+
|
| 1191 |
+
# Check that sample_weight and class_weight are multiplicative
|
| 1192 |
+
clf1 = TreeClassifier(random_state=0)
|
| 1193 |
+
clf1.fit(iris.data, iris.target, sample_weight**2)
|
| 1194 |
+
clf2 = TreeClassifier(class_weight=class_weight, random_state=0)
|
| 1195 |
+
clf2.fit(iris.data, iris.target, sample_weight)
|
| 1196 |
+
assert_almost_equal(clf1.feature_importances_, clf2.feature_importances_)
|
| 1197 |
+
|
| 1198 |
+
|
| 1199 |
+
@pytest.mark.parametrize("name", CLF_TREES)
|
| 1200 |
+
def test_class_weights(name):
|
| 1201 |
+
check_class_weights(name)
|
| 1202 |
+
|
| 1203 |
+
|
| 1204 |
+
def check_class_weight_errors(name):
|
| 1205 |
+
# Test if class_weight raises errors and warnings when expected.
|
| 1206 |
+
TreeClassifier = CLF_TREES[name]
|
| 1207 |
+
_y = np.vstack((y, np.array(y) * 2)).T
|
| 1208 |
+
|
| 1209 |
+
# Incorrect length list for multi-output
|
| 1210 |
+
clf = TreeClassifier(class_weight=[{-1: 0.5, 1: 1.0}], random_state=0)
|
| 1211 |
+
err_msg = "number of elements in class_weight should match number of outputs."
|
| 1212 |
+
with pytest.raises(ValueError, match=err_msg):
|
| 1213 |
+
clf.fit(X, _y)
|
| 1214 |
+
|
| 1215 |
+
|
| 1216 |
+
@pytest.mark.parametrize("name", CLF_TREES)
|
| 1217 |
+
def test_class_weight_errors(name):
|
| 1218 |
+
check_class_weight_errors(name)
|
| 1219 |
+
|
| 1220 |
+
|
| 1221 |
+
def test_max_leaf_nodes():
|
| 1222 |
+
# Test greedy trees with max_depth + 1 leafs.
|
| 1223 |
+
X, y = datasets.make_hastie_10_2(n_samples=100, random_state=1)
|
| 1224 |
+
k = 4
|
| 1225 |
+
for name, TreeEstimator in ALL_TREES.items():
|
| 1226 |
+
est = TreeEstimator(max_depth=None, max_leaf_nodes=k + 1).fit(X, y)
|
| 1227 |
+
assert est.get_n_leaves() == k + 1
|
| 1228 |
+
|
| 1229 |
+
|
| 1230 |
+
def test_max_leaf_nodes_max_depth():
|
| 1231 |
+
# Test precedence of max_leaf_nodes over max_depth.
|
| 1232 |
+
X, y = datasets.make_hastie_10_2(n_samples=100, random_state=1)
|
| 1233 |
+
k = 4
|
| 1234 |
+
for name, TreeEstimator in ALL_TREES.items():
|
| 1235 |
+
est = TreeEstimator(max_depth=1, max_leaf_nodes=k).fit(X, y)
|
| 1236 |
+
assert est.get_depth() == 1
|
| 1237 |
+
|
| 1238 |
+
|
| 1239 |
+
def test_arrays_persist():
|
| 1240 |
+
# Ensure property arrays' memory stays alive when tree disappears
|
| 1241 |
+
# non-regression for #2726
|
| 1242 |
+
for attr in [
|
| 1243 |
+
"n_classes",
|
| 1244 |
+
"value",
|
| 1245 |
+
"children_left",
|
| 1246 |
+
"children_right",
|
| 1247 |
+
"threshold",
|
| 1248 |
+
"impurity",
|
| 1249 |
+
"feature",
|
| 1250 |
+
"n_node_samples",
|
| 1251 |
+
]:
|
| 1252 |
+
value = getattr(DecisionTreeClassifier().fit([[0], [1]], [0, 1]).tree_, attr)
|
| 1253 |
+
# if pointing to freed memory, contents may be arbitrary
|
| 1254 |
+
assert -3 <= value.flat[0] < 3, "Array points to arbitrary memory"
|
| 1255 |
+
|
| 1256 |
+
|
| 1257 |
+
def test_only_constant_features():
|
| 1258 |
+
random_state = check_random_state(0)
|
| 1259 |
+
X = np.zeros((10, 20))
|
| 1260 |
+
y = random_state.randint(0, 2, (10,))
|
| 1261 |
+
for name, TreeEstimator in ALL_TREES.items():
|
| 1262 |
+
est = TreeEstimator(random_state=0)
|
| 1263 |
+
est.fit(X, y)
|
| 1264 |
+
assert est.tree_.max_depth == 0
|
| 1265 |
+
|
| 1266 |
+
|
| 1267 |
+
def test_behaviour_constant_feature_after_splits():
|
| 1268 |
+
X = np.transpose(
|
| 1269 |
+
np.vstack(([[0, 0, 0, 0, 0, 1, 2, 4, 5, 6, 7]], np.zeros((4, 11))))
|
| 1270 |
+
)
|
| 1271 |
+
y = [0, 0, 0, 1, 1, 2, 2, 2, 3, 3, 3]
|
| 1272 |
+
for name, TreeEstimator in ALL_TREES.items():
|
| 1273 |
+
# do not check extra random trees
|
| 1274 |
+
if "ExtraTree" not in name:
|
| 1275 |
+
est = TreeEstimator(random_state=0, max_features=1)
|
| 1276 |
+
est.fit(X, y)
|
| 1277 |
+
assert est.tree_.max_depth == 2
|
| 1278 |
+
assert est.tree_.node_count == 5
|
| 1279 |
+
|
| 1280 |
+
|
| 1281 |
+
def test_with_only_one_non_constant_features():
|
| 1282 |
+
X = np.hstack([np.array([[1.0], [1.0], [0.0], [0.0]]), np.zeros((4, 1000))])
|
| 1283 |
+
|
| 1284 |
+
y = np.array([0.0, 1.0, 0.0, 1.0])
|
| 1285 |
+
for name, TreeEstimator in CLF_TREES.items():
|
| 1286 |
+
est = TreeEstimator(random_state=0, max_features=1)
|
| 1287 |
+
est.fit(X, y)
|
| 1288 |
+
assert est.tree_.max_depth == 1
|
| 1289 |
+
assert_array_equal(est.predict_proba(X), np.full((4, 2), 0.5))
|
| 1290 |
+
|
| 1291 |
+
for name, TreeEstimator in REG_TREES.items():
|
| 1292 |
+
est = TreeEstimator(random_state=0, max_features=1)
|
| 1293 |
+
est.fit(X, y)
|
| 1294 |
+
assert est.tree_.max_depth == 1
|
| 1295 |
+
assert_array_equal(est.predict(X), np.full((4,), 0.5))
|
| 1296 |
+
|
| 1297 |
+
|
| 1298 |
+
def test_big_input():
|
| 1299 |
+
# Test if the warning for too large inputs is appropriate.
|
| 1300 |
+
X = np.repeat(10**40.0, 4).astype(np.float64).reshape(-1, 1)
|
| 1301 |
+
clf = DecisionTreeClassifier()
|
| 1302 |
+
with pytest.raises(ValueError, match="float32"):
|
| 1303 |
+
clf.fit(X, [0, 1, 0, 1])
|
| 1304 |
+
|
| 1305 |
+
|
| 1306 |
+
def test_realloc():
|
| 1307 |
+
from sklearn.tree._utils import _realloc_test
|
| 1308 |
+
|
| 1309 |
+
with pytest.raises(MemoryError):
|
| 1310 |
+
_realloc_test()
|
| 1311 |
+
|
| 1312 |
+
|
| 1313 |
+
def test_huge_allocations():
|
| 1314 |
+
n_bits = 8 * struct.calcsize("P")
|
| 1315 |
+
|
| 1316 |
+
X = np.random.randn(10, 2)
|
| 1317 |
+
y = np.random.randint(0, 2, 10)
|
| 1318 |
+
|
| 1319 |
+
# Sanity check: we cannot request more memory than the size of the address
|
| 1320 |
+
# space. Currently raises OverflowError.
|
| 1321 |
+
huge = 2 ** (n_bits + 1)
|
| 1322 |
+
clf = DecisionTreeClassifier(splitter="best", max_leaf_nodes=huge)
|
| 1323 |
+
with pytest.raises(Exception):
|
| 1324 |
+
clf.fit(X, y)
|
| 1325 |
+
|
| 1326 |
+
# Non-regression test: MemoryError used to be dropped by Cython
|
| 1327 |
+
# because of missing "except *".
|
| 1328 |
+
huge = 2 ** (n_bits - 1) - 1
|
| 1329 |
+
clf = DecisionTreeClassifier(splitter="best", max_leaf_nodes=huge)
|
| 1330 |
+
with pytest.raises(MemoryError):
|
| 1331 |
+
clf.fit(X, y)
|
| 1332 |
+
|
| 1333 |
+
|
| 1334 |
+
def check_sparse_input(tree, dataset, max_depth=None):
|
| 1335 |
+
TreeEstimator = ALL_TREES[tree]
|
| 1336 |
+
X = DATASETS[dataset]["X"]
|
| 1337 |
+
X_sparse = DATASETS[dataset]["X_sparse"]
|
| 1338 |
+
y = DATASETS[dataset]["y"]
|
| 1339 |
+
|
| 1340 |
+
# Gain testing time
|
| 1341 |
+
if dataset in ["digits", "diabetes"]:
|
| 1342 |
+
n_samples = X.shape[0] // 5
|
| 1343 |
+
X = X[:n_samples]
|
| 1344 |
+
X_sparse = X_sparse[:n_samples]
|
| 1345 |
+
y = y[:n_samples]
|
| 1346 |
+
|
| 1347 |
+
for sparse_format in (csr_matrix, csc_matrix, coo_matrix):
|
| 1348 |
+
X_sparse = sparse_format(X_sparse)
|
| 1349 |
+
|
| 1350 |
+
# Check the default (depth first search)
|
| 1351 |
+
d = TreeEstimator(random_state=0, max_depth=max_depth).fit(X, y)
|
| 1352 |
+
s = TreeEstimator(random_state=0, max_depth=max_depth).fit(X_sparse, y)
|
| 1353 |
+
|
| 1354 |
+
assert_tree_equal(
|
| 1355 |
+
d.tree_,
|
| 1356 |
+
s.tree_,
|
| 1357 |
+
"{0} with dense and sparse format gave different trees".format(tree),
|
| 1358 |
+
)
|
| 1359 |
+
|
| 1360 |
+
y_pred = d.predict(X)
|
| 1361 |
+
if tree in CLF_TREES:
|
| 1362 |
+
y_proba = d.predict_proba(X)
|
| 1363 |
+
y_log_proba = d.predict_log_proba(X)
|
| 1364 |
+
|
| 1365 |
+
for sparse_matrix in (csr_matrix, csc_matrix, coo_matrix):
|
| 1366 |
+
X_sparse_test = sparse_matrix(X_sparse, dtype=np.float32)
|
| 1367 |
+
|
| 1368 |
+
assert_array_almost_equal(s.predict(X_sparse_test), y_pred)
|
| 1369 |
+
|
| 1370 |
+
if tree in CLF_TREES:
|
| 1371 |
+
assert_array_almost_equal(s.predict_proba(X_sparse_test), y_proba)
|
| 1372 |
+
assert_array_almost_equal(
|
| 1373 |
+
s.predict_log_proba(X_sparse_test), y_log_proba
|
| 1374 |
+
)
|
| 1375 |
+
|
| 1376 |
+
|
| 1377 |
+
@pytest.mark.parametrize("tree_type", SPARSE_TREES)
|
| 1378 |
+
@pytest.mark.parametrize(
|
| 1379 |
+
"dataset",
|
| 1380 |
+
(
|
| 1381 |
+
"clf_small",
|
| 1382 |
+
"toy",
|
| 1383 |
+
"digits",
|
| 1384 |
+
"multilabel",
|
| 1385 |
+
"sparse-pos",
|
| 1386 |
+
"sparse-neg",
|
| 1387 |
+
"sparse-mix",
|
| 1388 |
+
"zeros",
|
| 1389 |
+
),
|
| 1390 |
+
)
|
| 1391 |
+
def test_sparse_input(tree_type, dataset):
|
| 1392 |
+
max_depth = 3 if dataset == "digits" else None
|
| 1393 |
+
check_sparse_input(tree_type, dataset, max_depth)
|
| 1394 |
+
|
| 1395 |
+
|
| 1396 |
+
@pytest.mark.parametrize("tree_type", sorted(set(SPARSE_TREES).intersection(REG_TREES)))
|
| 1397 |
+
@pytest.mark.parametrize("dataset", ["diabetes", "reg_small"])
|
| 1398 |
+
def test_sparse_input_reg_trees(tree_type, dataset):
|
| 1399 |
+
# Due to numerical instability of MSE and too strict test, we limit the
|
| 1400 |
+
# maximal depth
|
| 1401 |
+
check_sparse_input(tree_type, dataset, 2)
|
| 1402 |
+
|
| 1403 |
+
|
| 1404 |
+
def check_sparse_parameters(tree, dataset):
|
| 1405 |
+
TreeEstimator = ALL_TREES[tree]
|
| 1406 |
+
X = DATASETS[dataset]["X"]
|
| 1407 |
+
X_sparse = DATASETS[dataset]["X_sparse"]
|
| 1408 |
+
y = DATASETS[dataset]["y"]
|
| 1409 |
+
|
| 1410 |
+
# Check max_features
|
| 1411 |
+
d = TreeEstimator(random_state=0, max_features=1, max_depth=2).fit(X, y)
|
| 1412 |
+
s = TreeEstimator(random_state=0, max_features=1, max_depth=2).fit(X_sparse, y)
|
| 1413 |
+
assert_tree_equal(
|
| 1414 |
+
d.tree_,
|
| 1415 |
+
s.tree_,
|
| 1416 |
+
"{0} with dense and sparse format gave different trees".format(tree),
|
| 1417 |
+
)
|
| 1418 |
+
assert_array_almost_equal(s.predict(X), d.predict(X))
|
| 1419 |
+
|
| 1420 |
+
# Check min_samples_split
|
| 1421 |
+
d = TreeEstimator(random_state=0, max_features=1, min_samples_split=10).fit(X, y)
|
| 1422 |
+
s = TreeEstimator(random_state=0, max_features=1, min_samples_split=10).fit(
|
| 1423 |
+
X_sparse, y
|
| 1424 |
+
)
|
| 1425 |
+
assert_tree_equal(
|
| 1426 |
+
d.tree_,
|
| 1427 |
+
s.tree_,
|
| 1428 |
+
"{0} with dense and sparse format gave different trees".format(tree),
|
| 1429 |
+
)
|
| 1430 |
+
assert_array_almost_equal(s.predict(X), d.predict(X))
|
| 1431 |
+
|
| 1432 |
+
# Check min_samples_leaf
|
| 1433 |
+
d = TreeEstimator(random_state=0, min_samples_leaf=X_sparse.shape[0] // 2).fit(X, y)
|
| 1434 |
+
s = TreeEstimator(random_state=0, min_samples_leaf=X_sparse.shape[0] // 2).fit(
|
| 1435 |
+
X_sparse, y
|
| 1436 |
+
)
|
| 1437 |
+
assert_tree_equal(
|
| 1438 |
+
d.tree_,
|
| 1439 |
+
s.tree_,
|
| 1440 |
+
"{0} with dense and sparse format gave different trees".format(tree),
|
| 1441 |
+
)
|
| 1442 |
+
assert_array_almost_equal(s.predict(X), d.predict(X))
|
| 1443 |
+
|
| 1444 |
+
# Check best-first search
|
| 1445 |
+
d = TreeEstimator(random_state=0, max_leaf_nodes=3).fit(X, y)
|
| 1446 |
+
s = TreeEstimator(random_state=0, max_leaf_nodes=3).fit(X_sparse, y)
|
| 1447 |
+
assert_tree_equal(
|
| 1448 |
+
d.tree_,
|
| 1449 |
+
s.tree_,
|
| 1450 |
+
"{0} with dense and sparse format gave different trees".format(tree),
|
| 1451 |
+
)
|
| 1452 |
+
assert_array_almost_equal(s.predict(X), d.predict(X))
|
| 1453 |
+
|
| 1454 |
+
|
| 1455 |
+
def check_sparse_criterion(tree, dataset):
|
| 1456 |
+
TreeEstimator = ALL_TREES[tree]
|
| 1457 |
+
X = DATASETS[dataset]["X"]
|
| 1458 |
+
X_sparse = DATASETS[dataset]["X_sparse"]
|
| 1459 |
+
y = DATASETS[dataset]["y"]
|
| 1460 |
+
|
| 1461 |
+
# Check various criterion
|
| 1462 |
+
CRITERIONS = REG_CRITERIONS if tree in REG_TREES else CLF_CRITERIONS
|
| 1463 |
+
for criterion in CRITERIONS:
|
| 1464 |
+
d = TreeEstimator(random_state=0, max_depth=3, criterion=criterion).fit(X, y)
|
| 1465 |
+
s = TreeEstimator(random_state=0, max_depth=3, criterion=criterion).fit(
|
| 1466 |
+
X_sparse, y
|
| 1467 |
+
)
|
| 1468 |
+
|
| 1469 |
+
assert_tree_equal(
|
| 1470 |
+
d.tree_,
|
| 1471 |
+
s.tree_,
|
| 1472 |
+
"{0} with dense and sparse format gave different trees".format(tree),
|
| 1473 |
+
)
|
| 1474 |
+
assert_array_almost_equal(s.predict(X), d.predict(X))
|
| 1475 |
+
|
| 1476 |
+
|
| 1477 |
+
@pytest.mark.parametrize("tree_type", SPARSE_TREES)
|
| 1478 |
+
@pytest.mark.parametrize("dataset", ["sparse-pos", "sparse-neg", "sparse-mix", "zeros"])
|
| 1479 |
+
@pytest.mark.parametrize("check", [check_sparse_parameters, check_sparse_criterion])
|
| 1480 |
+
def test_sparse(tree_type, dataset, check):
|
| 1481 |
+
check(tree_type, dataset)
|
| 1482 |
+
|
| 1483 |
+
|
| 1484 |
+
def check_explicit_sparse_zeros(tree, max_depth=3, n_features=10):
|
| 1485 |
+
TreeEstimator = ALL_TREES[tree]
|
| 1486 |
+
|
| 1487 |
+
# n_samples set n_feature to ease construction of a simultaneous
|
| 1488 |
+
# construction of a csr and csc matrix
|
| 1489 |
+
n_samples = n_features
|
| 1490 |
+
samples = np.arange(n_samples)
|
| 1491 |
+
|
| 1492 |
+
# Generate X, y
|
| 1493 |
+
random_state = check_random_state(0)
|
| 1494 |
+
indices = []
|
| 1495 |
+
data = []
|
| 1496 |
+
offset = 0
|
| 1497 |
+
indptr = [offset]
|
| 1498 |
+
for i in range(n_features):
|
| 1499 |
+
n_nonzero_i = random_state.binomial(n_samples, 0.5)
|
| 1500 |
+
indices_i = random_state.permutation(samples)[:n_nonzero_i]
|
| 1501 |
+
indices.append(indices_i)
|
| 1502 |
+
data_i = random_state.binomial(3, 0.5, size=(n_nonzero_i,)) - 1
|
| 1503 |
+
data.append(data_i)
|
| 1504 |
+
offset += n_nonzero_i
|
| 1505 |
+
indptr.append(offset)
|
| 1506 |
+
|
| 1507 |
+
indices = np.concatenate(indices)
|
| 1508 |
+
data = np.array(np.concatenate(data), dtype=np.float32)
|
| 1509 |
+
X_sparse = csc_matrix((data, indices, indptr), shape=(n_samples, n_features))
|
| 1510 |
+
X = X_sparse.toarray()
|
| 1511 |
+
X_sparse_test = csr_matrix((data, indices, indptr), shape=(n_samples, n_features))
|
| 1512 |
+
X_test = X_sparse_test.toarray()
|
| 1513 |
+
y = random_state.randint(0, 3, size=(n_samples,))
|
| 1514 |
+
|
| 1515 |
+
# Ensure that X_sparse_test owns its data, indices and indptr array
|
| 1516 |
+
X_sparse_test = X_sparse_test.copy()
|
| 1517 |
+
|
| 1518 |
+
# Ensure that we have explicit zeros
|
| 1519 |
+
assert (X_sparse.data == 0.0).sum() > 0
|
| 1520 |
+
assert (X_sparse_test.data == 0.0).sum() > 0
|
| 1521 |
+
|
| 1522 |
+
# Perform the comparison
|
| 1523 |
+
d = TreeEstimator(random_state=0, max_depth=max_depth).fit(X, y)
|
| 1524 |
+
s = TreeEstimator(random_state=0, max_depth=max_depth).fit(X_sparse, y)
|
| 1525 |
+
|
| 1526 |
+
assert_tree_equal(
|
| 1527 |
+
d.tree_,
|
| 1528 |
+
s.tree_,
|
| 1529 |
+
"{0} with dense and sparse format gave different trees".format(tree),
|
| 1530 |
+
)
|
| 1531 |
+
|
| 1532 |
+
Xs = (X_test, X_sparse_test)
|
| 1533 |
+
for X1, X2 in product(Xs, Xs):
|
| 1534 |
+
assert_array_almost_equal(s.tree_.apply(X1), d.tree_.apply(X2))
|
| 1535 |
+
assert_array_almost_equal(s.apply(X1), d.apply(X2))
|
| 1536 |
+
assert_array_almost_equal(s.apply(X1), s.tree_.apply(X1))
|
| 1537 |
+
|
| 1538 |
+
assert_array_almost_equal(
|
| 1539 |
+
s.tree_.decision_path(X1).toarray(), d.tree_.decision_path(X2).toarray()
|
| 1540 |
+
)
|
| 1541 |
+
assert_array_almost_equal(
|
| 1542 |
+
s.decision_path(X1).toarray(), d.decision_path(X2).toarray()
|
| 1543 |
+
)
|
| 1544 |
+
assert_array_almost_equal(
|
| 1545 |
+
s.decision_path(X1).toarray(), s.tree_.decision_path(X1).toarray()
|
| 1546 |
+
)
|
| 1547 |
+
|
| 1548 |
+
assert_array_almost_equal(s.predict(X1), d.predict(X2))
|
| 1549 |
+
|
| 1550 |
+
if tree in CLF_TREES:
|
| 1551 |
+
assert_array_almost_equal(s.predict_proba(X1), d.predict_proba(X2))
|
| 1552 |
+
|
| 1553 |
+
|
| 1554 |
+
@pytest.mark.parametrize("tree_type", SPARSE_TREES)
|
| 1555 |
+
def test_explicit_sparse_zeros(tree_type):
|
| 1556 |
+
check_explicit_sparse_zeros(tree_type)
|
| 1557 |
+
|
| 1558 |
+
|
| 1559 |
+
@ignore_warnings
|
| 1560 |
+
def check_raise_error_on_1d_input(name):
|
| 1561 |
+
TreeEstimator = ALL_TREES[name]
|
| 1562 |
+
|
| 1563 |
+
X = iris.data[:, 0].ravel()
|
| 1564 |
+
X_2d = iris.data[:, 0].reshape((-1, 1))
|
| 1565 |
+
y = iris.target
|
| 1566 |
+
|
| 1567 |
+
with pytest.raises(ValueError):
|
| 1568 |
+
TreeEstimator(random_state=0).fit(X, y)
|
| 1569 |
+
|
| 1570 |
+
est = TreeEstimator(random_state=0)
|
| 1571 |
+
est.fit(X_2d, y)
|
| 1572 |
+
with pytest.raises(ValueError):
|
| 1573 |
+
est.predict([X])
|
| 1574 |
+
|
| 1575 |
+
|
| 1576 |
+
@pytest.mark.parametrize("name", ALL_TREES)
|
| 1577 |
+
def test_1d_input(name):
|
| 1578 |
+
with ignore_warnings():
|
| 1579 |
+
check_raise_error_on_1d_input(name)
|
| 1580 |
+
|
| 1581 |
+
|
| 1582 |
+
def _check_min_weight_leaf_split_level(TreeEstimator, X, y, sample_weight):
|
| 1583 |
+
est = TreeEstimator(random_state=0)
|
| 1584 |
+
est.fit(X, y, sample_weight=sample_weight)
|
| 1585 |
+
assert est.tree_.max_depth == 1
|
| 1586 |
+
|
| 1587 |
+
est = TreeEstimator(random_state=0, min_weight_fraction_leaf=0.4)
|
| 1588 |
+
est.fit(X, y, sample_weight=sample_weight)
|
| 1589 |
+
assert est.tree_.max_depth == 0
|
| 1590 |
+
|
| 1591 |
+
|
| 1592 |
+
def check_min_weight_leaf_split_level(name):
|
| 1593 |
+
TreeEstimator = ALL_TREES[name]
|
| 1594 |
+
|
| 1595 |
+
X = np.array([[0], [0], [0], [0], [1]])
|
| 1596 |
+
y = [0, 0, 0, 0, 1]
|
| 1597 |
+
sample_weight = [0.2, 0.2, 0.2, 0.2, 0.2]
|
| 1598 |
+
_check_min_weight_leaf_split_level(TreeEstimator, X, y, sample_weight)
|
| 1599 |
+
|
| 1600 |
+
_check_min_weight_leaf_split_level(TreeEstimator, csc_matrix(X), y, sample_weight)
|
| 1601 |
+
|
| 1602 |
+
|
| 1603 |
+
@pytest.mark.parametrize("name", ALL_TREES)
|
| 1604 |
+
def test_min_weight_leaf_split_level(name):
|
| 1605 |
+
check_min_weight_leaf_split_level(name)
|
| 1606 |
+
|
| 1607 |
+
|
| 1608 |
+
def check_public_apply(name):
|
| 1609 |
+
X_small32 = X_small.astype(tree._tree.DTYPE, copy=False)
|
| 1610 |
+
|
| 1611 |
+
est = ALL_TREES[name]()
|
| 1612 |
+
est.fit(X_small, y_small)
|
| 1613 |
+
assert_array_equal(est.apply(X_small), est.tree_.apply(X_small32))
|
| 1614 |
+
|
| 1615 |
+
|
| 1616 |
+
def check_public_apply_sparse(name):
|
| 1617 |
+
X_small32 = csr_matrix(X_small.astype(tree._tree.DTYPE, copy=False))
|
| 1618 |
+
|
| 1619 |
+
est = ALL_TREES[name]()
|
| 1620 |
+
est.fit(X_small, y_small)
|
| 1621 |
+
assert_array_equal(est.apply(X_small), est.tree_.apply(X_small32))
|
| 1622 |
+
|
| 1623 |
+
|
| 1624 |
+
@pytest.mark.parametrize("name", ALL_TREES)
|
| 1625 |
+
def test_public_apply_all_trees(name):
|
| 1626 |
+
check_public_apply(name)
|
| 1627 |
+
|
| 1628 |
+
|
| 1629 |
+
@pytest.mark.parametrize("name", SPARSE_TREES)
|
| 1630 |
+
def test_public_apply_sparse_trees(name):
|
| 1631 |
+
check_public_apply_sparse(name)
|
| 1632 |
+
|
| 1633 |
+
|
| 1634 |
+
def test_decision_path_hardcoded():
|
| 1635 |
+
X = iris.data
|
| 1636 |
+
y = iris.target
|
| 1637 |
+
est = DecisionTreeClassifier(random_state=0, max_depth=1).fit(X, y)
|
| 1638 |
+
node_indicator = est.decision_path(X[:2]).toarray()
|
| 1639 |
+
assert_array_equal(node_indicator, [[1, 1, 0], [1, 0, 1]])
|
| 1640 |
+
|
| 1641 |
+
|
| 1642 |
+
def check_decision_path(name):
|
| 1643 |
+
X = iris.data
|
| 1644 |
+
y = iris.target
|
| 1645 |
+
n_samples = X.shape[0]
|
| 1646 |
+
|
| 1647 |
+
TreeEstimator = ALL_TREES[name]
|
| 1648 |
+
est = TreeEstimator(random_state=0, max_depth=2)
|
| 1649 |
+
est.fit(X, y)
|
| 1650 |
+
|
| 1651 |
+
node_indicator_csr = est.decision_path(X)
|
| 1652 |
+
node_indicator = node_indicator_csr.toarray()
|
| 1653 |
+
assert node_indicator.shape == (n_samples, est.tree_.node_count)
|
| 1654 |
+
|
| 1655 |
+
# Assert that leaves index are correct
|
| 1656 |
+
leaves = est.apply(X)
|
| 1657 |
+
leave_indicator = [node_indicator[i, j] for i, j in enumerate(leaves)]
|
| 1658 |
+
assert_array_almost_equal(leave_indicator, np.ones(shape=n_samples))
|
| 1659 |
+
|
| 1660 |
+
# Ensure only one leave node per sample
|
| 1661 |
+
all_leaves = est.tree_.children_left == TREE_LEAF
|
| 1662 |
+
assert_array_almost_equal(
|
| 1663 |
+
np.dot(node_indicator, all_leaves), np.ones(shape=n_samples)
|
| 1664 |
+
)
|
| 1665 |
+
|
| 1666 |
+
# Ensure max depth is consistent with sum of indicator
|
| 1667 |
+
max_depth = node_indicator.sum(axis=1).max()
|
| 1668 |
+
assert est.tree_.max_depth <= max_depth
|
| 1669 |
+
|
| 1670 |
+
|
| 1671 |
+
@pytest.mark.parametrize("name", ALL_TREES)
|
| 1672 |
+
def test_decision_path(name):
|
| 1673 |
+
check_decision_path(name)
|
| 1674 |
+
|
| 1675 |
+
|
| 1676 |
+
def check_no_sparse_y_support(name):
|
| 1677 |
+
X, y = X_multilabel, csr_matrix(y_multilabel)
|
| 1678 |
+
TreeEstimator = ALL_TREES[name]
|
| 1679 |
+
with pytest.raises(TypeError):
|
| 1680 |
+
TreeEstimator(random_state=0).fit(X, y)
|
| 1681 |
+
|
| 1682 |
+
|
| 1683 |
+
@pytest.mark.parametrize("name", ALL_TREES)
|
| 1684 |
+
def test_no_sparse_y_support(name):
|
| 1685 |
+
# Currently we don't support sparse y
|
| 1686 |
+
check_no_sparse_y_support(name)
|
| 1687 |
+
|
| 1688 |
+
|
| 1689 |
+
def test_mae():
|
| 1690 |
+
"""Check MAE criterion produces correct results on small toy dataset:
|
| 1691 |
+
|
| 1692 |
+
------------------
|
| 1693 |
+
| X | y | weight |
|
| 1694 |
+
------------------
|
| 1695 |
+
| 3 | 3 | 0.1 |
|
| 1696 |
+
| 5 | 3 | 0.3 |
|
| 1697 |
+
| 8 | 4 | 1.0 |
|
| 1698 |
+
| 3 | 6 | 0.6 |
|
| 1699 |
+
| 5 | 7 | 0.3 |
|
| 1700 |
+
------------------
|
| 1701 |
+
|sum wt:| 2.3 |
|
| 1702 |
+
------------------
|
| 1703 |
+
|
| 1704 |
+
Because we are dealing with sample weights, we cannot find the median by
|
| 1705 |
+
simply choosing/averaging the centre value(s), instead we consider the
|
| 1706 |
+
median where 50% of the cumulative weight is found (in a y sorted data set)
|
| 1707 |
+
. Therefore with regards to this test data, the cumulative weight is >= 50%
|
| 1708 |
+
when y = 4. Therefore:
|
| 1709 |
+
Median = 4
|
| 1710 |
+
|
| 1711 |
+
For all the samples, we can get the total error by summing:
|
| 1712 |
+
Absolute(Median - y) * weight
|
| 1713 |
+
|
| 1714 |
+
I.e., total error = (Absolute(4 - 3) * 0.1)
|
| 1715 |
+
+ (Absolute(4 - 3) * 0.3)
|
| 1716 |
+
+ (Absolute(4 - 4) * 1.0)
|
| 1717 |
+
+ (Absolute(4 - 6) * 0.6)
|
| 1718 |
+
+ (Absolute(4 - 7) * 0.3)
|
| 1719 |
+
= 2.5
|
| 1720 |
+
|
| 1721 |
+
Impurity = Total error / total weight
|
| 1722 |
+
= 2.5 / 2.3
|
| 1723 |
+
= 1.08695652173913
|
| 1724 |
+
------------------
|
| 1725 |
+
|
| 1726 |
+
From this root node, the next best split is between X values of 3 and 5.
|
| 1727 |
+
Thus, we have left and right child nodes:
|
| 1728 |
+
|
| 1729 |
+
LEFT RIGHT
|
| 1730 |
+
------------------ ------------------
|
| 1731 |
+
| X | y | weight | | X | y | weight |
|
| 1732 |
+
------------------ ------------------
|
| 1733 |
+
| 3 | 3 | 0.1 | | 5 | 3 | 0.3 |
|
| 1734 |
+
| 3 | 6 | 0.6 | | 8 | 4 | 1.0 |
|
| 1735 |
+
------------------ | 5 | 7 | 0.3 |
|
| 1736 |
+
|sum wt:| 0.7 | ------------------
|
| 1737 |
+
------------------ |sum wt:| 1.6 |
|
| 1738 |
+
------------------
|
| 1739 |
+
|
| 1740 |
+
Impurity is found in the same way:
|
| 1741 |
+
Left node Median = 6
|
| 1742 |
+
Total error = (Absolute(6 - 3) * 0.1)
|
| 1743 |
+
+ (Absolute(6 - 6) * 0.6)
|
| 1744 |
+
= 0.3
|
| 1745 |
+
|
| 1746 |
+
Left Impurity = Total error / total weight
|
| 1747 |
+
= 0.3 / 0.7
|
| 1748 |
+
= 0.428571428571429
|
| 1749 |
+
-------------------
|
| 1750 |
+
|
| 1751 |
+
Likewise for Right node:
|
| 1752 |
+
Right node Median = 4
|
| 1753 |
+
Total error = (Absolute(4 - 3) * 0.3)
|
| 1754 |
+
+ (Absolute(4 - 4) * 1.0)
|
| 1755 |
+
+ (Absolute(4 - 7) * 0.3)
|
| 1756 |
+
= 1.2
|
| 1757 |
+
|
| 1758 |
+
Right Impurity = Total error / total weight
|
| 1759 |
+
= 1.2 / 1.6
|
| 1760 |
+
= 0.75
|
| 1761 |
+
------
|
| 1762 |
+
"""
|
| 1763 |
+
dt_mae = DecisionTreeRegressor(
|
| 1764 |
+
random_state=0, criterion="absolute_error", max_leaf_nodes=2
|
| 1765 |
+
)
|
| 1766 |
+
|
| 1767 |
+
# Test MAE where sample weights are non-uniform (as illustrated above):
|
| 1768 |
+
dt_mae.fit(
|
| 1769 |
+
X=[[3], [5], [3], [8], [5]],
|
| 1770 |
+
y=[6, 7, 3, 4, 3],
|
| 1771 |
+
sample_weight=[0.6, 0.3, 0.1, 1.0, 0.3],
|
| 1772 |
+
)
|
| 1773 |
+
assert_allclose(dt_mae.tree_.impurity, [2.5 / 2.3, 0.3 / 0.7, 1.2 / 1.6])
|
| 1774 |
+
assert_array_equal(dt_mae.tree_.value.flat, [4.0, 6.0, 4.0])
|
| 1775 |
+
|
| 1776 |
+
# Test MAE where all sample weights are uniform:
|
| 1777 |
+
dt_mae.fit(X=[[3], [5], [3], [8], [5]], y=[6, 7, 3, 4, 3], sample_weight=np.ones(5))
|
| 1778 |
+
assert_array_equal(dt_mae.tree_.impurity, [1.4, 1.5, 4.0 / 3.0])
|
| 1779 |
+
assert_array_equal(dt_mae.tree_.value.flat, [4, 4.5, 4.0])
|
| 1780 |
+
|
| 1781 |
+
# Test MAE where a `sample_weight` is not explicitly provided.
|
| 1782 |
+
# This is equivalent to providing uniform sample weights, though
|
| 1783 |
+
# the internal logic is different:
|
| 1784 |
+
dt_mae.fit(X=[[3], [5], [3], [8], [5]], y=[6, 7, 3, 4, 3])
|
| 1785 |
+
assert_array_equal(dt_mae.tree_.impurity, [1.4, 1.5, 4.0 / 3.0])
|
| 1786 |
+
assert_array_equal(dt_mae.tree_.value.flat, [4, 4.5, 4.0])
|
| 1787 |
+
|
| 1788 |
+
|
| 1789 |
+
def test_criterion_copy():
|
| 1790 |
+
# Let's check whether copy of our criterion has the same type
|
| 1791 |
+
# and properties as original
|
| 1792 |
+
n_outputs = 3
|
| 1793 |
+
n_classes = np.arange(3, dtype=np.intp)
|
| 1794 |
+
n_samples = 100
|
| 1795 |
+
|
| 1796 |
+
def _pickle_copy(obj):
|
| 1797 |
+
return pickle.loads(pickle.dumps(obj))
|
| 1798 |
+
|
| 1799 |
+
for copy_func in [copy.copy, copy.deepcopy, _pickle_copy]:
|
| 1800 |
+
for _, typename in CRITERIA_CLF.items():
|
| 1801 |
+
criteria = typename(n_outputs, n_classes)
|
| 1802 |
+
result = copy_func(criteria).__reduce__()
|
| 1803 |
+
typename_, (n_outputs_, n_classes_), _ = result
|
| 1804 |
+
assert typename == typename_
|
| 1805 |
+
assert n_outputs == n_outputs_
|
| 1806 |
+
assert_array_equal(n_classes, n_classes_)
|
| 1807 |
+
|
| 1808 |
+
for _, typename in CRITERIA_REG.items():
|
| 1809 |
+
criteria = typename(n_outputs, n_samples)
|
| 1810 |
+
result = copy_func(criteria).__reduce__()
|
| 1811 |
+
typename_, (n_outputs_, n_samples_), _ = result
|
| 1812 |
+
assert typename == typename_
|
| 1813 |
+
assert n_outputs == n_outputs_
|
| 1814 |
+
assert n_samples == n_samples_
|
| 1815 |
+
|
| 1816 |
+
|
| 1817 |
+
def test_empty_leaf_infinite_threshold():
|
| 1818 |
+
# try to make empty leaf by using near infinite value.
|
| 1819 |
+
data = np.random.RandomState(0).randn(100, 11) * 2e38
|
| 1820 |
+
data = np.nan_to_num(data.astype("float32"))
|
| 1821 |
+
X_full = data[:, :-1]
|
| 1822 |
+
X_sparse = csc_matrix(X_full)
|
| 1823 |
+
y = data[:, -1]
|
| 1824 |
+
for X in [X_full, X_sparse]:
|
| 1825 |
+
tree = DecisionTreeRegressor(random_state=0).fit(X, y)
|
| 1826 |
+
terminal_regions = tree.apply(X)
|
| 1827 |
+
left_leaf = set(np.where(tree.tree_.children_left == TREE_LEAF)[0])
|
| 1828 |
+
empty_leaf = left_leaf.difference(terminal_regions)
|
| 1829 |
+
infinite_threshold = np.where(~np.isfinite(tree.tree_.threshold))[0]
|
| 1830 |
+
assert len(infinite_threshold) == 0
|
| 1831 |
+
assert len(empty_leaf) == 0
|
| 1832 |
+
|
| 1833 |
+
|
| 1834 |
+
@pytest.mark.parametrize("criterion", CLF_CRITERIONS)
|
| 1835 |
+
@pytest.mark.parametrize(
|
| 1836 |
+
"dataset", sorted(set(DATASETS.keys()) - {"reg_small", "diabetes"})
|
| 1837 |
+
)
|
| 1838 |
+
@pytest.mark.parametrize("tree_cls", [DecisionTreeClassifier, ExtraTreeClassifier])
|
| 1839 |
+
def test_prune_tree_classifier_are_subtrees(criterion, dataset, tree_cls):
|
| 1840 |
+
dataset = DATASETS[dataset]
|
| 1841 |
+
X, y = dataset["X"], dataset["y"]
|
| 1842 |
+
est = tree_cls(max_leaf_nodes=20, random_state=0)
|
| 1843 |
+
info = est.cost_complexity_pruning_path(X, y)
|
| 1844 |
+
|
| 1845 |
+
pruning_path = info.ccp_alphas
|
| 1846 |
+
impurities = info.impurities
|
| 1847 |
+
assert np.all(np.diff(pruning_path) >= 0)
|
| 1848 |
+
assert np.all(np.diff(impurities) >= 0)
|
| 1849 |
+
|
| 1850 |
+
assert_pruning_creates_subtree(tree_cls, X, y, pruning_path)
|
| 1851 |
+
|
| 1852 |
+
|
| 1853 |
+
@pytest.mark.parametrize("criterion", REG_CRITERIONS)
|
| 1854 |
+
@pytest.mark.parametrize("dataset", DATASETS.keys())
|
| 1855 |
+
@pytest.mark.parametrize("tree_cls", [DecisionTreeRegressor, ExtraTreeRegressor])
|
| 1856 |
+
def test_prune_tree_regression_are_subtrees(criterion, dataset, tree_cls):
|
| 1857 |
+
dataset = DATASETS[dataset]
|
| 1858 |
+
X, y = dataset["X"], dataset["y"]
|
| 1859 |
+
|
| 1860 |
+
est = tree_cls(max_leaf_nodes=20, random_state=0)
|
| 1861 |
+
info = est.cost_complexity_pruning_path(X, y)
|
| 1862 |
+
|
| 1863 |
+
pruning_path = info.ccp_alphas
|
| 1864 |
+
impurities = info.impurities
|
| 1865 |
+
assert np.all(np.diff(pruning_path) >= 0)
|
| 1866 |
+
assert np.all(np.diff(impurities) >= 0)
|
| 1867 |
+
|
| 1868 |
+
assert_pruning_creates_subtree(tree_cls, X, y, pruning_path)
|
| 1869 |
+
|
| 1870 |
+
|
| 1871 |
+
def test_prune_single_node_tree():
|
| 1872 |
+
# single node tree
|
| 1873 |
+
clf1 = DecisionTreeClassifier(random_state=0)
|
| 1874 |
+
clf1.fit([[0], [1]], [0, 0])
|
| 1875 |
+
|
| 1876 |
+
# pruned single node tree
|
| 1877 |
+
clf2 = DecisionTreeClassifier(random_state=0, ccp_alpha=10)
|
| 1878 |
+
clf2.fit([[0], [1]], [0, 0])
|
| 1879 |
+
|
| 1880 |
+
assert_is_subtree(clf1.tree_, clf2.tree_)
|
| 1881 |
+
|
| 1882 |
+
|
| 1883 |
+
def assert_pruning_creates_subtree(estimator_cls, X, y, pruning_path):
|
| 1884 |
+
# generate trees with increasing alphas
|
| 1885 |
+
estimators = []
|
| 1886 |
+
for ccp_alpha in pruning_path:
|
| 1887 |
+
est = estimator_cls(max_leaf_nodes=20, ccp_alpha=ccp_alpha, random_state=0).fit(
|
| 1888 |
+
X, y
|
| 1889 |
+
)
|
| 1890 |
+
estimators.append(est)
|
| 1891 |
+
|
| 1892 |
+
# A pruned tree must be a subtree of the previous tree (which had a
|
| 1893 |
+
# smaller ccp_alpha)
|
| 1894 |
+
for prev_est, next_est in zip(estimators, estimators[1:]):
|
| 1895 |
+
assert_is_subtree(prev_est.tree_, next_est.tree_)
|
| 1896 |
+
|
| 1897 |
+
|
| 1898 |
+
def assert_is_subtree(tree, subtree):
|
| 1899 |
+
assert tree.node_count >= subtree.node_count
|
| 1900 |
+
assert tree.max_depth >= subtree.max_depth
|
| 1901 |
+
|
| 1902 |
+
tree_c_left = tree.children_left
|
| 1903 |
+
tree_c_right = tree.children_right
|
| 1904 |
+
subtree_c_left = subtree.children_left
|
| 1905 |
+
subtree_c_right = subtree.children_right
|
| 1906 |
+
|
| 1907 |
+
stack = [(0, 0)]
|
| 1908 |
+
while stack:
|
| 1909 |
+
tree_node_idx, subtree_node_idx = stack.pop()
|
| 1910 |
+
assert_array_almost_equal(
|
| 1911 |
+
tree.value[tree_node_idx], subtree.value[subtree_node_idx]
|
| 1912 |
+
)
|
| 1913 |
+
assert_almost_equal(
|
| 1914 |
+
tree.impurity[tree_node_idx], subtree.impurity[subtree_node_idx]
|
| 1915 |
+
)
|
| 1916 |
+
assert_almost_equal(
|
| 1917 |
+
tree.n_node_samples[tree_node_idx], subtree.n_node_samples[subtree_node_idx]
|
| 1918 |
+
)
|
| 1919 |
+
assert_almost_equal(
|
| 1920 |
+
tree.weighted_n_node_samples[tree_node_idx],
|
| 1921 |
+
subtree.weighted_n_node_samples[subtree_node_idx],
|
| 1922 |
+
)
|
| 1923 |
+
|
| 1924 |
+
if subtree_c_left[subtree_node_idx] == subtree_c_right[subtree_node_idx]:
|
| 1925 |
+
# is a leaf
|
| 1926 |
+
assert_almost_equal(TREE_UNDEFINED, subtree.threshold[subtree_node_idx])
|
| 1927 |
+
else:
|
| 1928 |
+
# not a leaf
|
| 1929 |
+
assert_almost_equal(
|
| 1930 |
+
tree.threshold[tree_node_idx], subtree.threshold[subtree_node_idx]
|
| 1931 |
+
)
|
| 1932 |
+
stack.append((tree_c_left[tree_node_idx], subtree_c_left[subtree_node_idx]))
|
| 1933 |
+
stack.append(
|
| 1934 |
+
(tree_c_right[tree_node_idx], subtree_c_right[subtree_node_idx])
|
| 1935 |
+
)
|
| 1936 |
+
|
| 1937 |
+
|
| 1938 |
+
@pytest.mark.parametrize("name", ALL_TREES)
|
| 1939 |
+
@pytest.mark.parametrize("splitter", ["best", "random"])
|
| 1940 |
+
@pytest.mark.parametrize("X_format", ["dense", "csr", "csc"])
|
| 1941 |
+
def test_apply_path_readonly_all_trees(name, splitter, X_format):
|
| 1942 |
+
dataset = DATASETS["clf_small"]
|
| 1943 |
+
X_small = dataset["X"].astype(tree._tree.DTYPE, copy=False)
|
| 1944 |
+
if X_format == "dense":
|
| 1945 |
+
X_readonly = create_memmap_backed_data(X_small)
|
| 1946 |
+
else:
|
| 1947 |
+
X_readonly = dataset["X_sparse"] # CSR
|
| 1948 |
+
if X_format == "csc":
|
| 1949 |
+
# Cheap CSR to CSC conversion
|
| 1950 |
+
X_readonly = X_readonly.tocsc()
|
| 1951 |
+
|
| 1952 |
+
X_readonly.data = np.array(X_readonly.data, dtype=tree._tree.DTYPE)
|
| 1953 |
+
(
|
| 1954 |
+
X_readonly.data,
|
| 1955 |
+
X_readonly.indices,
|
| 1956 |
+
X_readonly.indptr,
|
| 1957 |
+
) = create_memmap_backed_data(
|
| 1958 |
+
(X_readonly.data, X_readonly.indices, X_readonly.indptr)
|
| 1959 |
+
)
|
| 1960 |
+
|
| 1961 |
+
y_readonly = create_memmap_backed_data(np.array(y_small, dtype=tree._tree.DTYPE))
|
| 1962 |
+
est = ALL_TREES[name](splitter=splitter)
|
| 1963 |
+
est.fit(X_readonly, y_readonly)
|
| 1964 |
+
assert_array_equal(est.predict(X_readonly), est.predict(X_small))
|
| 1965 |
+
assert_array_equal(
|
| 1966 |
+
est.decision_path(X_readonly).todense(), est.decision_path(X_small).todense()
|
| 1967 |
+
)
|
| 1968 |
+
|
| 1969 |
+
|
| 1970 |
+
@pytest.mark.parametrize("criterion", ["squared_error", "friedman_mse", "poisson"])
|
| 1971 |
+
@pytest.mark.parametrize("Tree", REG_TREES.values())
|
| 1972 |
+
def test_balance_property(criterion, Tree):
|
| 1973 |
+
# Test that sum(y_pred)=sum(y_true) on training set.
|
| 1974 |
+
# This works if the mean is predicted (should even be true for each leaf).
|
| 1975 |
+
# MAE predicts the median and is therefore excluded from this test.
|
| 1976 |
+
|
| 1977 |
+
# Choose a training set with non-negative targets (for poisson)
|
| 1978 |
+
X, y = diabetes.data, diabetes.target
|
| 1979 |
+
reg = Tree(criterion=criterion)
|
| 1980 |
+
reg.fit(X, y)
|
| 1981 |
+
assert np.sum(reg.predict(X)) == pytest.approx(np.sum(y))
|
| 1982 |
+
|
| 1983 |
+
|
| 1984 |
+
@pytest.mark.parametrize("seed", range(3))
|
| 1985 |
+
def test_poisson_zero_nodes(seed):
|
| 1986 |
+
# Test that sum(y)=0 and therefore y_pred=0 is forbidden on nodes.
|
| 1987 |
+
X = [[0, 0], [0, 1], [0, 2], [0, 3], [1, 0], [1, 2], [1, 2], [1, 3]]
|
| 1988 |
+
y = [0, 0, 0, 0, 1, 2, 3, 4]
|
| 1989 |
+
# Note that X[:, 0] == 0 is a 100% indicator for y == 0. The tree can
|
| 1990 |
+
# easily learn that:
|
| 1991 |
+
reg = DecisionTreeRegressor(criterion="squared_error", random_state=seed)
|
| 1992 |
+
reg.fit(X, y)
|
| 1993 |
+
assert np.amin(reg.predict(X)) == 0
|
| 1994 |
+
# whereas Poisson must predict strictly positive numbers
|
| 1995 |
+
reg = DecisionTreeRegressor(criterion="poisson", random_state=seed)
|
| 1996 |
+
reg.fit(X, y)
|
| 1997 |
+
assert np.all(reg.predict(X) > 0)
|
| 1998 |
+
|
| 1999 |
+
# Test additional dataset where something could go wrong.
|
| 2000 |
+
n_features = 10
|
| 2001 |
+
X, y = datasets.make_regression(
|
| 2002 |
+
effective_rank=n_features * 2 // 3,
|
| 2003 |
+
tail_strength=0.6,
|
| 2004 |
+
n_samples=1_000,
|
| 2005 |
+
n_features=n_features,
|
| 2006 |
+
n_informative=n_features * 2 // 3,
|
| 2007 |
+
random_state=seed,
|
| 2008 |
+
)
|
| 2009 |
+
# some excess zeros
|
| 2010 |
+
y[(-1 < y) & (y < 0)] = 0
|
| 2011 |
+
# make sure the target is positive
|
| 2012 |
+
y = np.abs(y)
|
| 2013 |
+
reg = DecisionTreeRegressor(criterion="poisson", random_state=seed)
|
| 2014 |
+
reg.fit(X, y)
|
| 2015 |
+
assert np.all(reg.predict(X) > 0)
|
| 2016 |
+
|
| 2017 |
+
|
| 2018 |
+
def test_poisson_vs_mse():
|
| 2019 |
+
# For a Poisson distributed target, Poisson loss should give better results
|
| 2020 |
+
# than squared error measured in Poisson deviance as metric.
|
| 2021 |
+
# We have a similar test, test_poisson(), in
|
| 2022 |
+
# sklearn/ensemble/_hist_gradient_boosting/tests/test_gradient_boosting.py
|
| 2023 |
+
rng = np.random.RandomState(42)
|
| 2024 |
+
n_train, n_test, n_features = 500, 500, 10
|
| 2025 |
+
X = datasets.make_low_rank_matrix(
|
| 2026 |
+
n_samples=n_train + n_test, n_features=n_features, random_state=rng
|
| 2027 |
+
)
|
| 2028 |
+
# We create a log-linear Poisson model and downscale coef as it will get
|
| 2029 |
+
# exponentiated.
|
| 2030 |
+
coef = rng.uniform(low=-2, high=2, size=n_features) / np.max(X, axis=0)
|
| 2031 |
+
y = rng.poisson(lam=np.exp(X @ coef))
|
| 2032 |
+
X_train, X_test, y_train, y_test = train_test_split(
|
| 2033 |
+
X, y, test_size=n_test, random_state=rng
|
| 2034 |
+
)
|
| 2035 |
+
# We prevent some overfitting by setting min_samples_split=10.
|
| 2036 |
+
tree_poi = DecisionTreeRegressor(
|
| 2037 |
+
criterion="poisson", min_samples_split=10, random_state=rng
|
| 2038 |
+
)
|
| 2039 |
+
tree_mse = DecisionTreeRegressor(
|
| 2040 |
+
criterion="squared_error", min_samples_split=10, random_state=rng
|
| 2041 |
+
)
|
| 2042 |
+
|
| 2043 |
+
tree_poi.fit(X_train, y_train)
|
| 2044 |
+
tree_mse.fit(X_train, y_train)
|
| 2045 |
+
dummy = DummyRegressor(strategy="mean").fit(X_train, y_train)
|
| 2046 |
+
|
| 2047 |
+
for X, y, val in [(X_train, y_train, "train"), (X_test, y_test, "test")]:
|
| 2048 |
+
metric_poi = mean_poisson_deviance(y, tree_poi.predict(X))
|
| 2049 |
+
# squared_error might produce non-positive predictions => clip
|
| 2050 |
+
metric_mse = mean_poisson_deviance(y, np.clip(tree_mse.predict(X), 1e-15, None))
|
| 2051 |
+
metric_dummy = mean_poisson_deviance(y, dummy.predict(X))
|
| 2052 |
+
# As squared_error might correctly predict 0 in train set, its train
|
| 2053 |
+
# score can be better than Poisson. This is no longer the case for the
|
| 2054 |
+
# test set.
|
| 2055 |
+
if val == "test":
|
| 2056 |
+
assert metric_poi < 0.5 * metric_mse
|
| 2057 |
+
assert metric_poi < 0.75 * metric_dummy
|
| 2058 |
+
|
| 2059 |
+
|
| 2060 |
+
@pytest.mark.parametrize("criterion", REG_CRITERIONS)
|
| 2061 |
+
def test_decision_tree_regressor_sample_weight_consistency(criterion):
|
| 2062 |
+
"""Test that the impact of sample_weight is consistent."""
|
| 2063 |
+
tree_params = dict(criterion=criterion)
|
| 2064 |
+
tree = DecisionTreeRegressor(**tree_params, random_state=42)
|
| 2065 |
+
for kind in ["zeros", "ones"]:
|
| 2066 |
+
check_sample_weights_invariance(
|
| 2067 |
+
"DecisionTreeRegressor_" + criterion, tree, kind="zeros"
|
| 2068 |
+
)
|
| 2069 |
+
|
| 2070 |
+
rng = np.random.RandomState(0)
|
| 2071 |
+
n_samples, n_features = 10, 5
|
| 2072 |
+
|
| 2073 |
+
X = rng.rand(n_samples, n_features)
|
| 2074 |
+
y = np.mean(X, axis=1) + rng.rand(n_samples)
|
| 2075 |
+
# make it positive in order to work also for poisson criterion
|
| 2076 |
+
y += np.min(y) + 0.1
|
| 2077 |
+
|
| 2078 |
+
# check that multiplying sample_weight by 2 is equivalent
|
| 2079 |
+
# to repeating corresponding samples twice
|
| 2080 |
+
X2 = np.concatenate([X, X[: n_samples // 2]], axis=0)
|
| 2081 |
+
y2 = np.concatenate([y, y[: n_samples // 2]])
|
| 2082 |
+
sample_weight_1 = np.ones(len(y))
|
| 2083 |
+
sample_weight_1[: n_samples // 2] = 2
|
| 2084 |
+
|
| 2085 |
+
tree1 = DecisionTreeRegressor(**tree_params).fit(
|
| 2086 |
+
X, y, sample_weight=sample_weight_1
|
| 2087 |
+
)
|
| 2088 |
+
|
| 2089 |
+
tree2 = DecisionTreeRegressor(**tree_params).fit(X2, y2, sample_weight=None)
|
| 2090 |
+
|
| 2091 |
+
assert tree1.tree_.node_count == tree2.tree_.node_count
|
| 2092 |
+
# Thresholds, tree.tree_.threshold, and values, tree.tree_.value, are not
|
| 2093 |
+
# exactly the same, but on the training set, those differences do not
|
| 2094 |
+
# matter and thus predictions are the same.
|
| 2095 |
+
assert_allclose(tree1.predict(X), tree2.predict(X))
|
| 2096 |
+
|
| 2097 |
+
|
| 2098 |
+
@pytest.mark.parametrize("Tree", [DecisionTreeClassifier, ExtraTreeClassifier])
|
| 2099 |
+
@pytest.mark.parametrize("n_classes", [2, 4])
|
| 2100 |
+
def test_criterion_entropy_same_as_log_loss(Tree, n_classes):
|
| 2101 |
+
"""Test that criterion=entropy gives same as log_loss."""
|
| 2102 |
+
n_samples, n_features = 50, 5
|
| 2103 |
+
X, y = datasets.make_classification(
|
| 2104 |
+
n_classes=n_classes,
|
| 2105 |
+
n_samples=n_samples,
|
| 2106 |
+
n_features=n_features,
|
| 2107 |
+
n_informative=n_features,
|
| 2108 |
+
n_redundant=0,
|
| 2109 |
+
random_state=42,
|
| 2110 |
+
)
|
| 2111 |
+
tree_log_loss = Tree(criterion="log_loss", random_state=43).fit(X, y)
|
| 2112 |
+
tree_entropy = Tree(criterion="entropy", random_state=43).fit(X, y)
|
| 2113 |
+
|
| 2114 |
+
assert_tree_equal(
|
| 2115 |
+
tree_log_loss.tree_,
|
| 2116 |
+
tree_entropy.tree_,
|
| 2117 |
+
f"{Tree!r} with criterion 'entropy' and 'log_loss' gave different trees.",
|
| 2118 |
+
)
|
| 2119 |
+
assert_allclose(tree_log_loss.predict(X), tree_entropy.predict(X))
|
| 2120 |
+
|
| 2121 |
+
|
| 2122 |
+
def test_different_endianness_pickle():
|
| 2123 |
+
X, y = datasets.make_classification(random_state=0)
|
| 2124 |
+
|
| 2125 |
+
clf = DecisionTreeClassifier(random_state=0, max_depth=3)
|
| 2126 |
+
clf.fit(X, y)
|
| 2127 |
+
score = clf.score(X, y)
|
| 2128 |
+
|
| 2129 |
+
def reduce_ndarray(arr):
|
| 2130 |
+
return arr.byteswap().newbyteorder().__reduce__()
|
| 2131 |
+
|
| 2132 |
+
def get_pickle_non_native_endianness():
|
| 2133 |
+
f = io.BytesIO()
|
| 2134 |
+
p = pickle.Pickler(f)
|
| 2135 |
+
p.dispatch_table = copyreg.dispatch_table.copy()
|
| 2136 |
+
p.dispatch_table[np.ndarray] = reduce_ndarray
|
| 2137 |
+
|
| 2138 |
+
p.dump(clf)
|
| 2139 |
+
f.seek(0)
|
| 2140 |
+
return f
|
| 2141 |
+
|
| 2142 |
+
new_clf = pickle.load(get_pickle_non_native_endianness())
|
| 2143 |
+
new_score = new_clf.score(X, y)
|
| 2144 |
+
assert np.isclose(score, new_score)
|
| 2145 |
+
|
| 2146 |
+
|
| 2147 |
+
def test_different_endianness_joblib_pickle():
|
| 2148 |
+
X, y = datasets.make_classification(random_state=0)
|
| 2149 |
+
|
| 2150 |
+
clf = DecisionTreeClassifier(random_state=0, max_depth=3)
|
| 2151 |
+
clf.fit(X, y)
|
| 2152 |
+
score = clf.score(X, y)
|
| 2153 |
+
|
| 2154 |
+
class NonNativeEndiannessNumpyPickler(NumpyPickler):
|
| 2155 |
+
def save(self, obj):
|
| 2156 |
+
if isinstance(obj, np.ndarray):
|
| 2157 |
+
obj = obj.byteswap().newbyteorder()
|
| 2158 |
+
super().save(obj)
|
| 2159 |
+
|
| 2160 |
+
def get_joblib_pickle_non_native_endianness():
|
| 2161 |
+
f = io.BytesIO()
|
| 2162 |
+
p = NonNativeEndiannessNumpyPickler(f)
|
| 2163 |
+
|
| 2164 |
+
p.dump(clf)
|
| 2165 |
+
f.seek(0)
|
| 2166 |
+
return f
|
| 2167 |
+
|
| 2168 |
+
new_clf = joblib.load(get_joblib_pickle_non_native_endianness())
|
| 2169 |
+
new_score = new_clf.score(X, y)
|
| 2170 |
+
assert np.isclose(score, new_score)
|
| 2171 |
+
|
| 2172 |
+
|
| 2173 |
+
def get_different_bitness_node_ndarray(node_ndarray):
|
| 2174 |
+
new_dtype_for_indexing_fields = np.int64 if _IS_32BIT else np.int32
|
| 2175 |
+
|
| 2176 |
+
# field names in Node struct with SIZE_t types (see sklearn/tree/_tree.pxd)
|
| 2177 |
+
indexing_field_names = ["left_child", "right_child", "feature", "n_node_samples"]
|
| 2178 |
+
|
| 2179 |
+
new_dtype_dict = {
|
| 2180 |
+
name: dtype for name, (dtype, _) in node_ndarray.dtype.fields.items()
|
| 2181 |
+
}
|
| 2182 |
+
for name in indexing_field_names:
|
| 2183 |
+
new_dtype_dict[name] = new_dtype_for_indexing_fields
|
| 2184 |
+
|
| 2185 |
+
new_dtype = np.dtype(
|
| 2186 |
+
{"names": list(new_dtype_dict.keys()), "formats": list(new_dtype_dict.values())}
|
| 2187 |
+
)
|
| 2188 |
+
return node_ndarray.astype(new_dtype, casting="same_kind")
|
| 2189 |
+
|
| 2190 |
+
|
| 2191 |
+
def get_different_alignment_node_ndarray(node_ndarray):
|
| 2192 |
+
new_dtype_dict = {
|
| 2193 |
+
name: dtype for name, (dtype, _) in node_ndarray.dtype.fields.items()
|
| 2194 |
+
}
|
| 2195 |
+
offsets = [offset for dtype, offset in node_ndarray.dtype.fields.values()]
|
| 2196 |
+
shifted_offsets = [8 + offset for offset in offsets]
|
| 2197 |
+
|
| 2198 |
+
new_dtype = np.dtype(
|
| 2199 |
+
{
|
| 2200 |
+
"names": list(new_dtype_dict.keys()),
|
| 2201 |
+
"formats": list(new_dtype_dict.values()),
|
| 2202 |
+
"offsets": shifted_offsets,
|
| 2203 |
+
}
|
| 2204 |
+
)
|
| 2205 |
+
return node_ndarray.astype(new_dtype, casting="same_kind")
|
| 2206 |
+
|
| 2207 |
+
|
| 2208 |
+
def reduce_tree_with_different_bitness(tree):
|
| 2209 |
+
new_dtype = np.int64 if _IS_32BIT else np.int32
|
| 2210 |
+
tree_cls, (n_features, n_classes, n_outputs), state = tree.__reduce__()
|
| 2211 |
+
new_n_classes = n_classes.astype(new_dtype, casting="same_kind")
|
| 2212 |
+
|
| 2213 |
+
new_state = state.copy()
|
| 2214 |
+
new_state["nodes"] = get_different_bitness_node_ndarray(new_state["nodes"])
|
| 2215 |
+
|
| 2216 |
+
return (tree_cls, (n_features, new_n_classes, n_outputs), new_state)
|
| 2217 |
+
|
| 2218 |
+
|
| 2219 |
+
def test_different_bitness_pickle():
|
| 2220 |
+
X, y = datasets.make_classification(random_state=0)
|
| 2221 |
+
|
| 2222 |
+
clf = DecisionTreeClassifier(random_state=0, max_depth=3)
|
| 2223 |
+
clf.fit(X, y)
|
| 2224 |
+
score = clf.score(X, y)
|
| 2225 |
+
|
| 2226 |
+
def pickle_dump_with_different_bitness():
|
| 2227 |
+
f = io.BytesIO()
|
| 2228 |
+
p = pickle.Pickler(f)
|
| 2229 |
+
p.dispatch_table = copyreg.dispatch_table.copy()
|
| 2230 |
+
p.dispatch_table[CythonTree] = reduce_tree_with_different_bitness
|
| 2231 |
+
|
| 2232 |
+
p.dump(clf)
|
| 2233 |
+
f.seek(0)
|
| 2234 |
+
return f
|
| 2235 |
+
|
| 2236 |
+
new_clf = pickle.load(pickle_dump_with_different_bitness())
|
| 2237 |
+
new_score = new_clf.score(X, y)
|
| 2238 |
+
assert score == pytest.approx(new_score)
|
| 2239 |
+
|
| 2240 |
+
|
| 2241 |
+
def test_different_bitness_joblib_pickle():
|
| 2242 |
+
# Make sure that a platform specific pickle generated on a 64 bit
|
| 2243 |
+
# platform can be converted at pickle load time into an estimator
|
| 2244 |
+
# with Cython code that works with the host's native integer precision
|
| 2245 |
+
# to index nodes in the tree data structure when the host is a 32 bit
|
| 2246 |
+
# platform (and vice versa).
|
| 2247 |
+
X, y = datasets.make_classification(random_state=0)
|
| 2248 |
+
|
| 2249 |
+
clf = DecisionTreeClassifier(random_state=0, max_depth=3)
|
| 2250 |
+
clf.fit(X, y)
|
| 2251 |
+
score = clf.score(X, y)
|
| 2252 |
+
|
| 2253 |
+
def joblib_dump_with_different_bitness():
|
| 2254 |
+
f = io.BytesIO()
|
| 2255 |
+
p = NumpyPickler(f)
|
| 2256 |
+
p.dispatch_table = copyreg.dispatch_table.copy()
|
| 2257 |
+
p.dispatch_table[CythonTree] = reduce_tree_with_different_bitness
|
| 2258 |
+
|
| 2259 |
+
p.dump(clf)
|
| 2260 |
+
f.seek(0)
|
| 2261 |
+
return f
|
| 2262 |
+
|
| 2263 |
+
new_clf = joblib.load(joblib_dump_with_different_bitness())
|
| 2264 |
+
new_score = new_clf.score(X, y)
|
| 2265 |
+
assert score == pytest.approx(new_score)
|
| 2266 |
+
|
| 2267 |
+
|
| 2268 |
+
def test_check_n_classes():
|
| 2269 |
+
expected_dtype = np.dtype(np.int32) if _IS_32BIT else np.dtype(np.int64)
|
| 2270 |
+
allowed_dtypes = [np.dtype(np.int32), np.dtype(np.int64)]
|
| 2271 |
+
allowed_dtypes += [dt.newbyteorder() for dt in allowed_dtypes]
|
| 2272 |
+
|
| 2273 |
+
n_classes = np.array([0, 1], dtype=expected_dtype)
|
| 2274 |
+
for dt in allowed_dtypes:
|
| 2275 |
+
_check_n_classes(n_classes.astype(dt), expected_dtype)
|
| 2276 |
+
|
| 2277 |
+
with pytest.raises(ValueError, match="Wrong dimensions.+n_classes"):
|
| 2278 |
+
wrong_dim_n_classes = np.array([[0, 1]], dtype=expected_dtype)
|
| 2279 |
+
_check_n_classes(wrong_dim_n_classes, expected_dtype)
|
| 2280 |
+
|
| 2281 |
+
with pytest.raises(ValueError, match="n_classes.+incompatible dtype"):
|
| 2282 |
+
wrong_dtype_n_classes = n_classes.astype(np.float64)
|
| 2283 |
+
_check_n_classes(wrong_dtype_n_classes, expected_dtype)
|
| 2284 |
+
|
| 2285 |
+
|
| 2286 |
+
def test_check_value_ndarray():
|
| 2287 |
+
expected_dtype = np.dtype(np.float64)
|
| 2288 |
+
expected_shape = (5, 1, 2)
|
| 2289 |
+
value_ndarray = np.zeros(expected_shape, dtype=expected_dtype)
|
| 2290 |
+
|
| 2291 |
+
allowed_dtypes = [expected_dtype, expected_dtype.newbyteorder()]
|
| 2292 |
+
|
| 2293 |
+
for dt in allowed_dtypes:
|
| 2294 |
+
_check_value_ndarray(
|
| 2295 |
+
value_ndarray, expected_dtype=dt, expected_shape=expected_shape
|
| 2296 |
+
)
|
| 2297 |
+
|
| 2298 |
+
with pytest.raises(ValueError, match="Wrong shape.+value array"):
|
| 2299 |
+
_check_value_ndarray(
|
| 2300 |
+
value_ndarray, expected_dtype=expected_dtype, expected_shape=(1, 2)
|
| 2301 |
+
)
|
| 2302 |
+
|
| 2303 |
+
for problematic_arr in [value_ndarray[:, :, :1], np.asfortranarray(value_ndarray)]:
|
| 2304 |
+
with pytest.raises(ValueError, match="value array.+C-contiguous"):
|
| 2305 |
+
_check_value_ndarray(
|
| 2306 |
+
problematic_arr,
|
| 2307 |
+
expected_dtype=expected_dtype,
|
| 2308 |
+
expected_shape=problematic_arr.shape,
|
| 2309 |
+
)
|
| 2310 |
+
|
| 2311 |
+
with pytest.raises(ValueError, match="value array.+incompatible dtype"):
|
| 2312 |
+
_check_value_ndarray(
|
| 2313 |
+
value_ndarray.astype(np.float32),
|
| 2314 |
+
expected_dtype=expected_dtype,
|
| 2315 |
+
expected_shape=expected_shape,
|
| 2316 |
+
)
|
| 2317 |
+
|
| 2318 |
+
|
| 2319 |
+
def test_check_node_ndarray():
|
| 2320 |
+
expected_dtype = NODE_DTYPE
|
| 2321 |
+
|
| 2322 |
+
node_ndarray = np.zeros((5,), dtype=expected_dtype)
|
| 2323 |
+
|
| 2324 |
+
valid_node_ndarrays = [
|
| 2325 |
+
node_ndarray,
|
| 2326 |
+
get_different_bitness_node_ndarray(node_ndarray),
|
| 2327 |
+
get_different_alignment_node_ndarray(node_ndarray),
|
| 2328 |
+
]
|
| 2329 |
+
valid_node_ndarrays += [
|
| 2330 |
+
arr.astype(arr.dtype.newbyteorder()) for arr in valid_node_ndarrays
|
| 2331 |
+
]
|
| 2332 |
+
|
| 2333 |
+
for arr in valid_node_ndarrays:
|
| 2334 |
+
_check_node_ndarray(node_ndarray, expected_dtype=expected_dtype)
|
| 2335 |
+
|
| 2336 |
+
with pytest.raises(ValueError, match="Wrong dimensions.+node array"):
|
| 2337 |
+
problematic_node_ndarray = np.zeros((5, 2), dtype=expected_dtype)
|
| 2338 |
+
_check_node_ndarray(problematic_node_ndarray, expected_dtype=expected_dtype)
|
| 2339 |
+
|
| 2340 |
+
with pytest.raises(ValueError, match="node array.+C-contiguous"):
|
| 2341 |
+
problematic_node_ndarray = node_ndarray[::2]
|
| 2342 |
+
_check_node_ndarray(problematic_node_ndarray, expected_dtype=expected_dtype)
|
| 2343 |
+
|
| 2344 |
+
dtype_dict = {name: dtype for name, (dtype, _) in node_ndarray.dtype.fields.items()}
|
| 2345 |
+
|
| 2346 |
+
# array with wrong 'threshold' field dtype (int64 rather than float64)
|
| 2347 |
+
new_dtype_dict = dtype_dict.copy()
|
| 2348 |
+
new_dtype_dict["threshold"] = np.int64
|
| 2349 |
+
|
| 2350 |
+
new_dtype = np.dtype(
|
| 2351 |
+
{"names": list(new_dtype_dict.keys()), "formats": list(new_dtype_dict.values())}
|
| 2352 |
+
)
|
| 2353 |
+
problematic_node_ndarray = node_ndarray.astype(new_dtype)
|
| 2354 |
+
|
| 2355 |
+
with pytest.raises(ValueError, match="node array.+incompatible dtype"):
|
| 2356 |
+
_check_node_ndarray(problematic_node_ndarray, expected_dtype=expected_dtype)
|
| 2357 |
+
|
| 2358 |
+
# array with wrong 'left_child' field dtype (float64 rather than int64 or int32)
|
| 2359 |
+
new_dtype_dict = dtype_dict.copy()
|
| 2360 |
+
new_dtype_dict["left_child"] = np.float64
|
| 2361 |
+
new_dtype = np.dtype(
|
| 2362 |
+
{"names": list(new_dtype_dict.keys()), "formats": list(new_dtype_dict.values())}
|
| 2363 |
+
)
|
| 2364 |
+
|
| 2365 |
+
problematic_node_ndarray = node_ndarray.astype(new_dtype)
|
| 2366 |
+
|
| 2367 |
+
with pytest.raises(ValueError, match="node array.+incompatible dtype"):
|
| 2368 |
+
_check_node_ndarray(problematic_node_ndarray, expected_dtype=expected_dtype)
|
| 2369 |
+
|
| 2370 |
+
|
| 2371 |
+
# TODO(1.3): Remove
|
| 2372 |
+
def test_max_features_auto_deprecated():
|
| 2373 |
+
for Tree in CLF_TREES.values():
|
| 2374 |
+
tree = Tree(max_features="auto")
|
| 2375 |
+
msg = (
|
| 2376 |
+
"`max_features='auto'` has been deprecated in 1.1 and will be removed in"
|
| 2377 |
+
" 1.3. To keep the past behaviour, explicitly set `max_features='sqrt'`."
|
| 2378 |
+
)
|
| 2379 |
+
with pytest.warns(FutureWarning, match=msg):
|
| 2380 |
+
tree.fit(X, y)
|
| 2381 |
+
|
| 2382 |
+
for Tree in REG_TREES.values():
|
| 2383 |
+
tree = Tree(max_features="auto")
|
| 2384 |
+
msg = (
|
| 2385 |
+
"`max_features='auto'` has been deprecated in 1.1 and will be removed in"
|
| 2386 |
+
" 1.3. To keep the past behaviour, explicitly set `max_features=1.0'`."
|
| 2387 |
+
)
|
| 2388 |
+
with pytest.warns(FutureWarning, match=msg):
|
| 2389 |
+
tree.fit(X, y)
|
| 2390 |
+
|
| 2391 |
+
|
| 2392 |
+
def test_tree_deserialization_from_read_only_buffer(tmpdir):
|
| 2393 |
+
"""Check that Trees can be deserialized with read only buffers.
|
| 2394 |
+
|
| 2395 |
+
Non-regression test for gh-25584.
|
| 2396 |
+
"""
|
| 2397 |
+
pickle_path = str(tmpdir.join("clf.joblib"))
|
| 2398 |
+
clf = DecisionTreeClassifier(random_state=0)
|
| 2399 |
+
clf.fit(X_small, y_small)
|
| 2400 |
+
|
| 2401 |
+
joblib.dump(clf, pickle_path)
|
| 2402 |
+
loaded_clf = joblib.load(pickle_path, mmap_mode="r")
|
| 2403 |
+
|
| 2404 |
+
assert_tree_equal(
|
| 2405 |
+
loaded_clf.tree_,
|
| 2406 |
+
clf.tree_,
|
| 2407 |
+
"The trees of the original and loaded classifiers are not equal.",
|
| 2408 |
+
)
|
| 2409 |
+
|
| 2410 |
+
|
| 2411 |
+
@pytest.mark.parametrize("Tree", ALL_TREES.values())
|
| 2412 |
+
def test_min_sample_split_1_error(Tree):
|
| 2413 |
+
"""Check that an error is raised when min_sample_split=1.
|
| 2414 |
+
|
| 2415 |
+
non-regression test for issue gh-25481.
|
| 2416 |
+
"""
|
| 2417 |
+
X = np.array([[0, 0], [1, 1]])
|
| 2418 |
+
y = np.array([0, 1])
|
| 2419 |
+
|
| 2420 |
+
# min_samples_split=1.0 is valid
|
| 2421 |
+
Tree(min_samples_split=1.0).fit(X, y)
|
| 2422 |
+
|
| 2423 |
+
# min_samples_split=1 is invalid
|
| 2424 |
+
tree = Tree(min_samples_split=1)
|
| 2425 |
+
msg = (
|
| 2426 |
+
r"'min_samples_split' .* must be an int in the range \[2, inf\) "
|
| 2427 |
+
r"or a float in the range \(0.0, 1.0\]"
|
| 2428 |
+
)
|
| 2429 |
+
with pytest.raises(ValueError, match=msg):
|
| 2430 |
+
tree.fit(X, y)
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_arpack.py
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from .validation import check_random_state
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
def _init_arpack_v0(size, random_state):
|
| 5 |
+
"""Initialize the starting vector for iteration in ARPACK functions.
|
| 6 |
+
|
| 7 |
+
Initialize a ndarray with values sampled from the uniform distribution on
|
| 8 |
+
[-1, 1]. This initialization model has been chosen to be consistent with
|
| 9 |
+
the ARPACK one as another initialization can lead to convergence issues.
|
| 10 |
+
|
| 11 |
+
Parameters
|
| 12 |
+
----------
|
| 13 |
+
size : int
|
| 14 |
+
The size of the eigenvalue vector to be initialized.
|
| 15 |
+
|
| 16 |
+
random_state : int, RandomState instance or None, default=None
|
| 17 |
+
The seed of the pseudo random number generator used to generate a
|
| 18 |
+
uniform distribution. If int, random_state is the seed used by the
|
| 19 |
+
random number generator; If RandomState instance, random_state is the
|
| 20 |
+
random number generator; If None, the random number generator is the
|
| 21 |
+
RandomState instance used by `np.random`.
|
| 22 |
+
|
| 23 |
+
Returns
|
| 24 |
+
-------
|
| 25 |
+
v0 : ndarray of shape (size,)
|
| 26 |
+
The initialized vector.
|
| 27 |
+
"""
|
| 28 |
+
random_state = check_random_state(random_state)
|
| 29 |
+
v0 = random_state.uniform(-1, 1, size)
|
| 30 |
+
return v0
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_array_api.py
ADDED
|
@@ -0,0 +1,232 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Tools to support array_api."""
|
| 2 |
+
import numpy
|
| 3 |
+
from .._config import get_config
|
| 4 |
+
import scipy.special as special
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
class _ArrayAPIWrapper:
|
| 8 |
+
"""sklearn specific Array API compatibility wrapper
|
| 9 |
+
|
| 10 |
+
This wrapper makes it possible for scikit-learn maintainers to
|
| 11 |
+
deal with discrepancies between different implementations of the
|
| 12 |
+
Python array API standard and its evolution over time.
|
| 13 |
+
|
| 14 |
+
The Python array API standard specification:
|
| 15 |
+
https://data-apis.org/array-api/latest/
|
| 16 |
+
|
| 17 |
+
Documentation of the NumPy implementation:
|
| 18 |
+
https://numpy.org/neps/nep-0047-array-api-standard.html
|
| 19 |
+
"""
|
| 20 |
+
|
| 21 |
+
def __init__(self, array_namespace):
|
| 22 |
+
self._namespace = array_namespace
|
| 23 |
+
|
| 24 |
+
def __getattr__(self, name):
|
| 25 |
+
return getattr(self._namespace, name)
|
| 26 |
+
|
| 27 |
+
def take(self, X, indices, *, axis):
|
| 28 |
+
# When array_api supports `take` we can use this directly
|
| 29 |
+
# https://github.com/data-apis/array-api/issues/177
|
| 30 |
+
if self._namespace.__name__ == "numpy.array_api":
|
| 31 |
+
X_np = numpy.take(X, indices, axis=axis)
|
| 32 |
+
return self._namespace.asarray(X_np)
|
| 33 |
+
|
| 34 |
+
# We only support axis in (0, 1) and ndim in (1, 2) because that is all we need
|
| 35 |
+
# in scikit-learn
|
| 36 |
+
if axis not in {0, 1}:
|
| 37 |
+
raise ValueError(f"Only axis in (0, 1) is supported. Got {axis}")
|
| 38 |
+
|
| 39 |
+
if X.ndim not in {1, 2}:
|
| 40 |
+
raise ValueError(f"Only X.ndim in (1, 2) is supported. Got {X.ndim}")
|
| 41 |
+
|
| 42 |
+
if axis == 0:
|
| 43 |
+
if X.ndim == 1:
|
| 44 |
+
selected = [X[i] for i in indices]
|
| 45 |
+
else: # X.ndim == 2
|
| 46 |
+
selected = [X[i, :] for i in indices]
|
| 47 |
+
else: # axis == 1
|
| 48 |
+
selected = [X[:, i] for i in indices]
|
| 49 |
+
return self._namespace.stack(selected, axis=axis)
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
class _NumPyApiWrapper:
|
| 53 |
+
"""Array API compat wrapper for any numpy version
|
| 54 |
+
|
| 55 |
+
NumPy < 1.22 does not expose the numpy.array_api namespace. This
|
| 56 |
+
wrapper makes it possible to write code that uses the standard
|
| 57 |
+
Array API while working with any version of NumPy supported by
|
| 58 |
+
scikit-learn.
|
| 59 |
+
|
| 60 |
+
See the `get_namespace()` public function for more details.
|
| 61 |
+
"""
|
| 62 |
+
|
| 63 |
+
def __getattr__(self, name):
|
| 64 |
+
return getattr(numpy, name)
|
| 65 |
+
|
| 66 |
+
def astype(self, x, dtype, *, copy=True, casting="unsafe"):
|
| 67 |
+
# astype is not defined in the top level NumPy namespace
|
| 68 |
+
return x.astype(dtype, copy=copy, casting=casting)
|
| 69 |
+
|
| 70 |
+
def asarray(self, x, *, dtype=None, device=None, copy=None):
|
| 71 |
+
# Support copy in NumPy namespace
|
| 72 |
+
if copy is True:
|
| 73 |
+
return numpy.array(x, copy=True, dtype=dtype)
|
| 74 |
+
else:
|
| 75 |
+
return numpy.asarray(x, dtype=dtype)
|
| 76 |
+
|
| 77 |
+
def unique_inverse(self, x):
|
| 78 |
+
return numpy.unique(x, return_inverse=True)
|
| 79 |
+
|
| 80 |
+
def unique_counts(self, x):
|
| 81 |
+
return numpy.unique(x, return_counts=True)
|
| 82 |
+
|
| 83 |
+
def unique_values(self, x):
|
| 84 |
+
return numpy.unique(x)
|
| 85 |
+
|
| 86 |
+
def concat(self, arrays, *, axis=None):
|
| 87 |
+
return numpy.concatenate(arrays, axis=axis)
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
def get_namespace(*arrays):
|
| 91 |
+
"""Get namespace of arrays.
|
| 92 |
+
|
| 93 |
+
Introspect `arrays` arguments and return their common Array API
|
| 94 |
+
compatible namespace object, if any. NumPy 1.22 and later can
|
| 95 |
+
construct such containers using the `numpy.array_api` namespace
|
| 96 |
+
for instance.
|
| 97 |
+
|
| 98 |
+
See: https://numpy.org/neps/nep-0047-array-api-standard.html
|
| 99 |
+
|
| 100 |
+
If `arrays` are regular numpy arrays, an instance of the
|
| 101 |
+
`_NumPyApiWrapper` compatibility wrapper is returned instead.
|
| 102 |
+
|
| 103 |
+
Namespace support is not enabled by default. To enabled it
|
| 104 |
+
call:
|
| 105 |
+
|
| 106 |
+
sklearn.set_config(array_api_dispatch=True)
|
| 107 |
+
|
| 108 |
+
or:
|
| 109 |
+
|
| 110 |
+
with sklearn.config_context(array_api_dispatch=True):
|
| 111 |
+
# your code here
|
| 112 |
+
|
| 113 |
+
Otherwise an instance of the `_NumPyApiWrapper`
|
| 114 |
+
compatibility wrapper is always returned irrespective of
|
| 115 |
+
the fact that arrays implement the `__array_namespace__`
|
| 116 |
+
protocol or not.
|
| 117 |
+
|
| 118 |
+
Parameters
|
| 119 |
+
----------
|
| 120 |
+
*arrays : array objects
|
| 121 |
+
Array objects.
|
| 122 |
+
|
| 123 |
+
Returns
|
| 124 |
+
-------
|
| 125 |
+
namespace : module
|
| 126 |
+
Namespace shared by array objects.
|
| 127 |
+
|
| 128 |
+
is_array_api : bool
|
| 129 |
+
True of the arrays are containers that implement the Array API spec.
|
| 130 |
+
"""
|
| 131 |
+
# `arrays` contains one or more arrays, or possibly Python scalars (accepting
|
| 132 |
+
# those is a matter of taste, but doesn't seem unreasonable).
|
| 133 |
+
# Returns a tuple: (array_namespace, is_array_api)
|
| 134 |
+
|
| 135 |
+
if not get_config()["array_api_dispatch"]:
|
| 136 |
+
return _NumPyApiWrapper(), False
|
| 137 |
+
|
| 138 |
+
namespaces = {
|
| 139 |
+
x.__array_namespace__() if hasattr(x, "__array_namespace__") else None
|
| 140 |
+
for x in arrays
|
| 141 |
+
if not isinstance(x, (bool, int, float, complex))
|
| 142 |
+
}
|
| 143 |
+
|
| 144 |
+
if not namespaces:
|
| 145 |
+
# one could special-case np.ndarray above or use np.asarray here if
|
| 146 |
+
# older numpy versions need to be supported.
|
| 147 |
+
raise ValueError("Unrecognized array input")
|
| 148 |
+
|
| 149 |
+
if len(namespaces) != 1:
|
| 150 |
+
raise ValueError(f"Multiple namespaces for array inputs: {namespaces}")
|
| 151 |
+
|
| 152 |
+
(xp,) = namespaces
|
| 153 |
+
if xp is None:
|
| 154 |
+
# Use numpy as default
|
| 155 |
+
return _NumPyApiWrapper(), False
|
| 156 |
+
|
| 157 |
+
return _ArrayAPIWrapper(xp), True
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
def _expit(X):
|
| 161 |
+
xp, _ = get_namespace(X)
|
| 162 |
+
if xp.__name__ in {"numpy", "numpy.array_api"}:
|
| 163 |
+
return xp.asarray(special.expit(numpy.asarray(X)))
|
| 164 |
+
|
| 165 |
+
return 1.0 / (1.0 + xp.exp(-X))
|
| 166 |
+
|
| 167 |
+
|
| 168 |
+
def _asarray_with_order(array, dtype=None, order=None, copy=None, xp=None):
|
| 169 |
+
"""Helper to support the order kwarg only for NumPy-backed arrays
|
| 170 |
+
|
| 171 |
+
Memory layout parameter `order` is not exposed in the Array API standard,
|
| 172 |
+
however some input validation code in scikit-learn needs to work both
|
| 173 |
+
for classes and functions that will leverage Array API only operations
|
| 174 |
+
and for code that inherently relies on NumPy backed data containers with
|
| 175 |
+
specific memory layout constraints (e.g. our own Cython code). The
|
| 176 |
+
purpose of this helper is to make it possible to share code for data
|
| 177 |
+
container validation without memory copies for both downstream use cases:
|
| 178 |
+
the `order` parameter is only enforced if the input array implementation
|
| 179 |
+
is NumPy based, otherwise `order` is just silently ignored.
|
| 180 |
+
"""
|
| 181 |
+
if xp is None:
|
| 182 |
+
xp, _ = get_namespace(array)
|
| 183 |
+
if xp.__name__ in {"numpy", "numpy.array_api"}:
|
| 184 |
+
# Use NumPy API to support order
|
| 185 |
+
array = numpy.asarray(array, order=order, dtype=dtype)
|
| 186 |
+
return xp.asarray(array, copy=copy)
|
| 187 |
+
else:
|
| 188 |
+
return xp.asarray(array, dtype=dtype, copy=copy)
|
| 189 |
+
|
| 190 |
+
|
| 191 |
+
def _convert_to_numpy(array, xp):
|
| 192 |
+
"""Convert X into a NumPy ndarray.
|
| 193 |
+
|
| 194 |
+
Only works on cupy.array_api and numpy.array_api and is used for testing.
|
| 195 |
+
"""
|
| 196 |
+
supported_array_api = ["numpy.array_api", "cupy.array_api"]
|
| 197 |
+
if xp.__name__ not in supported_array_api:
|
| 198 |
+
support_array_api_str = ", ".join(supported_array_api)
|
| 199 |
+
raise ValueError(f"Supported namespaces are: {support_array_api_str}")
|
| 200 |
+
|
| 201 |
+
if xp.__name__ == "cupy.array_api":
|
| 202 |
+
return array._array.get()
|
| 203 |
+
else:
|
| 204 |
+
return numpy.asarray(array)
|
| 205 |
+
|
| 206 |
+
|
| 207 |
+
def _estimator_with_converted_arrays(estimator, converter):
|
| 208 |
+
"""Create new estimator which converting all attributes that are arrays.
|
| 209 |
+
|
| 210 |
+
Parameters
|
| 211 |
+
----------
|
| 212 |
+
estimator : Estimator
|
| 213 |
+
Estimator to convert
|
| 214 |
+
|
| 215 |
+
converter : callable
|
| 216 |
+
Callable that takes an array attribute and returns the converted array.
|
| 217 |
+
|
| 218 |
+
Returns
|
| 219 |
+
-------
|
| 220 |
+
new_estimator : Estimator
|
| 221 |
+
Convert estimator
|
| 222 |
+
"""
|
| 223 |
+
from sklearn.base import clone
|
| 224 |
+
|
| 225 |
+
new_estimator = clone(estimator)
|
| 226 |
+
for key, attribute in vars(estimator).items():
|
| 227 |
+
if hasattr(attribute, "__array_namespace__") or isinstance(
|
| 228 |
+
attribute, numpy.ndarray
|
| 229 |
+
):
|
| 230 |
+
attribute = converter(attribute)
|
| 231 |
+
setattr(new_estimator, key, attribute)
|
| 232 |
+
return new_estimator
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_joblib.py
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import warnings as _warnings
|
| 2 |
+
|
| 3 |
+
with _warnings.catch_warnings():
|
| 4 |
+
_warnings.simplefilter("ignore")
|
| 5 |
+
# joblib imports may raise DeprecationWarning on certain Python
|
| 6 |
+
# versions
|
| 7 |
+
import joblib
|
| 8 |
+
from joblib import logger
|
| 9 |
+
from joblib import dump, load
|
| 10 |
+
from joblib import __version__
|
| 11 |
+
from joblib import effective_n_jobs
|
| 12 |
+
from joblib import hash
|
| 13 |
+
from joblib import cpu_count, Parallel, Memory, delayed
|
| 14 |
+
from joblib import parallel_backend, register_parallel_backend
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
__all__ = [
|
| 18 |
+
"parallel_backend",
|
| 19 |
+
"register_parallel_backend",
|
| 20 |
+
"cpu_count",
|
| 21 |
+
"Parallel",
|
| 22 |
+
"Memory",
|
| 23 |
+
"delayed",
|
| 24 |
+
"effective_n_jobs",
|
| 25 |
+
"hash",
|
| 26 |
+
"logger",
|
| 27 |
+
"dump",
|
| 28 |
+
"load",
|
| 29 |
+
"joblib",
|
| 30 |
+
"__version__",
|
| 31 |
+
]
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_logistic_sigmoid.cpython-310-x86_64-linux-gnu.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a289dcf0576538ce83dae8afd304a7eabfae9ac387e1c2603cc0acde379d6c73
|
| 3 |
+
size 187849
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_mask.py
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
from scipy import sparse as sp
|
| 3 |
+
from contextlib import suppress
|
| 4 |
+
|
| 5 |
+
from . import is_scalar_nan
|
| 6 |
+
from .fixes import _object_dtype_isnan
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
def _get_dense_mask(X, value_to_mask):
|
| 10 |
+
with suppress(ImportError, AttributeError):
|
| 11 |
+
# We also suppress `AttributeError` because older versions of pandas do
|
| 12 |
+
# not have `NA`.
|
| 13 |
+
import pandas
|
| 14 |
+
|
| 15 |
+
if value_to_mask is pandas.NA:
|
| 16 |
+
return pandas.isna(X)
|
| 17 |
+
|
| 18 |
+
if is_scalar_nan(value_to_mask):
|
| 19 |
+
if X.dtype.kind == "f":
|
| 20 |
+
Xt = np.isnan(X)
|
| 21 |
+
elif X.dtype.kind in ("i", "u"):
|
| 22 |
+
# can't have NaNs in integer array.
|
| 23 |
+
Xt = np.zeros(X.shape, dtype=bool)
|
| 24 |
+
else:
|
| 25 |
+
# np.isnan does not work on object dtypes.
|
| 26 |
+
Xt = _object_dtype_isnan(X)
|
| 27 |
+
else:
|
| 28 |
+
Xt = X == value_to_mask
|
| 29 |
+
|
| 30 |
+
return Xt
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def _get_mask(X, value_to_mask):
|
| 34 |
+
"""Compute the boolean mask X == value_to_mask.
|
| 35 |
+
|
| 36 |
+
Parameters
|
| 37 |
+
----------
|
| 38 |
+
X : {ndarray, sparse matrix} of shape (n_samples, n_features)
|
| 39 |
+
Input data, where ``n_samples`` is the number of samples and
|
| 40 |
+
``n_features`` is the number of features.
|
| 41 |
+
|
| 42 |
+
value_to_mask : {int, float}
|
| 43 |
+
The value which is to be masked in X.
|
| 44 |
+
|
| 45 |
+
Returns
|
| 46 |
+
-------
|
| 47 |
+
X_mask : {ndarray, sparse matrix} of shape (n_samples, n_features)
|
| 48 |
+
Missing mask.
|
| 49 |
+
"""
|
| 50 |
+
if not sp.issparse(X):
|
| 51 |
+
# For all cases apart of a sparse input where we need to reconstruct
|
| 52 |
+
# a sparse output
|
| 53 |
+
return _get_dense_mask(X, value_to_mask)
|
| 54 |
+
|
| 55 |
+
Xt = _get_dense_mask(X.data, value_to_mask)
|
| 56 |
+
|
| 57 |
+
sparse_constructor = sp.csr_matrix if X.format == "csr" else sp.csc_matrix
|
| 58 |
+
Xt_sparse = sparse_constructor(
|
| 59 |
+
(Xt, X.indices.copy(), X.indptr.copy()), shape=X.shape, dtype=bool
|
| 60 |
+
)
|
| 61 |
+
|
| 62 |
+
return Xt_sparse
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_openmp_helpers.cpython-310-x86_64-linux-gnu.so
ADDED
|
Binary file (47.2 kB). View file
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_openmp_helpers.pxd
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Helpers to access OpenMP threads information
|
| 2 |
+
#
|
| 3 |
+
# Those interfaces act as indirections which allows the non-support of OpenMP
|
| 4 |
+
# for implementations which have been written for it.
|
| 5 |
+
|
| 6 |
+
cdef int _openmp_thread_num() nogil
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_param_validation.py
ADDED
|
@@ -0,0 +1,943 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from abc import ABC
|
| 2 |
+
from abc import abstractmethod
|
| 3 |
+
from collections.abc import Iterable
|
| 4 |
+
import functools
|
| 5 |
+
import math
|
| 6 |
+
from inspect import signature
|
| 7 |
+
from numbers import Integral
|
| 8 |
+
from numbers import Real
|
| 9 |
+
import operator
|
| 10 |
+
import re
|
| 11 |
+
import warnings
|
| 12 |
+
|
| 13 |
+
import numpy as np
|
| 14 |
+
from scipy.sparse import issparse
|
| 15 |
+
from scipy.sparse import csr_matrix
|
| 16 |
+
|
| 17 |
+
from .validation import _is_arraylike_not_scalar
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
class InvalidParameterError(ValueError, TypeError):
|
| 21 |
+
"""Custom exception to be raised when the parameter of a class/method/function
|
| 22 |
+
does not have a valid type or value.
|
| 23 |
+
"""
|
| 24 |
+
|
| 25 |
+
# Inherits from ValueError and TypeError to keep backward compatibility.
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def validate_parameter_constraints(parameter_constraints, params, caller_name):
|
| 29 |
+
"""Validate types and values of given parameters.
|
| 30 |
+
|
| 31 |
+
Parameters
|
| 32 |
+
----------
|
| 33 |
+
parameter_constraints : dict or {"no_validation"}
|
| 34 |
+
If "no_validation", validation is skipped for this parameter.
|
| 35 |
+
|
| 36 |
+
If a dict, it must be a dictionary `param_name: list of constraints`.
|
| 37 |
+
A parameter is valid if it satisfies one of the constraints from the list.
|
| 38 |
+
Constraints can be:
|
| 39 |
+
- an Interval object, representing a continuous or discrete range of numbers
|
| 40 |
+
- the string "array-like"
|
| 41 |
+
- the string "sparse matrix"
|
| 42 |
+
- the string "random_state"
|
| 43 |
+
- callable
|
| 44 |
+
- None, meaning that None is a valid value for the parameter
|
| 45 |
+
- any type, meaning that any instance of this type is valid
|
| 46 |
+
- an Options object, representing a set of elements of a given type
|
| 47 |
+
- a StrOptions object, representing a set of strings
|
| 48 |
+
- the string "boolean"
|
| 49 |
+
- the string "verbose"
|
| 50 |
+
- the string "cv_object"
|
| 51 |
+
- the string "missing_values"
|
| 52 |
+
- a HasMethods object, representing method(s) an object must have
|
| 53 |
+
- a Hidden object, representing a constraint not meant to be exposed to the user
|
| 54 |
+
|
| 55 |
+
params : dict
|
| 56 |
+
A dictionary `param_name: param_value`. The parameters to validate against the
|
| 57 |
+
constraints.
|
| 58 |
+
|
| 59 |
+
caller_name : str
|
| 60 |
+
The name of the estimator or function or method that called this function.
|
| 61 |
+
"""
|
| 62 |
+
for param_name, param_val in params.items():
|
| 63 |
+
# We allow parameters to not have a constraint so that third party estimators
|
| 64 |
+
# can inherit from sklearn estimators without having to necessarily use the
|
| 65 |
+
# validation tools.
|
| 66 |
+
if param_name not in parameter_constraints:
|
| 67 |
+
continue
|
| 68 |
+
|
| 69 |
+
constraints = parameter_constraints[param_name]
|
| 70 |
+
|
| 71 |
+
if constraints == "no_validation":
|
| 72 |
+
continue
|
| 73 |
+
|
| 74 |
+
constraints = [make_constraint(constraint) for constraint in constraints]
|
| 75 |
+
|
| 76 |
+
for constraint in constraints:
|
| 77 |
+
if constraint.is_satisfied_by(param_val):
|
| 78 |
+
# this constraint is satisfied, no need to check further.
|
| 79 |
+
break
|
| 80 |
+
else:
|
| 81 |
+
# No constraint is satisfied, raise with an informative message.
|
| 82 |
+
|
| 83 |
+
# Ignore constraints that we don't want to expose in the error message,
|
| 84 |
+
# i.e. options that are for internal purpose or not officially supported.
|
| 85 |
+
constraints = [
|
| 86 |
+
constraint for constraint in constraints if not constraint.hidden
|
| 87 |
+
]
|
| 88 |
+
|
| 89 |
+
if len(constraints) == 1:
|
| 90 |
+
constraints_str = f"{constraints[0]}"
|
| 91 |
+
else:
|
| 92 |
+
constraints_str = (
|
| 93 |
+
f"{', '.join([str(c) for c in constraints[:-1]])} or"
|
| 94 |
+
f" {constraints[-1]}"
|
| 95 |
+
)
|
| 96 |
+
|
| 97 |
+
raise InvalidParameterError(
|
| 98 |
+
f"The {param_name!r} parameter of {caller_name} must be"
|
| 99 |
+
f" {constraints_str}. Got {param_val!r} instead."
|
| 100 |
+
)
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
def make_constraint(constraint):
|
| 104 |
+
"""Convert the constraint into the appropriate Constraint object.
|
| 105 |
+
|
| 106 |
+
Parameters
|
| 107 |
+
----------
|
| 108 |
+
constraint : object
|
| 109 |
+
The constraint to convert.
|
| 110 |
+
|
| 111 |
+
Returns
|
| 112 |
+
-------
|
| 113 |
+
constraint : instance of _Constraint
|
| 114 |
+
The converted constraint.
|
| 115 |
+
"""
|
| 116 |
+
if isinstance(constraint, str) and constraint == "array-like":
|
| 117 |
+
return _ArrayLikes()
|
| 118 |
+
if isinstance(constraint, str) and constraint == "sparse matrix":
|
| 119 |
+
return _SparseMatrices()
|
| 120 |
+
if isinstance(constraint, str) and constraint == "random_state":
|
| 121 |
+
return _RandomStates()
|
| 122 |
+
if constraint is callable:
|
| 123 |
+
return _Callables()
|
| 124 |
+
if constraint is None:
|
| 125 |
+
return _NoneConstraint()
|
| 126 |
+
if isinstance(constraint, type):
|
| 127 |
+
return _InstancesOf(constraint)
|
| 128 |
+
if isinstance(constraint, (Interval, StrOptions, Options, HasMethods)):
|
| 129 |
+
return constraint
|
| 130 |
+
if isinstance(constraint, str) and constraint == "boolean":
|
| 131 |
+
return _Booleans()
|
| 132 |
+
if isinstance(constraint, str) and constraint == "verbose":
|
| 133 |
+
return _VerboseHelper()
|
| 134 |
+
if isinstance(constraint, str) and constraint == "missing_values":
|
| 135 |
+
return _MissingValues()
|
| 136 |
+
if isinstance(constraint, str) and constraint == "cv_object":
|
| 137 |
+
return _CVObjects()
|
| 138 |
+
if isinstance(constraint, Hidden):
|
| 139 |
+
constraint = make_constraint(constraint.constraint)
|
| 140 |
+
constraint.hidden = True
|
| 141 |
+
return constraint
|
| 142 |
+
raise ValueError(f"Unknown constraint type: {constraint}")
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
def validate_params(parameter_constraints):
|
| 146 |
+
"""Decorator to validate types and values of functions and methods.
|
| 147 |
+
|
| 148 |
+
Parameters
|
| 149 |
+
----------
|
| 150 |
+
parameter_constraints : dict
|
| 151 |
+
A dictionary `param_name: list of constraints`. See the docstring of
|
| 152 |
+
`validate_parameter_constraints` for a description of the accepted constraints.
|
| 153 |
+
|
| 154 |
+
Note that the *args and **kwargs parameters are not validated and must not be
|
| 155 |
+
present in the parameter_constraints dictionary.
|
| 156 |
+
|
| 157 |
+
Returns
|
| 158 |
+
-------
|
| 159 |
+
decorated_function : function or method
|
| 160 |
+
The decorated function.
|
| 161 |
+
"""
|
| 162 |
+
|
| 163 |
+
def decorator(func):
|
| 164 |
+
# The dict of parameter constraints is set as an attribute of the function
|
| 165 |
+
# to make it possible to dynamically introspect the constraints for
|
| 166 |
+
# automatic testing.
|
| 167 |
+
setattr(func, "_skl_parameter_constraints", parameter_constraints)
|
| 168 |
+
|
| 169 |
+
@functools.wraps(func)
|
| 170 |
+
def wrapper(*args, **kwargs):
|
| 171 |
+
|
| 172 |
+
func_sig = signature(func)
|
| 173 |
+
|
| 174 |
+
# Map *args/**kwargs to the function signature
|
| 175 |
+
params = func_sig.bind(*args, **kwargs)
|
| 176 |
+
params.apply_defaults()
|
| 177 |
+
|
| 178 |
+
# ignore self/cls and positional/keyword markers
|
| 179 |
+
to_ignore = [
|
| 180 |
+
p.name
|
| 181 |
+
for p in func_sig.parameters.values()
|
| 182 |
+
if p.kind in (p.VAR_POSITIONAL, p.VAR_KEYWORD)
|
| 183 |
+
]
|
| 184 |
+
to_ignore += ["self", "cls"]
|
| 185 |
+
params = {k: v for k, v in params.arguments.items() if k not in to_ignore}
|
| 186 |
+
|
| 187 |
+
validate_parameter_constraints(
|
| 188 |
+
parameter_constraints, params, caller_name=func.__qualname__
|
| 189 |
+
)
|
| 190 |
+
|
| 191 |
+
try:
|
| 192 |
+
return func(*args, **kwargs)
|
| 193 |
+
except InvalidParameterError as e:
|
| 194 |
+
# When the function is just a wrapper around an estimator, we allow
|
| 195 |
+
# the function to delegate validation to the estimator, but we replace
|
| 196 |
+
# the name of the estimator by the name of the function in the error
|
| 197 |
+
# message to avoid confusion.
|
| 198 |
+
msg = re.sub(
|
| 199 |
+
r"parameter of \w+ must be",
|
| 200 |
+
f"parameter of {func.__qualname__} must be",
|
| 201 |
+
str(e),
|
| 202 |
+
)
|
| 203 |
+
raise InvalidParameterError(msg) from e
|
| 204 |
+
|
| 205 |
+
return wrapper
|
| 206 |
+
|
| 207 |
+
return decorator
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
def _type_name(t):
|
| 211 |
+
"""Convert type into human readable string."""
|
| 212 |
+
module = t.__module__
|
| 213 |
+
qualname = t.__qualname__
|
| 214 |
+
if module == "builtins":
|
| 215 |
+
return qualname
|
| 216 |
+
elif t == Real:
|
| 217 |
+
return "float"
|
| 218 |
+
elif t == Integral:
|
| 219 |
+
return "int"
|
| 220 |
+
return f"{module}.{qualname}"
|
| 221 |
+
|
| 222 |
+
|
| 223 |
+
class _Constraint(ABC):
|
| 224 |
+
"""Base class for the constraint objects."""
|
| 225 |
+
|
| 226 |
+
def __init__(self):
|
| 227 |
+
self.hidden = False
|
| 228 |
+
|
| 229 |
+
@abstractmethod
|
| 230 |
+
def is_satisfied_by(self, val):
|
| 231 |
+
"""Whether or not a value satisfies the constraint.
|
| 232 |
+
|
| 233 |
+
Parameters
|
| 234 |
+
----------
|
| 235 |
+
val : object
|
| 236 |
+
The value to check.
|
| 237 |
+
|
| 238 |
+
Returns
|
| 239 |
+
-------
|
| 240 |
+
is_satisfied : bool
|
| 241 |
+
Whether or not the constraint is satisfied by this value.
|
| 242 |
+
"""
|
| 243 |
+
|
| 244 |
+
@abstractmethod
|
| 245 |
+
def __str__(self):
|
| 246 |
+
"""A human readable representational string of the constraint."""
|
| 247 |
+
|
| 248 |
+
|
| 249 |
+
class _InstancesOf(_Constraint):
|
| 250 |
+
"""Constraint representing instances of a given type.
|
| 251 |
+
|
| 252 |
+
Parameters
|
| 253 |
+
----------
|
| 254 |
+
type : type
|
| 255 |
+
The valid type.
|
| 256 |
+
"""
|
| 257 |
+
|
| 258 |
+
def __init__(self, type):
|
| 259 |
+
super().__init__()
|
| 260 |
+
self.type = type
|
| 261 |
+
|
| 262 |
+
def is_satisfied_by(self, val):
|
| 263 |
+
return isinstance(val, self.type)
|
| 264 |
+
|
| 265 |
+
def __str__(self):
|
| 266 |
+
return f"an instance of {_type_name(self.type)!r}"
|
| 267 |
+
|
| 268 |
+
|
| 269 |
+
class _NoneConstraint(_Constraint):
|
| 270 |
+
"""Constraint representing the None singleton."""
|
| 271 |
+
|
| 272 |
+
def is_satisfied_by(self, val):
|
| 273 |
+
return val is None
|
| 274 |
+
|
| 275 |
+
def __str__(self):
|
| 276 |
+
return "None"
|
| 277 |
+
|
| 278 |
+
|
| 279 |
+
class _NanConstraint(_Constraint):
|
| 280 |
+
"""Constraint representing the indicator `np.nan`."""
|
| 281 |
+
|
| 282 |
+
def is_satisfied_by(self, val):
|
| 283 |
+
return isinstance(val, Real) and math.isnan(val)
|
| 284 |
+
|
| 285 |
+
def __str__(self):
|
| 286 |
+
return "numpy.nan"
|
| 287 |
+
|
| 288 |
+
|
| 289 |
+
class _PandasNAConstraint(_Constraint):
|
| 290 |
+
"""Constraint representing the indicator `pd.NA`."""
|
| 291 |
+
|
| 292 |
+
def is_satisfied_by(self, val):
|
| 293 |
+
try:
|
| 294 |
+
import pandas as pd
|
| 295 |
+
|
| 296 |
+
return isinstance(val, type(pd.NA)) and pd.isna(val)
|
| 297 |
+
except ImportError:
|
| 298 |
+
return False
|
| 299 |
+
|
| 300 |
+
def __str__(self):
|
| 301 |
+
return "pandas.NA"
|
| 302 |
+
|
| 303 |
+
|
| 304 |
+
class Options(_Constraint):
|
| 305 |
+
"""Constraint representing a finite set of instances of a given type.
|
| 306 |
+
|
| 307 |
+
Parameters
|
| 308 |
+
----------
|
| 309 |
+
type : type
|
| 310 |
+
|
| 311 |
+
options : set
|
| 312 |
+
The set of valid scalars.
|
| 313 |
+
|
| 314 |
+
deprecated : set or None, default=None
|
| 315 |
+
A subset of the `options` to mark as deprecated in the string
|
| 316 |
+
representation of the constraint.
|
| 317 |
+
"""
|
| 318 |
+
|
| 319 |
+
def __init__(self, type, options, *, deprecated=None):
|
| 320 |
+
super().__init__()
|
| 321 |
+
self.type = type
|
| 322 |
+
self.options = options
|
| 323 |
+
self.deprecated = deprecated or set()
|
| 324 |
+
|
| 325 |
+
if self.deprecated - self.options:
|
| 326 |
+
raise ValueError("The deprecated options must be a subset of the options.")
|
| 327 |
+
|
| 328 |
+
def is_satisfied_by(self, val):
|
| 329 |
+
return isinstance(val, self.type) and val in self.options
|
| 330 |
+
|
| 331 |
+
def _mark_if_deprecated(self, option):
|
| 332 |
+
"""Add a deprecated mark to an option if needed."""
|
| 333 |
+
option_str = f"{option!r}"
|
| 334 |
+
if option in self.deprecated:
|
| 335 |
+
option_str = f"{option_str} (deprecated)"
|
| 336 |
+
return option_str
|
| 337 |
+
|
| 338 |
+
def __str__(self):
|
| 339 |
+
options_str = (
|
| 340 |
+
f"{', '.join([self._mark_if_deprecated(o) for o in self.options])}"
|
| 341 |
+
)
|
| 342 |
+
return f"a {_type_name(self.type)} among {{{options_str}}}"
|
| 343 |
+
|
| 344 |
+
|
| 345 |
+
class StrOptions(Options):
|
| 346 |
+
"""Constraint representing a finite set of strings.
|
| 347 |
+
|
| 348 |
+
Parameters
|
| 349 |
+
----------
|
| 350 |
+
options : set of str
|
| 351 |
+
The set of valid strings.
|
| 352 |
+
|
| 353 |
+
deprecated : set of str or None, default=None
|
| 354 |
+
A subset of the `options` to mark as deprecated in the string
|
| 355 |
+
representation of the constraint.
|
| 356 |
+
"""
|
| 357 |
+
|
| 358 |
+
def __init__(self, options, *, deprecated=None):
|
| 359 |
+
super().__init__(type=str, options=options, deprecated=deprecated)
|
| 360 |
+
|
| 361 |
+
|
| 362 |
+
class Interval(_Constraint):
|
| 363 |
+
"""Constraint representing a typed interval.
|
| 364 |
+
|
| 365 |
+
Parameters
|
| 366 |
+
----------
|
| 367 |
+
type : {numbers.Integral, numbers.Real, "real_not_int"}
|
| 368 |
+
The set of numbers in which to set the interval.
|
| 369 |
+
|
| 370 |
+
If "real_not_int", only reals that don't have the integer type
|
| 371 |
+
are allowed. For example 1.0 is allowed but 1 is not.
|
| 372 |
+
|
| 373 |
+
left : float or int or None
|
| 374 |
+
The left bound of the interval. None means left bound is -∞.
|
| 375 |
+
|
| 376 |
+
right : float, int or None
|
| 377 |
+
The right bound of the interval. None means right bound is +∞.
|
| 378 |
+
|
| 379 |
+
closed : {"left", "right", "both", "neither"}
|
| 380 |
+
Whether the interval is open or closed. Possible choices are:
|
| 381 |
+
|
| 382 |
+
- `"left"`: the interval is closed on the left and open on the right.
|
| 383 |
+
It is equivalent to the interval `[ left, right )`.
|
| 384 |
+
- `"right"`: the interval is closed on the right and open on the left.
|
| 385 |
+
It is equivalent to the interval `( left, right ]`.
|
| 386 |
+
- `"both"`: the interval is closed.
|
| 387 |
+
It is equivalent to the interval `[ left, right ]`.
|
| 388 |
+
- `"neither"`: the interval is open.
|
| 389 |
+
It is equivalent to the interval `( left, right )`.
|
| 390 |
+
|
| 391 |
+
Notes
|
| 392 |
+
-----
|
| 393 |
+
Setting a bound to `None` and setting the interval closed is valid. For instance,
|
| 394 |
+
strictly speaking, `Interval(Real, 0, None, closed="both")` corresponds to
|
| 395 |
+
`[0, +∞) U {+∞}`.
|
| 396 |
+
"""
|
| 397 |
+
|
| 398 |
+
def __init__(self, type, left, right, *, closed):
|
| 399 |
+
super().__init__()
|
| 400 |
+
self.type = type
|
| 401 |
+
self.left = left
|
| 402 |
+
self.right = right
|
| 403 |
+
self.closed = closed
|
| 404 |
+
|
| 405 |
+
self._check_params()
|
| 406 |
+
|
| 407 |
+
def _check_params(self):
|
| 408 |
+
if self.type not in (Integral, Real, "real_not_int"):
|
| 409 |
+
raise ValueError(
|
| 410 |
+
"type must be either numbers.Integral, numbers.Real or 'real_not_int'."
|
| 411 |
+
f" Got {self.type} instead."
|
| 412 |
+
)
|
| 413 |
+
|
| 414 |
+
if self.closed not in ("left", "right", "both", "neither"):
|
| 415 |
+
raise ValueError(
|
| 416 |
+
"closed must be either 'left', 'right', 'both' or 'neither'. "
|
| 417 |
+
f"Got {self.closed} instead."
|
| 418 |
+
)
|
| 419 |
+
|
| 420 |
+
if self.type is Integral:
|
| 421 |
+
suffix = "for an interval over the integers."
|
| 422 |
+
if self.left is not None and not isinstance(self.left, Integral):
|
| 423 |
+
raise TypeError(f"Expecting left to be an int {suffix}")
|
| 424 |
+
if self.right is not None and not isinstance(self.right, Integral):
|
| 425 |
+
raise TypeError(f"Expecting right to be an int {suffix}")
|
| 426 |
+
if self.left is None and self.closed in ("left", "both"):
|
| 427 |
+
raise ValueError(
|
| 428 |
+
f"left can't be None when closed == {self.closed} {suffix}"
|
| 429 |
+
)
|
| 430 |
+
if self.right is None and self.closed in ("right", "both"):
|
| 431 |
+
raise ValueError(
|
| 432 |
+
f"right can't be None when closed == {self.closed} {suffix}"
|
| 433 |
+
)
|
| 434 |
+
else:
|
| 435 |
+
if self.left is not None and not isinstance(self.left, Real):
|
| 436 |
+
raise TypeError("Expecting left to be a real number.")
|
| 437 |
+
if self.right is not None and not isinstance(self.right, Real):
|
| 438 |
+
raise TypeError("Expecting right to be a real number.")
|
| 439 |
+
|
| 440 |
+
if self.right is not None and self.left is not None and self.right <= self.left:
|
| 441 |
+
raise ValueError(
|
| 442 |
+
f"right can't be less than left. Got left={self.left} and "
|
| 443 |
+
f"right={self.right}"
|
| 444 |
+
)
|
| 445 |
+
|
| 446 |
+
def __contains__(self, val):
|
| 447 |
+
if np.isnan(val):
|
| 448 |
+
return False
|
| 449 |
+
|
| 450 |
+
left_cmp = operator.lt if self.closed in ("left", "both") else operator.le
|
| 451 |
+
right_cmp = operator.gt if self.closed in ("right", "both") else operator.ge
|
| 452 |
+
|
| 453 |
+
left = -np.inf if self.left is None else self.left
|
| 454 |
+
right = np.inf if self.right is None else self.right
|
| 455 |
+
|
| 456 |
+
if left_cmp(val, left):
|
| 457 |
+
return False
|
| 458 |
+
if right_cmp(val, right):
|
| 459 |
+
return False
|
| 460 |
+
return True
|
| 461 |
+
|
| 462 |
+
def _has_valid_type(self, val):
|
| 463 |
+
if self.type == "real_not_int":
|
| 464 |
+
return isinstance(val, Real) and not isinstance(val, Integral)
|
| 465 |
+
return isinstance(val, self.type)
|
| 466 |
+
|
| 467 |
+
def is_satisfied_by(self, val):
|
| 468 |
+
if not self._has_valid_type(val):
|
| 469 |
+
return False
|
| 470 |
+
|
| 471 |
+
return val in self
|
| 472 |
+
|
| 473 |
+
def __str__(self):
|
| 474 |
+
type_str = "an int" if self.type is Integral else "a float"
|
| 475 |
+
left_bracket = "[" if self.closed in ("left", "both") else "("
|
| 476 |
+
left_bound = "-inf" if self.left is None else self.left
|
| 477 |
+
right_bound = "inf" if self.right is None else self.right
|
| 478 |
+
right_bracket = "]" if self.closed in ("right", "both") else ")"
|
| 479 |
+
return (
|
| 480 |
+
f"{type_str} in the range "
|
| 481 |
+
f"{left_bracket}{left_bound}, {right_bound}{right_bracket}"
|
| 482 |
+
)
|
| 483 |
+
|
| 484 |
+
|
| 485 |
+
class _ArrayLikes(_Constraint):
|
| 486 |
+
"""Constraint representing array-likes"""
|
| 487 |
+
|
| 488 |
+
def is_satisfied_by(self, val):
|
| 489 |
+
return _is_arraylike_not_scalar(val)
|
| 490 |
+
|
| 491 |
+
def __str__(self):
|
| 492 |
+
return "an array-like"
|
| 493 |
+
|
| 494 |
+
|
| 495 |
+
class _SparseMatrices(_Constraint):
|
| 496 |
+
"""Constraint representing sparse matrices."""
|
| 497 |
+
|
| 498 |
+
def is_satisfied_by(self, val):
|
| 499 |
+
return issparse(val)
|
| 500 |
+
|
| 501 |
+
def __str__(self):
|
| 502 |
+
return "a sparse matrix"
|
| 503 |
+
|
| 504 |
+
|
| 505 |
+
class _Callables(_Constraint):
|
| 506 |
+
"""Constraint representing callables."""
|
| 507 |
+
|
| 508 |
+
def is_satisfied_by(self, val):
|
| 509 |
+
return callable(val)
|
| 510 |
+
|
| 511 |
+
def __str__(self):
|
| 512 |
+
return "a callable"
|
| 513 |
+
|
| 514 |
+
|
| 515 |
+
class _RandomStates(_Constraint):
|
| 516 |
+
"""Constraint representing random states.
|
| 517 |
+
|
| 518 |
+
Convenience class for
|
| 519 |
+
[Interval(Integral, 0, 2**32 - 1, closed="both"), np.random.RandomState, None]
|
| 520 |
+
"""
|
| 521 |
+
|
| 522 |
+
def __init__(self):
|
| 523 |
+
super().__init__()
|
| 524 |
+
self._constraints = [
|
| 525 |
+
Interval(Integral, 0, 2**32 - 1, closed="both"),
|
| 526 |
+
_InstancesOf(np.random.RandomState),
|
| 527 |
+
_NoneConstraint(),
|
| 528 |
+
]
|
| 529 |
+
|
| 530 |
+
def is_satisfied_by(self, val):
|
| 531 |
+
return any(c.is_satisfied_by(val) for c in self._constraints)
|
| 532 |
+
|
| 533 |
+
def __str__(self):
|
| 534 |
+
return (
|
| 535 |
+
f"{', '.join([str(c) for c in self._constraints[:-1]])} or"
|
| 536 |
+
f" {self._constraints[-1]}"
|
| 537 |
+
)
|
| 538 |
+
|
| 539 |
+
|
| 540 |
+
class _Booleans(_Constraint):
|
| 541 |
+
"""Constraint representing boolean likes.
|
| 542 |
+
|
| 543 |
+
Convenience class for
|
| 544 |
+
[bool, np.bool_, Integral (deprecated)]
|
| 545 |
+
"""
|
| 546 |
+
|
| 547 |
+
def __init__(self):
|
| 548 |
+
super().__init__()
|
| 549 |
+
self._constraints = [
|
| 550 |
+
_InstancesOf(bool),
|
| 551 |
+
_InstancesOf(np.bool_),
|
| 552 |
+
_InstancesOf(Integral),
|
| 553 |
+
]
|
| 554 |
+
|
| 555 |
+
def is_satisfied_by(self, val):
|
| 556 |
+
# TODO(1.4) remove support for Integral.
|
| 557 |
+
if isinstance(val, Integral) and not isinstance(val, bool):
|
| 558 |
+
warnings.warn(
|
| 559 |
+
"Passing an int for a boolean parameter is deprecated in version 1.2 "
|
| 560 |
+
"and won't be supported anymore in version 1.4.",
|
| 561 |
+
FutureWarning,
|
| 562 |
+
)
|
| 563 |
+
|
| 564 |
+
return any(c.is_satisfied_by(val) for c in self._constraints)
|
| 565 |
+
|
| 566 |
+
def __str__(self):
|
| 567 |
+
return (
|
| 568 |
+
f"{', '.join([str(c) for c in self._constraints[:-1]])} or"
|
| 569 |
+
f" {self._constraints[-1]}"
|
| 570 |
+
)
|
| 571 |
+
|
| 572 |
+
|
| 573 |
+
class _VerboseHelper(_Constraint):
|
| 574 |
+
"""Helper constraint for the verbose parameter.
|
| 575 |
+
|
| 576 |
+
Convenience class for
|
| 577 |
+
[Interval(Integral, 0, None, closed="left"), bool, numpy.bool_]
|
| 578 |
+
"""
|
| 579 |
+
|
| 580 |
+
def __init__(self):
|
| 581 |
+
super().__init__()
|
| 582 |
+
self._constraints = [
|
| 583 |
+
Interval(Integral, 0, None, closed="left"),
|
| 584 |
+
_InstancesOf(bool),
|
| 585 |
+
_InstancesOf(np.bool_),
|
| 586 |
+
]
|
| 587 |
+
|
| 588 |
+
def is_satisfied_by(self, val):
|
| 589 |
+
return any(c.is_satisfied_by(val) for c in self._constraints)
|
| 590 |
+
|
| 591 |
+
def __str__(self):
|
| 592 |
+
return (
|
| 593 |
+
f"{', '.join([str(c) for c in self._constraints[:-1]])} or"
|
| 594 |
+
f" {self._constraints[-1]}"
|
| 595 |
+
)
|
| 596 |
+
|
| 597 |
+
|
| 598 |
+
class _MissingValues(_Constraint):
|
| 599 |
+
"""Helper constraint for the `missing_values` parameters.
|
| 600 |
+
|
| 601 |
+
Convenience for
|
| 602 |
+
[
|
| 603 |
+
Integral,
|
| 604 |
+
Interval(Real, None, None, closed="both"),
|
| 605 |
+
str,
|
| 606 |
+
None,
|
| 607 |
+
_NanConstraint(),
|
| 608 |
+
_PandasNAConstraint(),
|
| 609 |
+
]
|
| 610 |
+
"""
|
| 611 |
+
|
| 612 |
+
def __init__(self):
|
| 613 |
+
super().__init__()
|
| 614 |
+
self._constraints = [
|
| 615 |
+
_InstancesOf(Integral),
|
| 616 |
+
# we use an interval of Real to ignore np.nan that has its own constraint
|
| 617 |
+
Interval(Real, None, None, closed="both"),
|
| 618 |
+
_InstancesOf(str),
|
| 619 |
+
_NoneConstraint(),
|
| 620 |
+
_NanConstraint(),
|
| 621 |
+
_PandasNAConstraint(),
|
| 622 |
+
]
|
| 623 |
+
|
| 624 |
+
def is_satisfied_by(self, val):
|
| 625 |
+
return any(c.is_satisfied_by(val) for c in self._constraints)
|
| 626 |
+
|
| 627 |
+
def __str__(self):
|
| 628 |
+
return (
|
| 629 |
+
f"{', '.join([str(c) for c in self._constraints[:-1]])} or"
|
| 630 |
+
f" {self._constraints[-1]}"
|
| 631 |
+
)
|
| 632 |
+
|
| 633 |
+
|
| 634 |
+
class HasMethods(_Constraint):
|
| 635 |
+
"""Constraint representing objects that expose specific methods.
|
| 636 |
+
|
| 637 |
+
It is useful for parameters following a protocol and where we don't want to impose
|
| 638 |
+
an affiliation to a specific module or class.
|
| 639 |
+
|
| 640 |
+
Parameters
|
| 641 |
+
----------
|
| 642 |
+
methods : str or list of str
|
| 643 |
+
The method(s) that the object is expected to expose.
|
| 644 |
+
"""
|
| 645 |
+
|
| 646 |
+
@validate_params({"methods": [str, list]})
|
| 647 |
+
def __init__(self, methods):
|
| 648 |
+
super().__init__()
|
| 649 |
+
if isinstance(methods, str):
|
| 650 |
+
methods = [methods]
|
| 651 |
+
self.methods = methods
|
| 652 |
+
|
| 653 |
+
def is_satisfied_by(self, val):
|
| 654 |
+
return all(callable(getattr(val, method, None)) for method in self.methods)
|
| 655 |
+
|
| 656 |
+
def __str__(self):
|
| 657 |
+
if len(self.methods) == 1:
|
| 658 |
+
methods = f"{self.methods[0]!r}"
|
| 659 |
+
else:
|
| 660 |
+
methods = (
|
| 661 |
+
f"{', '.join([repr(m) for m in self.methods[:-1]])} and"
|
| 662 |
+
f" {self.methods[-1]!r}"
|
| 663 |
+
)
|
| 664 |
+
return f"an object implementing {methods}"
|
| 665 |
+
|
| 666 |
+
|
| 667 |
+
class _IterablesNotString(_Constraint):
|
| 668 |
+
"""Constraint representing iterables that are not strings."""
|
| 669 |
+
|
| 670 |
+
def is_satisfied_by(self, val):
|
| 671 |
+
return isinstance(val, Iterable) and not isinstance(val, str)
|
| 672 |
+
|
| 673 |
+
def __str__(self):
|
| 674 |
+
return "an iterable"
|
| 675 |
+
|
| 676 |
+
|
| 677 |
+
class _CVObjects(_Constraint):
|
| 678 |
+
"""Constraint representing cv objects.
|
| 679 |
+
|
| 680 |
+
Convenient class for
|
| 681 |
+
[
|
| 682 |
+
Interval(Integral, 2, None, closed="left"),
|
| 683 |
+
HasMethods(["split", "get_n_splits"]),
|
| 684 |
+
_IterablesNotString(),
|
| 685 |
+
None,
|
| 686 |
+
]
|
| 687 |
+
"""
|
| 688 |
+
|
| 689 |
+
def __init__(self):
|
| 690 |
+
super().__init__()
|
| 691 |
+
self._constraints = [
|
| 692 |
+
Interval(Integral, 2, None, closed="left"),
|
| 693 |
+
HasMethods(["split", "get_n_splits"]),
|
| 694 |
+
_IterablesNotString(),
|
| 695 |
+
_NoneConstraint(),
|
| 696 |
+
]
|
| 697 |
+
|
| 698 |
+
def is_satisfied_by(self, val):
|
| 699 |
+
return any(c.is_satisfied_by(val) for c in self._constraints)
|
| 700 |
+
|
| 701 |
+
def __str__(self):
|
| 702 |
+
return (
|
| 703 |
+
f"{', '.join([str(c) for c in self._constraints[:-1]])} or"
|
| 704 |
+
f" {self._constraints[-1]}"
|
| 705 |
+
)
|
| 706 |
+
|
| 707 |
+
|
| 708 |
+
class Hidden:
|
| 709 |
+
"""Class encapsulating a constraint not meant to be exposed to the user.
|
| 710 |
+
|
| 711 |
+
Parameters
|
| 712 |
+
----------
|
| 713 |
+
constraint : str or _Constraint instance
|
| 714 |
+
The constraint to be used internally.
|
| 715 |
+
"""
|
| 716 |
+
|
| 717 |
+
def __init__(self, constraint):
|
| 718 |
+
self.constraint = constraint
|
| 719 |
+
|
| 720 |
+
|
| 721 |
+
def generate_invalid_param_val(constraint, constraints=None):
|
| 722 |
+
"""Return a value that does not satisfy the constraint.
|
| 723 |
+
|
| 724 |
+
Raises a NotImplementedError if there exists no invalid value for this constraint.
|
| 725 |
+
|
| 726 |
+
This is only useful for testing purpose.
|
| 727 |
+
|
| 728 |
+
Parameters
|
| 729 |
+
----------
|
| 730 |
+
constraint : _Constraint instance
|
| 731 |
+
The constraint to generate a value for.
|
| 732 |
+
|
| 733 |
+
constraints : list of _Constraint instances or None, default=None
|
| 734 |
+
The list of all constraints for this parameter. If None, the list only
|
| 735 |
+
containing `constraint` is used.
|
| 736 |
+
|
| 737 |
+
Returns
|
| 738 |
+
-------
|
| 739 |
+
val : object
|
| 740 |
+
A value that does not satisfy the constraint.
|
| 741 |
+
"""
|
| 742 |
+
if isinstance(constraint, StrOptions):
|
| 743 |
+
return f"not {' or '.join(constraint.options)}"
|
| 744 |
+
|
| 745 |
+
if isinstance(constraint, _MissingValues):
|
| 746 |
+
return np.array([1, 2, 3])
|
| 747 |
+
|
| 748 |
+
if isinstance(constraint, _VerboseHelper):
|
| 749 |
+
return -1
|
| 750 |
+
|
| 751 |
+
if isinstance(constraint, HasMethods):
|
| 752 |
+
return type("HasNotMethods", (), {})()
|
| 753 |
+
|
| 754 |
+
if isinstance(constraint, _IterablesNotString):
|
| 755 |
+
return "a string"
|
| 756 |
+
|
| 757 |
+
if isinstance(constraint, _CVObjects):
|
| 758 |
+
return "not a cv object"
|
| 759 |
+
|
| 760 |
+
if not isinstance(constraint, Interval):
|
| 761 |
+
raise NotImplementedError
|
| 762 |
+
|
| 763 |
+
# constraint is an interval
|
| 764 |
+
constraints = [constraint] if constraints is None else constraints
|
| 765 |
+
return _generate_invalid_param_val_interval(constraint, constraints)
|
| 766 |
+
|
| 767 |
+
|
| 768 |
+
def _generate_invalid_param_val_interval(interval, constraints):
|
| 769 |
+
"""Return a value that does not satisfy an interval constraint.
|
| 770 |
+
|
| 771 |
+
Generating an invalid value for an integer interval depends on the other constraints
|
| 772 |
+
since an int is a real, meaning that it can be valid for a real interval.
|
| 773 |
+
Assumes that there can be at most 2 interval constraints: one integer interval
|
| 774 |
+
and/or one real interval.
|
| 775 |
+
|
| 776 |
+
This is only useful for testing purpose.
|
| 777 |
+
|
| 778 |
+
Parameters
|
| 779 |
+
----------
|
| 780 |
+
interval : Interval instance
|
| 781 |
+
The interval to generate a value for.
|
| 782 |
+
|
| 783 |
+
constraints : list of _Constraint instances
|
| 784 |
+
The list of all constraints for this parameter.
|
| 785 |
+
|
| 786 |
+
Returns
|
| 787 |
+
-------
|
| 788 |
+
val : object
|
| 789 |
+
A value that does not satisfy the interval constraint.
|
| 790 |
+
"""
|
| 791 |
+
if interval.type is Real:
|
| 792 |
+
# generate a non-integer value such that it can't be valid even if there's also
|
| 793 |
+
# an integer interval constraint.
|
| 794 |
+
if interval.left is None and interval.right is None:
|
| 795 |
+
if interval.closed in ("left", "neither"):
|
| 796 |
+
return np.inf
|
| 797 |
+
elif interval.closed in ("right", "neither"):
|
| 798 |
+
return -np.inf
|
| 799 |
+
else:
|
| 800 |
+
raise NotImplementedError
|
| 801 |
+
|
| 802 |
+
if interval.left is not None:
|
| 803 |
+
return np.floor(interval.left) - 0.5
|
| 804 |
+
else: # right is not None
|
| 805 |
+
return np.ceil(interval.right) + 0.5
|
| 806 |
+
|
| 807 |
+
else: # interval.type is Integral
|
| 808 |
+
if interval.left is None and interval.right is None:
|
| 809 |
+
raise NotImplementedError
|
| 810 |
+
|
| 811 |
+
# We need to check if there's also a real interval constraint to generate a
|
| 812 |
+
# value that is not valid for any of the 2 interval constraints.
|
| 813 |
+
real_intervals = [
|
| 814 |
+
i for i in constraints if isinstance(i, Interval) and i.type is Real
|
| 815 |
+
]
|
| 816 |
+
real_interval = real_intervals[0] if real_intervals else None
|
| 817 |
+
|
| 818 |
+
if real_interval is None:
|
| 819 |
+
# Only the integer interval constraint -> easy
|
| 820 |
+
if interval.left is not None:
|
| 821 |
+
return interval.left - 1
|
| 822 |
+
else: # interval.right is not None
|
| 823 |
+
return interval.right + 1
|
| 824 |
+
|
| 825 |
+
# There's also a real interval constraint. Try to find a value left to both or
|
| 826 |
+
# right to both or in between them.
|
| 827 |
+
|
| 828 |
+
# redefine left and right bounds to be smallest and largest valid integers in
|
| 829 |
+
# both intervals.
|
| 830 |
+
int_left = interval.left
|
| 831 |
+
if int_left is not None and interval.closed in ("right", "neither"):
|
| 832 |
+
int_left = int_left + 1
|
| 833 |
+
|
| 834 |
+
int_right = interval.right
|
| 835 |
+
if int_right is not None and interval.closed in ("left", "neither"):
|
| 836 |
+
int_right = int_right - 1
|
| 837 |
+
|
| 838 |
+
real_left = real_interval.left
|
| 839 |
+
if real_interval.left is not None:
|
| 840 |
+
real_left = int(np.ceil(real_interval.left))
|
| 841 |
+
if real_interval.closed in ("right", "neither"):
|
| 842 |
+
real_left = real_left + 1
|
| 843 |
+
|
| 844 |
+
real_right = real_interval.right
|
| 845 |
+
if real_interval.right is not None:
|
| 846 |
+
real_right = int(np.floor(real_interval.right))
|
| 847 |
+
if real_interval.closed in ("left", "neither"):
|
| 848 |
+
real_right = real_right - 1
|
| 849 |
+
|
| 850 |
+
if int_left is not None and real_left is not None:
|
| 851 |
+
# there exists an int left to both intervals
|
| 852 |
+
return min(int_left, real_left) - 1
|
| 853 |
+
|
| 854 |
+
if int_right is not None and real_right is not None:
|
| 855 |
+
# there exists an int right to both intervals
|
| 856 |
+
return max(int_right, real_right) + 1
|
| 857 |
+
|
| 858 |
+
if int_left is not None:
|
| 859 |
+
if real_right is not None and int_left - real_right >= 2:
|
| 860 |
+
# there exists an int between the 2 intervals
|
| 861 |
+
return int_left - 1
|
| 862 |
+
else:
|
| 863 |
+
raise NotImplementedError
|
| 864 |
+
else: # int_right is not None
|
| 865 |
+
if real_left is not None and real_left - int_right >= 2:
|
| 866 |
+
# there exists an int between the 2 intervals
|
| 867 |
+
return int_right + 1
|
| 868 |
+
else:
|
| 869 |
+
raise NotImplementedError
|
| 870 |
+
|
| 871 |
+
|
| 872 |
+
def generate_valid_param(constraint):
|
| 873 |
+
"""Return a value that does satisfy a constraint.
|
| 874 |
+
|
| 875 |
+
This is only useful for testing purpose.
|
| 876 |
+
|
| 877 |
+
Parameters
|
| 878 |
+
----------
|
| 879 |
+
constraint : Constraint instance
|
| 880 |
+
The constraint to generate a value for.
|
| 881 |
+
|
| 882 |
+
Returns
|
| 883 |
+
-------
|
| 884 |
+
val : object
|
| 885 |
+
A value that does satisfy the constraint.
|
| 886 |
+
"""
|
| 887 |
+
if isinstance(constraint, _ArrayLikes):
|
| 888 |
+
return np.array([1, 2, 3])
|
| 889 |
+
|
| 890 |
+
if isinstance(constraint, _SparseMatrices):
|
| 891 |
+
return csr_matrix([[0, 1], [1, 0]])
|
| 892 |
+
|
| 893 |
+
if isinstance(constraint, _RandomStates):
|
| 894 |
+
return np.random.RandomState(42)
|
| 895 |
+
|
| 896 |
+
if isinstance(constraint, _Callables):
|
| 897 |
+
return lambda x: x
|
| 898 |
+
|
| 899 |
+
if isinstance(constraint, _NoneConstraint):
|
| 900 |
+
return None
|
| 901 |
+
|
| 902 |
+
if isinstance(constraint, _InstancesOf):
|
| 903 |
+
return constraint.type()
|
| 904 |
+
|
| 905 |
+
if isinstance(constraint, _Booleans):
|
| 906 |
+
return True
|
| 907 |
+
|
| 908 |
+
if isinstance(constraint, _VerboseHelper):
|
| 909 |
+
return 1
|
| 910 |
+
|
| 911 |
+
if isinstance(constraint, _MissingValues):
|
| 912 |
+
return np.nan
|
| 913 |
+
|
| 914 |
+
if isinstance(constraint, HasMethods):
|
| 915 |
+
return type(
|
| 916 |
+
"ValidHasMethods", (), {m: lambda self: None for m in constraint.methods}
|
| 917 |
+
)()
|
| 918 |
+
|
| 919 |
+
if isinstance(constraint, _IterablesNotString):
|
| 920 |
+
return [1, 2, 3]
|
| 921 |
+
|
| 922 |
+
if isinstance(constraint, _CVObjects):
|
| 923 |
+
return 5
|
| 924 |
+
|
| 925 |
+
if isinstance(constraint, Options): # includes StrOptions
|
| 926 |
+
for option in constraint.options:
|
| 927 |
+
return option
|
| 928 |
+
|
| 929 |
+
if isinstance(constraint, Interval):
|
| 930 |
+
interval = constraint
|
| 931 |
+
if interval.left is None and interval.right is None:
|
| 932 |
+
return 0
|
| 933 |
+
elif interval.left is None:
|
| 934 |
+
return interval.right - 1
|
| 935 |
+
elif interval.right is None:
|
| 936 |
+
return interval.left + 1
|
| 937 |
+
else:
|
| 938 |
+
if interval.type is Real:
|
| 939 |
+
return (interval.left + interval.right) / 2
|
| 940 |
+
else:
|
| 941 |
+
return interval.left + 1
|
| 942 |
+
|
| 943 |
+
raise ValueError(f"Unknown constraint type: {constraint}")
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_random.cpython-310-x86_64-linux-gnu.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a9c4f5f42a71d5792e4b1f8abd4f21f4bec97f97760346667faa773f35b1070a
|
| 3 |
+
size 109121
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_readonly_array_wrapper.cpython-310-x86_64-linux-gnu.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:de92483727ba785a6a37bf6d157763cf5e743f632437bcbf25d8a5d81448a8f4
|
| 3 |
+
size 237217
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_seq_dataset.cpython-310-x86_64-linux-gnu.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:41af9dbb90669834261523880b5bbb692b99d301148fa010b3eed979e8c99303
|
| 3 |
+
size 158369
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_show_versions.py
ADDED
|
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Utility methods to print system info for debugging
|
| 3 |
+
|
| 4 |
+
adapted from :func:`pandas.show_versions`
|
| 5 |
+
"""
|
| 6 |
+
# License: BSD 3 clause
|
| 7 |
+
|
| 8 |
+
import platform
|
| 9 |
+
import sys
|
| 10 |
+
from ..utils.fixes import threadpool_info
|
| 11 |
+
from .. import __version__
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
from ._openmp_helpers import _openmp_parallelism_enabled
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
def _get_sys_info():
|
| 18 |
+
"""System information
|
| 19 |
+
|
| 20 |
+
Returns
|
| 21 |
+
-------
|
| 22 |
+
sys_info : dict
|
| 23 |
+
system and Python version information
|
| 24 |
+
|
| 25 |
+
"""
|
| 26 |
+
python = sys.version.replace("\n", " ")
|
| 27 |
+
|
| 28 |
+
blob = [
|
| 29 |
+
("python", python),
|
| 30 |
+
("executable", sys.executable),
|
| 31 |
+
("machine", platform.platform()),
|
| 32 |
+
]
|
| 33 |
+
|
| 34 |
+
return dict(blob)
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def _get_deps_info():
|
| 38 |
+
"""Overview of the installed version of main dependencies
|
| 39 |
+
|
| 40 |
+
This function does not import the modules to collect the version numbers
|
| 41 |
+
but instead relies on standard Python package metadata.
|
| 42 |
+
|
| 43 |
+
Returns
|
| 44 |
+
-------
|
| 45 |
+
deps_info: dict
|
| 46 |
+
version information on relevant Python libraries
|
| 47 |
+
|
| 48 |
+
"""
|
| 49 |
+
deps = [
|
| 50 |
+
"pip",
|
| 51 |
+
"setuptools",
|
| 52 |
+
"numpy",
|
| 53 |
+
"scipy",
|
| 54 |
+
"Cython",
|
| 55 |
+
"pandas",
|
| 56 |
+
"matplotlib",
|
| 57 |
+
"joblib",
|
| 58 |
+
"threadpoolctl",
|
| 59 |
+
]
|
| 60 |
+
|
| 61 |
+
deps_info = {
|
| 62 |
+
"sklearn": __version__,
|
| 63 |
+
}
|
| 64 |
+
|
| 65 |
+
from importlib.metadata import version, PackageNotFoundError
|
| 66 |
+
|
| 67 |
+
for modname in deps:
|
| 68 |
+
try:
|
| 69 |
+
deps_info[modname] = version(modname)
|
| 70 |
+
except PackageNotFoundError:
|
| 71 |
+
deps_info[modname] = None
|
| 72 |
+
return deps_info
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def show_versions():
|
| 76 |
+
"""Print useful debugging information"
|
| 77 |
+
|
| 78 |
+
.. versionadded:: 0.20
|
| 79 |
+
"""
|
| 80 |
+
|
| 81 |
+
sys_info = _get_sys_info()
|
| 82 |
+
deps_info = _get_deps_info()
|
| 83 |
+
|
| 84 |
+
print("\nSystem:")
|
| 85 |
+
for k, stat in sys_info.items():
|
| 86 |
+
print("{k:>10}: {stat}".format(k=k, stat=stat))
|
| 87 |
+
|
| 88 |
+
print("\nPython dependencies:")
|
| 89 |
+
for k, stat in deps_info.items():
|
| 90 |
+
print("{k:>13}: {stat}".format(k=k, stat=stat))
|
| 91 |
+
|
| 92 |
+
print(
|
| 93 |
+
"\n{k}: {stat}".format(
|
| 94 |
+
k="Built with OpenMP", stat=_openmp_parallelism_enabled()
|
| 95 |
+
)
|
| 96 |
+
)
|
| 97 |
+
|
| 98 |
+
# show threadpoolctl results
|
| 99 |
+
threadpool_results = threadpool_info()
|
| 100 |
+
if threadpool_results:
|
| 101 |
+
print()
|
| 102 |
+
print("threadpoolctl info:")
|
| 103 |
+
|
| 104 |
+
for i, result in enumerate(threadpool_results):
|
| 105 |
+
for key, val in result.items():
|
| 106 |
+
print(f"{key:>15}: {val}")
|
| 107 |
+
if i != len(threadpool_results) - 1:
|
| 108 |
+
print()
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_sorting.cpython-310-x86_64-linux-gnu.so
ADDED
|
Binary file (38.6 kB). View file
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_tags.py
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
|
| 3 |
+
_DEFAULT_TAGS = {
|
| 4 |
+
"non_deterministic": False,
|
| 5 |
+
"requires_positive_X": False,
|
| 6 |
+
"requires_positive_y": False,
|
| 7 |
+
"X_types": ["2darray"],
|
| 8 |
+
"poor_score": False,
|
| 9 |
+
"no_validation": False,
|
| 10 |
+
"multioutput": False,
|
| 11 |
+
"allow_nan": False,
|
| 12 |
+
"stateless": False,
|
| 13 |
+
"multilabel": False,
|
| 14 |
+
"_skip_test": False,
|
| 15 |
+
"_xfail_checks": False,
|
| 16 |
+
"multioutput_only": False,
|
| 17 |
+
"binary_only": False,
|
| 18 |
+
"requires_fit": True,
|
| 19 |
+
"preserves_dtype": [np.float64],
|
| 20 |
+
"requires_y": False,
|
| 21 |
+
"pairwise": False,
|
| 22 |
+
}
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
def _safe_tags(estimator, key=None):
|
| 26 |
+
"""Safely get estimator tags.
|
| 27 |
+
|
| 28 |
+
:class:`~sklearn.BaseEstimator` provides the estimator tags machinery.
|
| 29 |
+
However, if an estimator does not inherit from this base class, we should
|
| 30 |
+
fall-back to the default tags.
|
| 31 |
+
|
| 32 |
+
For scikit-learn built-in estimators, we should still rely on
|
| 33 |
+
`self._get_tags()`. `_safe_tags(est)` should be used when we are not sure
|
| 34 |
+
where `est` comes from: typically `_safe_tags(self.base_estimator)` where
|
| 35 |
+
`self` is a meta-estimator, or in the common checks.
|
| 36 |
+
|
| 37 |
+
Parameters
|
| 38 |
+
----------
|
| 39 |
+
estimator : estimator object
|
| 40 |
+
The estimator from which to get the tag.
|
| 41 |
+
|
| 42 |
+
key : str, default=None
|
| 43 |
+
Tag name to get. By default (`None`), all tags are returned.
|
| 44 |
+
|
| 45 |
+
Returns
|
| 46 |
+
-------
|
| 47 |
+
tags : dict or tag value
|
| 48 |
+
The estimator tags. A single value is returned if `key` is not None.
|
| 49 |
+
"""
|
| 50 |
+
if hasattr(estimator, "_get_tags"):
|
| 51 |
+
tags_provider = "_get_tags()"
|
| 52 |
+
tags = estimator._get_tags()
|
| 53 |
+
elif hasattr(estimator, "_more_tags"):
|
| 54 |
+
tags_provider = "_more_tags()"
|
| 55 |
+
tags = {**_DEFAULT_TAGS, **estimator._more_tags()}
|
| 56 |
+
else:
|
| 57 |
+
tags_provider = "_DEFAULT_TAGS"
|
| 58 |
+
tags = _DEFAULT_TAGS
|
| 59 |
+
|
| 60 |
+
if key is not None:
|
| 61 |
+
if key not in tags:
|
| 62 |
+
raise ValueError(
|
| 63 |
+
f"The key {key} is not defined in {tags_provider} for the "
|
| 64 |
+
f"class {estimator.__class__.__name__}."
|
| 65 |
+
)
|
| 66 |
+
return tags[key]
|
| 67 |
+
return tags
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_testing.py
ADDED
|
@@ -0,0 +1,1057 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Testing utilities."""
|
| 2 |
+
|
| 3 |
+
# Copyright (c) 2011, 2012
|
| 4 |
+
# Authors: Pietro Berkes,
|
| 5 |
+
# Andreas Muller
|
| 6 |
+
# Mathieu Blondel
|
| 7 |
+
# Olivier Grisel
|
| 8 |
+
# Arnaud Joly
|
| 9 |
+
# Denis Engemann
|
| 10 |
+
# Giorgio Patrini
|
| 11 |
+
# Thierry Guillemot
|
| 12 |
+
# License: BSD 3 clause
|
| 13 |
+
import os
|
| 14 |
+
import os.path as op
|
| 15 |
+
import inspect
|
| 16 |
+
import warnings
|
| 17 |
+
import sys
|
| 18 |
+
import functools
|
| 19 |
+
import tempfile
|
| 20 |
+
from subprocess import check_output, STDOUT, CalledProcessError
|
| 21 |
+
from subprocess import TimeoutExpired
|
| 22 |
+
import re
|
| 23 |
+
import contextlib
|
| 24 |
+
from collections.abc import Iterable
|
| 25 |
+
from collections.abc import Sequence
|
| 26 |
+
|
| 27 |
+
import scipy as sp
|
| 28 |
+
from functools import wraps
|
| 29 |
+
from inspect import signature
|
| 30 |
+
|
| 31 |
+
import shutil
|
| 32 |
+
import atexit
|
| 33 |
+
import unittest
|
| 34 |
+
from unittest import TestCase
|
| 35 |
+
|
| 36 |
+
# WindowsError only exist on Windows
|
| 37 |
+
try:
|
| 38 |
+
WindowsError # type: ignore
|
| 39 |
+
except NameError:
|
| 40 |
+
WindowsError = None
|
| 41 |
+
|
| 42 |
+
from numpy.testing import assert_allclose as np_assert_allclose
|
| 43 |
+
from numpy.testing import assert_almost_equal
|
| 44 |
+
from numpy.testing import assert_approx_equal
|
| 45 |
+
from numpy.testing import assert_array_equal
|
| 46 |
+
from numpy.testing import assert_array_almost_equal
|
| 47 |
+
from numpy.testing import assert_array_less
|
| 48 |
+
import numpy as np
|
| 49 |
+
import joblib
|
| 50 |
+
|
| 51 |
+
import sklearn
|
| 52 |
+
from sklearn.utils import (
|
| 53 |
+
IS_PYPY,
|
| 54 |
+
_IS_32BIT,
|
| 55 |
+
_in_unstable_openblas_configuration,
|
| 56 |
+
)
|
| 57 |
+
from sklearn.utils.multiclass import check_classification_targets
|
| 58 |
+
from sklearn.utils.validation import (
|
| 59 |
+
check_array,
|
| 60 |
+
check_is_fitted,
|
| 61 |
+
check_X_y,
|
| 62 |
+
)
|
| 63 |
+
from sklearn.utils.fixes import threadpool_info
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
__all__ = [
|
| 67 |
+
"assert_raises",
|
| 68 |
+
"assert_raises_regexp",
|
| 69 |
+
"assert_array_equal",
|
| 70 |
+
"assert_almost_equal",
|
| 71 |
+
"assert_array_almost_equal",
|
| 72 |
+
"assert_array_less",
|
| 73 |
+
"assert_approx_equal",
|
| 74 |
+
"assert_allclose",
|
| 75 |
+
"assert_run_python_script",
|
| 76 |
+
"SkipTest",
|
| 77 |
+
]
|
| 78 |
+
|
| 79 |
+
_dummy = TestCase("__init__")
|
| 80 |
+
assert_raises = _dummy.assertRaises
|
| 81 |
+
SkipTest = unittest.case.SkipTest
|
| 82 |
+
assert_dict_equal = _dummy.assertDictEqual
|
| 83 |
+
|
| 84 |
+
assert_raises_regex = _dummy.assertRaisesRegex
|
| 85 |
+
# assert_raises_regexp is deprecated in Python 3.4 in favor of
|
| 86 |
+
# assert_raises_regex but lets keep the backward compat in scikit-learn with
|
| 87 |
+
# the old name for now
|
| 88 |
+
assert_raises_regexp = assert_raises_regex
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
# To remove when we support numpy 1.7
|
| 92 |
+
def assert_no_warnings(func, *args, **kw):
|
| 93 |
+
"""
|
| 94 |
+
Parameters
|
| 95 |
+
----------
|
| 96 |
+
func
|
| 97 |
+
*args
|
| 98 |
+
**kw
|
| 99 |
+
"""
|
| 100 |
+
# very important to avoid uncontrolled state propagation
|
| 101 |
+
with warnings.catch_warnings(record=True) as w:
|
| 102 |
+
warnings.simplefilter("always")
|
| 103 |
+
|
| 104 |
+
result = func(*args, **kw)
|
| 105 |
+
if hasattr(np, "FutureWarning"):
|
| 106 |
+
# Filter out numpy-specific warnings in numpy >= 1.9
|
| 107 |
+
w = [e for e in w if e.category is not np.VisibleDeprecationWarning]
|
| 108 |
+
|
| 109 |
+
if len(w) > 0:
|
| 110 |
+
raise AssertionError(
|
| 111 |
+
"Got warnings when calling %s: [%s]"
|
| 112 |
+
% (func.__name__, ", ".join(str(warning) for warning in w))
|
| 113 |
+
)
|
| 114 |
+
return result
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
def ignore_warnings(obj=None, category=Warning):
|
| 118 |
+
"""Context manager and decorator to ignore warnings.
|
| 119 |
+
|
| 120 |
+
Note: Using this (in both variants) will clear all warnings
|
| 121 |
+
from all python modules loaded. In case you need to test
|
| 122 |
+
cross-module-warning-logging, this is not your tool of choice.
|
| 123 |
+
|
| 124 |
+
Parameters
|
| 125 |
+
----------
|
| 126 |
+
obj : callable, default=None
|
| 127 |
+
callable where you want to ignore the warnings.
|
| 128 |
+
category : warning class, default=Warning
|
| 129 |
+
The category to filter. If Warning, all categories will be muted.
|
| 130 |
+
|
| 131 |
+
Examples
|
| 132 |
+
--------
|
| 133 |
+
>>> import warnings
|
| 134 |
+
>>> from sklearn.utils._testing import ignore_warnings
|
| 135 |
+
>>> with ignore_warnings():
|
| 136 |
+
... warnings.warn('buhuhuhu')
|
| 137 |
+
|
| 138 |
+
>>> def nasty_warn():
|
| 139 |
+
... warnings.warn('buhuhuhu')
|
| 140 |
+
... print(42)
|
| 141 |
+
|
| 142 |
+
>>> ignore_warnings(nasty_warn)()
|
| 143 |
+
42
|
| 144 |
+
"""
|
| 145 |
+
if isinstance(obj, type) and issubclass(obj, Warning):
|
| 146 |
+
# Avoid common pitfall of passing category as the first positional
|
| 147 |
+
# argument which result in the test not being run
|
| 148 |
+
warning_name = obj.__name__
|
| 149 |
+
raise ValueError(
|
| 150 |
+
"'obj' should be a callable where you want to ignore warnings. "
|
| 151 |
+
"You passed a warning class instead: 'obj={warning_name}'. "
|
| 152 |
+
"If you want to pass a warning class to ignore_warnings, "
|
| 153 |
+
"you should use 'category={warning_name}'".format(warning_name=warning_name)
|
| 154 |
+
)
|
| 155 |
+
elif callable(obj):
|
| 156 |
+
return _IgnoreWarnings(category=category)(obj)
|
| 157 |
+
else:
|
| 158 |
+
return _IgnoreWarnings(category=category)
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
class _IgnoreWarnings:
|
| 162 |
+
"""Improved and simplified Python warnings context manager and decorator.
|
| 163 |
+
|
| 164 |
+
This class allows the user to ignore the warnings raised by a function.
|
| 165 |
+
Copied from Python 2.7.5 and modified as required.
|
| 166 |
+
|
| 167 |
+
Parameters
|
| 168 |
+
----------
|
| 169 |
+
category : tuple of warning class, default=Warning
|
| 170 |
+
The category to filter. By default, all the categories will be muted.
|
| 171 |
+
|
| 172 |
+
"""
|
| 173 |
+
|
| 174 |
+
def __init__(self, category):
|
| 175 |
+
self._record = True
|
| 176 |
+
self._module = sys.modules["warnings"]
|
| 177 |
+
self._entered = False
|
| 178 |
+
self.log = []
|
| 179 |
+
self.category = category
|
| 180 |
+
|
| 181 |
+
def __call__(self, fn):
|
| 182 |
+
"""Decorator to catch and hide warnings without visual nesting."""
|
| 183 |
+
|
| 184 |
+
@wraps(fn)
|
| 185 |
+
def wrapper(*args, **kwargs):
|
| 186 |
+
with warnings.catch_warnings():
|
| 187 |
+
warnings.simplefilter("ignore", self.category)
|
| 188 |
+
return fn(*args, **kwargs)
|
| 189 |
+
|
| 190 |
+
return wrapper
|
| 191 |
+
|
| 192 |
+
def __repr__(self):
|
| 193 |
+
args = []
|
| 194 |
+
if self._record:
|
| 195 |
+
args.append("record=True")
|
| 196 |
+
if self._module is not sys.modules["warnings"]:
|
| 197 |
+
args.append("module=%r" % self._module)
|
| 198 |
+
name = type(self).__name__
|
| 199 |
+
return "%s(%s)" % (name, ", ".join(args))
|
| 200 |
+
|
| 201 |
+
def __enter__(self):
|
| 202 |
+
if self._entered:
|
| 203 |
+
raise RuntimeError("Cannot enter %r twice" % self)
|
| 204 |
+
self._entered = True
|
| 205 |
+
self._filters = self._module.filters
|
| 206 |
+
self._module.filters = self._filters[:]
|
| 207 |
+
self._showwarning = self._module.showwarning
|
| 208 |
+
warnings.simplefilter("ignore", self.category)
|
| 209 |
+
|
| 210 |
+
def __exit__(self, *exc_info):
|
| 211 |
+
if not self._entered:
|
| 212 |
+
raise RuntimeError("Cannot exit %r without entering first" % self)
|
| 213 |
+
self._module.filters = self._filters
|
| 214 |
+
self._module.showwarning = self._showwarning
|
| 215 |
+
self.log[:] = []
|
| 216 |
+
|
| 217 |
+
|
| 218 |
+
def assert_raise_message(exceptions, message, function, *args, **kwargs):
|
| 219 |
+
"""Helper function to test the message raised in an exception.
|
| 220 |
+
|
| 221 |
+
Given an exception, a callable to raise the exception, and
|
| 222 |
+
a message string, tests that the correct exception is raised and
|
| 223 |
+
that the message is a substring of the error thrown. Used to test
|
| 224 |
+
that the specific message thrown during an exception is correct.
|
| 225 |
+
|
| 226 |
+
Parameters
|
| 227 |
+
----------
|
| 228 |
+
exceptions : exception or tuple of exception
|
| 229 |
+
An Exception object.
|
| 230 |
+
|
| 231 |
+
message : str
|
| 232 |
+
The error message or a substring of the error message.
|
| 233 |
+
|
| 234 |
+
function : callable
|
| 235 |
+
Callable object to raise error.
|
| 236 |
+
|
| 237 |
+
*args : the positional arguments to `function`.
|
| 238 |
+
|
| 239 |
+
**kwargs : the keyword arguments to `function`.
|
| 240 |
+
"""
|
| 241 |
+
try:
|
| 242 |
+
function(*args, **kwargs)
|
| 243 |
+
except exceptions as e:
|
| 244 |
+
error_message = str(e)
|
| 245 |
+
if message not in error_message:
|
| 246 |
+
raise AssertionError(
|
| 247 |
+
"Error message does not include the expected"
|
| 248 |
+
" string: %r. Observed error message: %r" % (message, error_message)
|
| 249 |
+
)
|
| 250 |
+
else:
|
| 251 |
+
# concatenate exception names
|
| 252 |
+
if isinstance(exceptions, tuple):
|
| 253 |
+
names = " or ".join(e.__name__ for e in exceptions)
|
| 254 |
+
else:
|
| 255 |
+
names = exceptions.__name__
|
| 256 |
+
|
| 257 |
+
raise AssertionError("%s not raised by %s" % (names, function.__name__))
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
def assert_allclose(
|
| 261 |
+
actual, desired, rtol=None, atol=0.0, equal_nan=True, err_msg="", verbose=True
|
| 262 |
+
):
|
| 263 |
+
"""dtype-aware variant of numpy.testing.assert_allclose
|
| 264 |
+
|
| 265 |
+
This variant introspects the least precise floating point dtype
|
| 266 |
+
in the input argument and automatically sets the relative tolerance
|
| 267 |
+
parameter to 1e-4 float32 and use 1e-7 otherwise (typically float64
|
| 268 |
+
in scikit-learn).
|
| 269 |
+
|
| 270 |
+
`atol` is always left to 0. by default. It should be adjusted manually
|
| 271 |
+
to an assertion-specific value in case there are null values expected
|
| 272 |
+
in `desired`.
|
| 273 |
+
|
| 274 |
+
The aggregate tolerance is `atol + rtol * abs(desired)`.
|
| 275 |
+
|
| 276 |
+
Parameters
|
| 277 |
+
----------
|
| 278 |
+
actual : array_like
|
| 279 |
+
Array obtained.
|
| 280 |
+
desired : array_like
|
| 281 |
+
Array desired.
|
| 282 |
+
rtol : float, optional, default=None
|
| 283 |
+
Relative tolerance.
|
| 284 |
+
If None, it is set based on the provided arrays' dtypes.
|
| 285 |
+
atol : float, optional, default=0.
|
| 286 |
+
Absolute tolerance.
|
| 287 |
+
equal_nan : bool, optional, default=True
|
| 288 |
+
If True, NaNs will compare equal.
|
| 289 |
+
err_msg : str, optional, default=''
|
| 290 |
+
The error message to be printed in case of failure.
|
| 291 |
+
verbose : bool, optional, default=True
|
| 292 |
+
If True, the conflicting values are appended to the error message.
|
| 293 |
+
|
| 294 |
+
Raises
|
| 295 |
+
------
|
| 296 |
+
AssertionError
|
| 297 |
+
If actual and desired are not equal up to specified precision.
|
| 298 |
+
|
| 299 |
+
See Also
|
| 300 |
+
--------
|
| 301 |
+
numpy.testing.assert_allclose
|
| 302 |
+
|
| 303 |
+
Examples
|
| 304 |
+
--------
|
| 305 |
+
>>> import numpy as np
|
| 306 |
+
>>> from sklearn.utils._testing import assert_allclose
|
| 307 |
+
>>> x = [1e-5, 1e-3, 1e-1]
|
| 308 |
+
>>> y = np.arccos(np.cos(x))
|
| 309 |
+
>>> assert_allclose(x, y, rtol=1e-5, atol=0)
|
| 310 |
+
>>> a = np.full(shape=10, fill_value=1e-5, dtype=np.float32)
|
| 311 |
+
>>> assert_allclose(a, 1e-5)
|
| 312 |
+
"""
|
| 313 |
+
dtypes = []
|
| 314 |
+
|
| 315 |
+
actual, desired = np.asanyarray(actual), np.asanyarray(desired)
|
| 316 |
+
dtypes = [actual.dtype, desired.dtype]
|
| 317 |
+
|
| 318 |
+
if rtol is None:
|
| 319 |
+
rtols = [1e-4 if dtype == np.float32 else 1e-7 for dtype in dtypes]
|
| 320 |
+
rtol = max(rtols)
|
| 321 |
+
|
| 322 |
+
np_assert_allclose(
|
| 323 |
+
actual,
|
| 324 |
+
desired,
|
| 325 |
+
rtol=rtol,
|
| 326 |
+
atol=atol,
|
| 327 |
+
equal_nan=equal_nan,
|
| 328 |
+
err_msg=err_msg,
|
| 329 |
+
verbose=verbose,
|
| 330 |
+
)
|
| 331 |
+
|
| 332 |
+
|
| 333 |
+
def assert_allclose_dense_sparse(x, y, rtol=1e-07, atol=1e-9, err_msg=""):
|
| 334 |
+
"""Assert allclose for sparse and dense data.
|
| 335 |
+
|
| 336 |
+
Both x and y need to be either sparse or dense, they
|
| 337 |
+
can't be mixed.
|
| 338 |
+
|
| 339 |
+
Parameters
|
| 340 |
+
----------
|
| 341 |
+
x : {array-like, sparse matrix}
|
| 342 |
+
First array to compare.
|
| 343 |
+
|
| 344 |
+
y : {array-like, sparse matrix}
|
| 345 |
+
Second array to compare.
|
| 346 |
+
|
| 347 |
+
rtol : float, default=1e-07
|
| 348 |
+
relative tolerance; see numpy.allclose.
|
| 349 |
+
|
| 350 |
+
atol : float, default=1e-9
|
| 351 |
+
absolute tolerance; see numpy.allclose. Note that the default here is
|
| 352 |
+
more tolerant than the default for numpy.testing.assert_allclose, where
|
| 353 |
+
atol=0.
|
| 354 |
+
|
| 355 |
+
err_msg : str, default=''
|
| 356 |
+
Error message to raise.
|
| 357 |
+
"""
|
| 358 |
+
if sp.sparse.issparse(x) and sp.sparse.issparse(y):
|
| 359 |
+
x = x.tocsr()
|
| 360 |
+
y = y.tocsr()
|
| 361 |
+
x.sum_duplicates()
|
| 362 |
+
y.sum_duplicates()
|
| 363 |
+
assert_array_equal(x.indices, y.indices, err_msg=err_msg)
|
| 364 |
+
assert_array_equal(x.indptr, y.indptr, err_msg=err_msg)
|
| 365 |
+
assert_allclose(x.data, y.data, rtol=rtol, atol=atol, err_msg=err_msg)
|
| 366 |
+
elif not sp.sparse.issparse(x) and not sp.sparse.issparse(y):
|
| 367 |
+
# both dense
|
| 368 |
+
assert_allclose(x, y, rtol=rtol, atol=atol, err_msg=err_msg)
|
| 369 |
+
else:
|
| 370 |
+
raise ValueError(
|
| 371 |
+
"Can only compare two sparse matrices, not a sparse matrix and an array."
|
| 372 |
+
)
|
| 373 |
+
|
| 374 |
+
|
| 375 |
+
def set_random_state(estimator, random_state=0):
|
| 376 |
+
"""Set random state of an estimator if it has the `random_state` param.
|
| 377 |
+
|
| 378 |
+
Parameters
|
| 379 |
+
----------
|
| 380 |
+
estimator : object
|
| 381 |
+
The estimator.
|
| 382 |
+
random_state : int, RandomState instance or None, default=0
|
| 383 |
+
Pseudo random number generator state.
|
| 384 |
+
Pass an int for reproducible results across multiple function calls.
|
| 385 |
+
See :term:`Glossary <random_state>`.
|
| 386 |
+
"""
|
| 387 |
+
if "random_state" in estimator.get_params():
|
| 388 |
+
estimator.set_params(random_state=random_state)
|
| 389 |
+
|
| 390 |
+
|
| 391 |
+
try:
|
| 392 |
+
import pytest
|
| 393 |
+
|
| 394 |
+
skip_if_32bit = pytest.mark.skipif(_IS_32BIT, reason="skipped on 32bit platforms")
|
| 395 |
+
fails_if_pypy = pytest.mark.xfail(IS_PYPY, reason="not compatible with PyPy")
|
| 396 |
+
fails_if_unstable_openblas = pytest.mark.xfail(
|
| 397 |
+
_in_unstable_openblas_configuration(),
|
| 398 |
+
reason="OpenBLAS is unstable for this configuration",
|
| 399 |
+
)
|
| 400 |
+
skip_if_no_parallel = pytest.mark.skipif(
|
| 401 |
+
not joblib.parallel.mp, reason="joblib is in serial mode"
|
| 402 |
+
)
|
| 403 |
+
|
| 404 |
+
# Decorator for tests involving both BLAS calls and multiprocessing.
|
| 405 |
+
#
|
| 406 |
+
# Under POSIX (e.g. Linux or OSX), using multiprocessing in conjunction
|
| 407 |
+
# with some implementation of BLAS (or other libraries that manage an
|
| 408 |
+
# internal posix thread pool) can cause a crash or a freeze of the Python
|
| 409 |
+
# process.
|
| 410 |
+
#
|
| 411 |
+
# In practice all known packaged distributions (from Linux distros or
|
| 412 |
+
# Anaconda) of BLAS under Linux seems to be safe. So we this problem seems
|
| 413 |
+
# to only impact OSX users.
|
| 414 |
+
#
|
| 415 |
+
# This wrapper makes it possible to skip tests that can possibly cause
|
| 416 |
+
# this crash under OS X with.
|
| 417 |
+
#
|
| 418 |
+
# Under Python 3.4+ it is possible to use the `forkserver` start method
|
| 419 |
+
# for multiprocessing to avoid this issue. However it can cause pickling
|
| 420 |
+
# errors on interactively defined functions. It therefore not enabled by
|
| 421 |
+
# default.
|
| 422 |
+
|
| 423 |
+
if_safe_multiprocessing_with_blas = pytest.mark.skipif(
|
| 424 |
+
sys.platform == "darwin", reason="Possible multi-process bug with some BLAS"
|
| 425 |
+
)
|
| 426 |
+
except ImportError:
|
| 427 |
+
pass
|
| 428 |
+
|
| 429 |
+
|
| 430 |
+
def check_skip_network():
|
| 431 |
+
if int(os.environ.get("SKLEARN_SKIP_NETWORK_TESTS", 0)):
|
| 432 |
+
raise SkipTest("Text tutorial requires large dataset download")
|
| 433 |
+
|
| 434 |
+
|
| 435 |
+
def _delete_folder(folder_path, warn=False):
|
| 436 |
+
"""Utility function to cleanup a temporary folder if still existing.
|
| 437 |
+
|
| 438 |
+
Copy from joblib.pool (for independence).
|
| 439 |
+
"""
|
| 440 |
+
try:
|
| 441 |
+
if os.path.exists(folder_path):
|
| 442 |
+
# This can fail under windows,
|
| 443 |
+
# but will succeed when called by atexit
|
| 444 |
+
shutil.rmtree(folder_path)
|
| 445 |
+
except WindowsError:
|
| 446 |
+
if warn:
|
| 447 |
+
warnings.warn("Could not delete temporary folder %s" % folder_path)
|
| 448 |
+
|
| 449 |
+
|
| 450 |
+
class TempMemmap:
|
| 451 |
+
"""
|
| 452 |
+
Parameters
|
| 453 |
+
----------
|
| 454 |
+
data
|
| 455 |
+
mmap_mode : str, default='r'
|
| 456 |
+
"""
|
| 457 |
+
|
| 458 |
+
def __init__(self, data, mmap_mode="r"):
|
| 459 |
+
self.mmap_mode = mmap_mode
|
| 460 |
+
self.data = data
|
| 461 |
+
|
| 462 |
+
def __enter__(self):
|
| 463 |
+
data_read_only, self.temp_folder = create_memmap_backed_data(
|
| 464 |
+
self.data, mmap_mode=self.mmap_mode, return_folder=True
|
| 465 |
+
)
|
| 466 |
+
return data_read_only
|
| 467 |
+
|
| 468 |
+
def __exit__(self, exc_type, exc_val, exc_tb):
|
| 469 |
+
_delete_folder(self.temp_folder)
|
| 470 |
+
|
| 471 |
+
|
| 472 |
+
def _create_memmap_backed_array(array, filename, mmap_mode):
|
| 473 |
+
# https://numpy.org/doc/stable/reference/generated/numpy.memmap.html
|
| 474 |
+
fp = np.memmap(filename, dtype=array.dtype, mode="w+", shape=array.shape)
|
| 475 |
+
fp[:] = array[:] # write array to memmap array
|
| 476 |
+
fp.flush()
|
| 477 |
+
memmap_backed_array = np.memmap(
|
| 478 |
+
filename, dtype=array.dtype, mode=mmap_mode, shape=array.shape
|
| 479 |
+
)
|
| 480 |
+
return memmap_backed_array
|
| 481 |
+
|
| 482 |
+
|
| 483 |
+
def _create_aligned_memmap_backed_arrays(data, mmap_mode, folder):
|
| 484 |
+
if isinstance(data, np.ndarray):
|
| 485 |
+
filename = op.join(folder, "data.dat")
|
| 486 |
+
return _create_memmap_backed_array(data, filename, mmap_mode)
|
| 487 |
+
|
| 488 |
+
if isinstance(data, Sequence) and all(
|
| 489 |
+
isinstance(each, np.ndarray) for each in data
|
| 490 |
+
):
|
| 491 |
+
return [
|
| 492 |
+
_create_memmap_backed_array(
|
| 493 |
+
array, op.join(folder, f"data{index}.dat"), mmap_mode
|
| 494 |
+
)
|
| 495 |
+
for index, array in enumerate(data)
|
| 496 |
+
]
|
| 497 |
+
|
| 498 |
+
raise ValueError(
|
| 499 |
+
"When creating aligned memmap-backed arrays, input must be a single array or a"
|
| 500 |
+
" sequence of arrays"
|
| 501 |
+
)
|
| 502 |
+
|
| 503 |
+
|
| 504 |
+
def create_memmap_backed_data(data, mmap_mode="r", return_folder=False, aligned=False):
|
| 505 |
+
"""
|
| 506 |
+
Parameters
|
| 507 |
+
----------
|
| 508 |
+
data
|
| 509 |
+
mmap_mode : str, default='r'
|
| 510 |
+
return_folder : bool, default=False
|
| 511 |
+
aligned : bool, default=False
|
| 512 |
+
If True, if input is a single numpy array and if the input array is aligned,
|
| 513 |
+
the memory mapped array will also be aligned. This is a workaround for
|
| 514 |
+
https://github.com/joblib/joblib/issues/563.
|
| 515 |
+
"""
|
| 516 |
+
temp_folder = tempfile.mkdtemp(prefix="sklearn_testing_")
|
| 517 |
+
atexit.register(functools.partial(_delete_folder, temp_folder, warn=True))
|
| 518 |
+
# OpenBLAS is known to segfault with unaligned data on the Prescott
|
| 519 |
+
# architecture so force aligned=True on Prescott. For more details, see:
|
| 520 |
+
# https://github.com/scipy/scipy/issues/14886
|
| 521 |
+
has_prescott_openblas = any(
|
| 522 |
+
True
|
| 523 |
+
for info in threadpool_info()
|
| 524 |
+
if info["internal_api"] == "openblas"
|
| 525 |
+
# Prudently assume Prescott might be the architecture if it is unknown.
|
| 526 |
+
and info.get("architecture", "prescott").lower() == "prescott"
|
| 527 |
+
)
|
| 528 |
+
if has_prescott_openblas:
|
| 529 |
+
aligned = True
|
| 530 |
+
|
| 531 |
+
if aligned:
|
| 532 |
+
memmap_backed_data = _create_aligned_memmap_backed_arrays(
|
| 533 |
+
data, mmap_mode, temp_folder
|
| 534 |
+
)
|
| 535 |
+
else:
|
| 536 |
+
filename = op.join(temp_folder, "data.pkl")
|
| 537 |
+
joblib.dump(data, filename)
|
| 538 |
+
memmap_backed_data = joblib.load(filename, mmap_mode=mmap_mode)
|
| 539 |
+
result = (
|
| 540 |
+
memmap_backed_data if not return_folder else (memmap_backed_data, temp_folder)
|
| 541 |
+
)
|
| 542 |
+
return result
|
| 543 |
+
|
| 544 |
+
|
| 545 |
+
# Utils to test docstrings
|
| 546 |
+
|
| 547 |
+
|
| 548 |
+
def _get_args(function, varargs=False):
|
| 549 |
+
"""Helper to get function arguments."""
|
| 550 |
+
|
| 551 |
+
try:
|
| 552 |
+
params = signature(function).parameters
|
| 553 |
+
except ValueError:
|
| 554 |
+
# Error on builtin C function
|
| 555 |
+
return []
|
| 556 |
+
args = [
|
| 557 |
+
key
|
| 558 |
+
for key, param in params.items()
|
| 559 |
+
if param.kind not in (param.VAR_POSITIONAL, param.VAR_KEYWORD)
|
| 560 |
+
]
|
| 561 |
+
if varargs:
|
| 562 |
+
varargs = [
|
| 563 |
+
param.name
|
| 564 |
+
for param in params.values()
|
| 565 |
+
if param.kind == param.VAR_POSITIONAL
|
| 566 |
+
]
|
| 567 |
+
if len(varargs) == 0:
|
| 568 |
+
varargs = None
|
| 569 |
+
return args, varargs
|
| 570 |
+
else:
|
| 571 |
+
return args
|
| 572 |
+
|
| 573 |
+
|
| 574 |
+
def _get_func_name(func):
|
| 575 |
+
"""Get function full name.
|
| 576 |
+
|
| 577 |
+
Parameters
|
| 578 |
+
----------
|
| 579 |
+
func : callable
|
| 580 |
+
The function object.
|
| 581 |
+
|
| 582 |
+
Returns
|
| 583 |
+
-------
|
| 584 |
+
name : str
|
| 585 |
+
The function name.
|
| 586 |
+
"""
|
| 587 |
+
parts = []
|
| 588 |
+
module = inspect.getmodule(func)
|
| 589 |
+
if module:
|
| 590 |
+
parts.append(module.__name__)
|
| 591 |
+
|
| 592 |
+
qualname = func.__qualname__
|
| 593 |
+
if qualname != func.__name__:
|
| 594 |
+
parts.append(qualname[: qualname.find(".")])
|
| 595 |
+
|
| 596 |
+
parts.append(func.__name__)
|
| 597 |
+
return ".".join(parts)
|
| 598 |
+
|
| 599 |
+
|
| 600 |
+
def check_docstring_parameters(func, doc=None, ignore=None):
|
| 601 |
+
"""Helper to check docstring.
|
| 602 |
+
|
| 603 |
+
Parameters
|
| 604 |
+
----------
|
| 605 |
+
func : callable
|
| 606 |
+
The function object to test.
|
| 607 |
+
doc : str, default=None
|
| 608 |
+
Docstring if it is passed manually to the test.
|
| 609 |
+
ignore : list, default=None
|
| 610 |
+
Parameters to ignore.
|
| 611 |
+
|
| 612 |
+
Returns
|
| 613 |
+
-------
|
| 614 |
+
incorrect : list
|
| 615 |
+
A list of string describing the incorrect results.
|
| 616 |
+
"""
|
| 617 |
+
from numpydoc import docscrape
|
| 618 |
+
|
| 619 |
+
incorrect = []
|
| 620 |
+
ignore = [] if ignore is None else ignore
|
| 621 |
+
|
| 622 |
+
func_name = _get_func_name(func)
|
| 623 |
+
if not func_name.startswith("sklearn.") or func_name.startswith(
|
| 624 |
+
"sklearn.externals"
|
| 625 |
+
):
|
| 626 |
+
return incorrect
|
| 627 |
+
# Don't check docstring for property-functions
|
| 628 |
+
if inspect.isdatadescriptor(func):
|
| 629 |
+
return incorrect
|
| 630 |
+
# Don't check docstring for setup / teardown pytest functions
|
| 631 |
+
if func_name.split(".")[-1] in ("setup_module", "teardown_module"):
|
| 632 |
+
return incorrect
|
| 633 |
+
# Dont check estimator_checks module
|
| 634 |
+
if func_name.split(".")[2] == "estimator_checks":
|
| 635 |
+
return incorrect
|
| 636 |
+
# Get the arguments from the function signature
|
| 637 |
+
param_signature = list(filter(lambda x: x not in ignore, _get_args(func)))
|
| 638 |
+
# drop self
|
| 639 |
+
if len(param_signature) > 0 and param_signature[0] == "self":
|
| 640 |
+
param_signature.remove("self")
|
| 641 |
+
|
| 642 |
+
# Analyze function's docstring
|
| 643 |
+
if doc is None:
|
| 644 |
+
records = []
|
| 645 |
+
with warnings.catch_warnings(record=True):
|
| 646 |
+
warnings.simplefilter("error", UserWarning)
|
| 647 |
+
try:
|
| 648 |
+
doc = docscrape.FunctionDoc(func)
|
| 649 |
+
except UserWarning as exp:
|
| 650 |
+
if "potentially wrong underline length" in str(exp):
|
| 651 |
+
# Catch warning raised as of numpydoc 1.2 when
|
| 652 |
+
# the underline length for a section of a docstring
|
| 653 |
+
# is not consistent.
|
| 654 |
+
message = str(exp).split("\n")[:3]
|
| 655 |
+
incorrect += [f"In function: {func_name}"] + message
|
| 656 |
+
return incorrect
|
| 657 |
+
records.append(str(exp))
|
| 658 |
+
except Exception as exp:
|
| 659 |
+
incorrect += [func_name + " parsing error: " + str(exp)]
|
| 660 |
+
return incorrect
|
| 661 |
+
if len(records):
|
| 662 |
+
raise RuntimeError("Error for %s:\n%s" % (func_name, records[0]))
|
| 663 |
+
|
| 664 |
+
param_docs = []
|
| 665 |
+
for name, type_definition, param_doc in doc["Parameters"]:
|
| 666 |
+
# Type hints are empty only if parameter name ended with :
|
| 667 |
+
if not type_definition.strip():
|
| 668 |
+
if ":" in name and name[: name.index(":")][-1:].strip():
|
| 669 |
+
incorrect += [
|
| 670 |
+
func_name
|
| 671 |
+
+ " There was no space between the param name and colon (%r)" % name
|
| 672 |
+
]
|
| 673 |
+
elif name.rstrip().endswith(":"):
|
| 674 |
+
incorrect += [
|
| 675 |
+
func_name
|
| 676 |
+
+ " Parameter %r has an empty type spec. Remove the colon"
|
| 677 |
+
% (name.lstrip())
|
| 678 |
+
]
|
| 679 |
+
|
| 680 |
+
# Create a list of parameters to compare with the parameters gotten
|
| 681 |
+
# from the func signature
|
| 682 |
+
if "*" not in name:
|
| 683 |
+
param_docs.append(name.split(":")[0].strip("` "))
|
| 684 |
+
|
| 685 |
+
# If one of the docstring's parameters had an error then return that
|
| 686 |
+
# incorrect message
|
| 687 |
+
if len(incorrect) > 0:
|
| 688 |
+
return incorrect
|
| 689 |
+
|
| 690 |
+
# Remove the parameters that should be ignored from list
|
| 691 |
+
param_docs = list(filter(lambda x: x not in ignore, param_docs))
|
| 692 |
+
|
| 693 |
+
# The following is derived from pytest, Copyright (c) 2004-2017 Holger
|
| 694 |
+
# Krekel and others, Licensed under MIT License. See
|
| 695 |
+
# https://github.com/pytest-dev/pytest
|
| 696 |
+
|
| 697 |
+
message = []
|
| 698 |
+
for i in range(min(len(param_docs), len(param_signature))):
|
| 699 |
+
if param_signature[i] != param_docs[i]:
|
| 700 |
+
message += [
|
| 701 |
+
"There's a parameter name mismatch in function"
|
| 702 |
+
" docstring w.r.t. function signature, at index %s"
|
| 703 |
+
" diff: %r != %r" % (i, param_signature[i], param_docs[i])
|
| 704 |
+
]
|
| 705 |
+
break
|
| 706 |
+
if len(param_signature) > len(param_docs):
|
| 707 |
+
message += [
|
| 708 |
+
"Parameters in function docstring have less items w.r.t."
|
| 709 |
+
" function signature, first missing item: %s"
|
| 710 |
+
% param_signature[len(param_docs)]
|
| 711 |
+
]
|
| 712 |
+
|
| 713 |
+
elif len(param_signature) < len(param_docs):
|
| 714 |
+
message += [
|
| 715 |
+
"Parameters in function docstring have more items w.r.t."
|
| 716 |
+
" function signature, first extra item: %s"
|
| 717 |
+
% param_docs[len(param_signature)]
|
| 718 |
+
]
|
| 719 |
+
|
| 720 |
+
# If there wasn't any difference in the parameters themselves between
|
| 721 |
+
# docstring and signature including having the same length then return
|
| 722 |
+
# empty list
|
| 723 |
+
if len(message) == 0:
|
| 724 |
+
return []
|
| 725 |
+
|
| 726 |
+
import difflib
|
| 727 |
+
import pprint
|
| 728 |
+
|
| 729 |
+
param_docs_formatted = pprint.pformat(param_docs).splitlines()
|
| 730 |
+
param_signature_formatted = pprint.pformat(param_signature).splitlines()
|
| 731 |
+
|
| 732 |
+
message += ["Full diff:"]
|
| 733 |
+
|
| 734 |
+
message.extend(
|
| 735 |
+
line.strip()
|
| 736 |
+
for line in difflib.ndiff(param_signature_formatted, param_docs_formatted)
|
| 737 |
+
)
|
| 738 |
+
|
| 739 |
+
incorrect.extend(message)
|
| 740 |
+
|
| 741 |
+
# Prepend function name
|
| 742 |
+
incorrect = ["In function: " + func_name] + incorrect
|
| 743 |
+
|
| 744 |
+
return incorrect
|
| 745 |
+
|
| 746 |
+
|
| 747 |
+
def assert_run_python_script(source_code, timeout=60):
|
| 748 |
+
"""Utility to check assertions in an independent Python subprocess.
|
| 749 |
+
|
| 750 |
+
The script provided in the source code should return 0 and not print
|
| 751 |
+
anything on stderr or stdout.
|
| 752 |
+
|
| 753 |
+
This is a port from cloudpickle https://github.com/cloudpipe/cloudpickle
|
| 754 |
+
|
| 755 |
+
Parameters
|
| 756 |
+
----------
|
| 757 |
+
source_code : str
|
| 758 |
+
The Python source code to execute.
|
| 759 |
+
timeout : int, default=60
|
| 760 |
+
Time in seconds before timeout.
|
| 761 |
+
"""
|
| 762 |
+
fd, source_file = tempfile.mkstemp(suffix="_src_test_sklearn.py")
|
| 763 |
+
os.close(fd)
|
| 764 |
+
try:
|
| 765 |
+
with open(source_file, "wb") as f:
|
| 766 |
+
f.write(source_code.encode("utf-8"))
|
| 767 |
+
cmd = [sys.executable, source_file]
|
| 768 |
+
cwd = op.normpath(op.join(op.dirname(sklearn.__file__), ".."))
|
| 769 |
+
env = os.environ.copy()
|
| 770 |
+
try:
|
| 771 |
+
env["PYTHONPATH"] = os.pathsep.join([cwd, env["PYTHONPATH"]])
|
| 772 |
+
except KeyError:
|
| 773 |
+
env["PYTHONPATH"] = cwd
|
| 774 |
+
kwargs = {"cwd": cwd, "stderr": STDOUT, "env": env}
|
| 775 |
+
# If coverage is running, pass the config file to the subprocess
|
| 776 |
+
coverage_rc = os.environ.get("COVERAGE_PROCESS_START")
|
| 777 |
+
if coverage_rc:
|
| 778 |
+
kwargs["env"]["COVERAGE_PROCESS_START"] = coverage_rc
|
| 779 |
+
|
| 780 |
+
kwargs["timeout"] = timeout
|
| 781 |
+
try:
|
| 782 |
+
try:
|
| 783 |
+
out = check_output(cmd, **kwargs)
|
| 784 |
+
except CalledProcessError as e:
|
| 785 |
+
raise RuntimeError(
|
| 786 |
+
"script errored with output:\n%s" % e.output.decode("utf-8")
|
| 787 |
+
)
|
| 788 |
+
if out != b"":
|
| 789 |
+
raise AssertionError(out.decode("utf-8"))
|
| 790 |
+
except TimeoutExpired as e:
|
| 791 |
+
raise RuntimeError(
|
| 792 |
+
"script timeout, output so far:\n%s" % e.output.decode("utf-8")
|
| 793 |
+
)
|
| 794 |
+
finally:
|
| 795 |
+
os.unlink(source_file)
|
| 796 |
+
|
| 797 |
+
|
| 798 |
+
def _convert_container(container, constructor_name, columns_name=None, dtype=None):
|
| 799 |
+
"""Convert a given container to a specific array-like with a dtype.
|
| 800 |
+
|
| 801 |
+
Parameters
|
| 802 |
+
----------
|
| 803 |
+
container : array-like
|
| 804 |
+
The container to convert.
|
| 805 |
+
constructor_name : {"list", "tuple", "array", "sparse", "dataframe", \
|
| 806 |
+
"series", "index", "slice", "sparse_csr", "sparse_csc"}
|
| 807 |
+
The type of the returned container.
|
| 808 |
+
columns_name : index or array-like, default=None
|
| 809 |
+
For pandas container supporting `columns_names`, it will affect
|
| 810 |
+
specific names.
|
| 811 |
+
dtype : dtype, default=None
|
| 812 |
+
Force the dtype of the container. Does not apply to `"slice"`
|
| 813 |
+
container.
|
| 814 |
+
|
| 815 |
+
Returns
|
| 816 |
+
-------
|
| 817 |
+
converted_container
|
| 818 |
+
"""
|
| 819 |
+
if constructor_name == "list":
|
| 820 |
+
if dtype is None:
|
| 821 |
+
return list(container)
|
| 822 |
+
else:
|
| 823 |
+
return np.asarray(container, dtype=dtype).tolist()
|
| 824 |
+
elif constructor_name == "tuple":
|
| 825 |
+
if dtype is None:
|
| 826 |
+
return tuple(container)
|
| 827 |
+
else:
|
| 828 |
+
return tuple(np.asarray(container, dtype=dtype).tolist())
|
| 829 |
+
elif constructor_name == "array":
|
| 830 |
+
return np.asarray(container, dtype=dtype)
|
| 831 |
+
elif constructor_name == "sparse":
|
| 832 |
+
return sp.sparse.csr_matrix(container, dtype=dtype)
|
| 833 |
+
elif constructor_name == "dataframe":
|
| 834 |
+
pd = pytest.importorskip("pandas")
|
| 835 |
+
return pd.DataFrame(container, columns=columns_name, dtype=dtype)
|
| 836 |
+
elif constructor_name == "series":
|
| 837 |
+
pd = pytest.importorskip("pandas")
|
| 838 |
+
return pd.Series(container, dtype=dtype)
|
| 839 |
+
elif constructor_name == "index":
|
| 840 |
+
pd = pytest.importorskip("pandas")
|
| 841 |
+
return pd.Index(container, dtype=dtype)
|
| 842 |
+
elif constructor_name == "slice":
|
| 843 |
+
return slice(container[0], container[1])
|
| 844 |
+
elif constructor_name == "sparse_csr":
|
| 845 |
+
return sp.sparse.csr_matrix(container, dtype=dtype)
|
| 846 |
+
elif constructor_name == "sparse_csc":
|
| 847 |
+
return sp.sparse.csc_matrix(container, dtype=dtype)
|
| 848 |
+
|
| 849 |
+
|
| 850 |
+
def raises(expected_exc_type, match=None, may_pass=False, err_msg=None):
|
| 851 |
+
"""Context manager to ensure exceptions are raised within a code block.
|
| 852 |
+
|
| 853 |
+
This is similar to and inspired from pytest.raises, but supports a few
|
| 854 |
+
other cases.
|
| 855 |
+
|
| 856 |
+
This is only intended to be used in estimator_checks.py where we don't
|
| 857 |
+
want to use pytest. In the rest of the code base, just use pytest.raises
|
| 858 |
+
instead.
|
| 859 |
+
|
| 860 |
+
Parameters
|
| 861 |
+
----------
|
| 862 |
+
excepted_exc_type : Exception or list of Exception
|
| 863 |
+
The exception that should be raised by the block. If a list, the block
|
| 864 |
+
should raise one of the exceptions.
|
| 865 |
+
match : str or list of str, default=None
|
| 866 |
+
A regex that the exception message should match. If a list, one of
|
| 867 |
+
the entries must match. If None, match isn't enforced.
|
| 868 |
+
may_pass : bool, default=False
|
| 869 |
+
If True, the block is allowed to not raise an exception. Useful in
|
| 870 |
+
cases where some estimators may support a feature but others must
|
| 871 |
+
fail with an appropriate error message. By default, the context
|
| 872 |
+
manager will raise an exception if the block does not raise an
|
| 873 |
+
exception.
|
| 874 |
+
err_msg : str, default=None
|
| 875 |
+
If the context manager fails (e.g. the block fails to raise the
|
| 876 |
+
proper exception, or fails to match), then an AssertionError is
|
| 877 |
+
raised with this message. By default, an AssertionError is raised
|
| 878 |
+
with a default error message (depends on the kind of failure). Use
|
| 879 |
+
this to indicate how users should fix their estimators to pass the
|
| 880 |
+
checks.
|
| 881 |
+
|
| 882 |
+
Attributes
|
| 883 |
+
----------
|
| 884 |
+
raised_and_matched : bool
|
| 885 |
+
True if an exception was raised and a match was found, False otherwise.
|
| 886 |
+
"""
|
| 887 |
+
return _Raises(expected_exc_type, match, may_pass, err_msg)
|
| 888 |
+
|
| 889 |
+
|
| 890 |
+
class _Raises(contextlib.AbstractContextManager):
|
| 891 |
+
# see raises() for parameters
|
| 892 |
+
def __init__(self, expected_exc_type, match, may_pass, err_msg):
|
| 893 |
+
self.expected_exc_types = (
|
| 894 |
+
expected_exc_type
|
| 895 |
+
if isinstance(expected_exc_type, Iterable)
|
| 896 |
+
else [expected_exc_type]
|
| 897 |
+
)
|
| 898 |
+
self.matches = [match] if isinstance(match, str) else match
|
| 899 |
+
self.may_pass = may_pass
|
| 900 |
+
self.err_msg = err_msg
|
| 901 |
+
self.raised_and_matched = False
|
| 902 |
+
|
| 903 |
+
def __exit__(self, exc_type, exc_value, _):
|
| 904 |
+
# see
|
| 905 |
+
# https://docs.python.org/2.5/whatsnew/pep-343.html#SECTION000910000000000000000
|
| 906 |
+
|
| 907 |
+
if exc_type is None: # No exception was raised in the block
|
| 908 |
+
if self.may_pass:
|
| 909 |
+
return True # CM is happy
|
| 910 |
+
else:
|
| 911 |
+
err_msg = self.err_msg or f"Did not raise: {self.expected_exc_types}"
|
| 912 |
+
raise AssertionError(err_msg)
|
| 913 |
+
|
| 914 |
+
if not any(
|
| 915 |
+
issubclass(exc_type, expected_type)
|
| 916 |
+
for expected_type in self.expected_exc_types
|
| 917 |
+
):
|
| 918 |
+
if self.err_msg is not None:
|
| 919 |
+
raise AssertionError(self.err_msg) from exc_value
|
| 920 |
+
else:
|
| 921 |
+
return False # will re-raise the original exception
|
| 922 |
+
|
| 923 |
+
if self.matches is not None:
|
| 924 |
+
err_msg = self.err_msg or (
|
| 925 |
+
"The error message should contain one of the following "
|
| 926 |
+
"patterns:\n{}\nGot {}".format("\n".join(self.matches), str(exc_value))
|
| 927 |
+
)
|
| 928 |
+
if not any(re.search(match, str(exc_value)) for match in self.matches):
|
| 929 |
+
raise AssertionError(err_msg) from exc_value
|
| 930 |
+
self.raised_and_matched = True
|
| 931 |
+
|
| 932 |
+
return True
|
| 933 |
+
|
| 934 |
+
|
| 935 |
+
class MinimalClassifier:
|
| 936 |
+
"""Minimal classifier implementation with inheriting from BaseEstimator.
|
| 937 |
+
|
| 938 |
+
This estimator should be tested with:
|
| 939 |
+
|
| 940 |
+
* `check_estimator` in `test_estimator_checks.py`;
|
| 941 |
+
* within a `Pipeline` in `test_pipeline.py`;
|
| 942 |
+
* within a `SearchCV` in `test_search.py`.
|
| 943 |
+
"""
|
| 944 |
+
|
| 945 |
+
_estimator_type = "classifier"
|
| 946 |
+
|
| 947 |
+
def __init__(self, param=None):
|
| 948 |
+
self.param = param
|
| 949 |
+
|
| 950 |
+
def get_params(self, deep=True):
|
| 951 |
+
return {"param": self.param}
|
| 952 |
+
|
| 953 |
+
def set_params(self, **params):
|
| 954 |
+
for key, value in params.items():
|
| 955 |
+
setattr(self, key, value)
|
| 956 |
+
return self
|
| 957 |
+
|
| 958 |
+
def fit(self, X, y):
|
| 959 |
+
X, y = check_X_y(X, y)
|
| 960 |
+
check_classification_targets(y)
|
| 961 |
+
self.classes_, counts = np.unique(y, return_counts=True)
|
| 962 |
+
self._most_frequent_class_idx = counts.argmax()
|
| 963 |
+
return self
|
| 964 |
+
|
| 965 |
+
def predict_proba(self, X):
|
| 966 |
+
check_is_fitted(self)
|
| 967 |
+
X = check_array(X)
|
| 968 |
+
proba_shape = (X.shape[0], self.classes_.size)
|
| 969 |
+
y_proba = np.zeros(shape=proba_shape, dtype=np.float64)
|
| 970 |
+
y_proba[:, self._most_frequent_class_idx] = 1.0
|
| 971 |
+
return y_proba
|
| 972 |
+
|
| 973 |
+
def predict(self, X):
|
| 974 |
+
y_proba = self.predict_proba(X)
|
| 975 |
+
y_pred = y_proba.argmax(axis=1)
|
| 976 |
+
return self.classes_[y_pred]
|
| 977 |
+
|
| 978 |
+
def score(self, X, y):
|
| 979 |
+
from sklearn.metrics import accuracy_score
|
| 980 |
+
|
| 981 |
+
return accuracy_score(y, self.predict(X))
|
| 982 |
+
|
| 983 |
+
|
| 984 |
+
class MinimalRegressor:
|
| 985 |
+
"""Minimal regressor implementation with inheriting from BaseEstimator.
|
| 986 |
+
|
| 987 |
+
This estimator should be tested with:
|
| 988 |
+
|
| 989 |
+
* `check_estimator` in `test_estimator_checks.py`;
|
| 990 |
+
* within a `Pipeline` in `test_pipeline.py`;
|
| 991 |
+
* within a `SearchCV` in `test_search.py`.
|
| 992 |
+
"""
|
| 993 |
+
|
| 994 |
+
_estimator_type = "regressor"
|
| 995 |
+
|
| 996 |
+
def __init__(self, param=None):
|
| 997 |
+
self.param = param
|
| 998 |
+
|
| 999 |
+
def get_params(self, deep=True):
|
| 1000 |
+
return {"param": self.param}
|
| 1001 |
+
|
| 1002 |
+
def set_params(self, **params):
|
| 1003 |
+
for key, value in params.items():
|
| 1004 |
+
setattr(self, key, value)
|
| 1005 |
+
return self
|
| 1006 |
+
|
| 1007 |
+
def fit(self, X, y):
|
| 1008 |
+
X, y = check_X_y(X, y)
|
| 1009 |
+
self.is_fitted_ = True
|
| 1010 |
+
self._mean = np.mean(y)
|
| 1011 |
+
return self
|
| 1012 |
+
|
| 1013 |
+
def predict(self, X):
|
| 1014 |
+
check_is_fitted(self)
|
| 1015 |
+
X = check_array(X)
|
| 1016 |
+
return np.ones(shape=(X.shape[0],)) * self._mean
|
| 1017 |
+
|
| 1018 |
+
def score(self, X, y):
|
| 1019 |
+
from sklearn.metrics import r2_score
|
| 1020 |
+
|
| 1021 |
+
return r2_score(y, self.predict(X))
|
| 1022 |
+
|
| 1023 |
+
|
| 1024 |
+
class MinimalTransformer:
|
| 1025 |
+
"""Minimal transformer implementation with inheriting from
|
| 1026 |
+
BaseEstimator.
|
| 1027 |
+
|
| 1028 |
+
This estimator should be tested with:
|
| 1029 |
+
|
| 1030 |
+
* `check_estimator` in `test_estimator_checks.py`;
|
| 1031 |
+
* within a `Pipeline` in `test_pipeline.py`;
|
| 1032 |
+
* within a `SearchCV` in `test_search.py`.
|
| 1033 |
+
"""
|
| 1034 |
+
|
| 1035 |
+
def __init__(self, param=None):
|
| 1036 |
+
self.param = param
|
| 1037 |
+
|
| 1038 |
+
def get_params(self, deep=True):
|
| 1039 |
+
return {"param": self.param}
|
| 1040 |
+
|
| 1041 |
+
def set_params(self, **params):
|
| 1042 |
+
for key, value in params.items():
|
| 1043 |
+
setattr(self, key, value)
|
| 1044 |
+
return self
|
| 1045 |
+
|
| 1046 |
+
def fit(self, X, y=None):
|
| 1047 |
+
check_array(X)
|
| 1048 |
+
self.is_fitted_ = True
|
| 1049 |
+
return self
|
| 1050 |
+
|
| 1051 |
+
def transform(self, X, y=None):
|
| 1052 |
+
check_is_fitted(self)
|
| 1053 |
+
X = check_array(X)
|
| 1054 |
+
return X
|
| 1055 |
+
|
| 1056 |
+
def fit_transform(self, X, y=None):
|
| 1057 |
+
return self.fit(X, y).transform(X, y)
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_typedefs.pxd
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!python
|
| 2 |
+
cimport numpy as cnp
|
| 3 |
+
|
| 4 |
+
# Floating point/data type
|
| 5 |
+
ctypedef cnp.float64_t DTYPE_t # WARNING: should match DTYPE in typedefs.pyx
|
| 6 |
+
|
| 7 |
+
cdef enum:
|
| 8 |
+
DTYPECODE = cnp.NPY_FLOAT64
|
| 9 |
+
ITYPECODE = cnp.NPY_INTP
|
| 10 |
+
INT32TYPECODE = cnp.NPY_INT32
|
| 11 |
+
INT64TYPECODE = cnp.NPY_INT64
|
| 12 |
+
|
| 13 |
+
# Index/integer type.
|
| 14 |
+
# WARNING: ITYPE_t must be a signed integer type or you will have a bad time!
|
| 15 |
+
ctypedef cnp.intp_t ITYPE_t # WARNING: should match ITYPE in typedefs.pyx
|
| 16 |
+
ctypedef cnp.int32_t INT32TYPE_t # WARNING: should match INT32TYPE in typedefs.pyx
|
| 17 |
+
ctypedef cnp.int64_t INT64TYPE_t # WARNING: should match INT32TYPE in typedefs.pyx
|
| 18 |
+
|
| 19 |
+
# scipy matrices indices dtype (namely for indptr and indices arrays)
|
| 20 |
+
#
|
| 21 |
+
# Note that indices might need to be represented as cnp.int64_t.
|
| 22 |
+
# Currently, we use Cython classes which do not handle fused types
|
| 23 |
+
# so we hardcode this type to cnp.int32_t, supporting all but edge
|
| 24 |
+
# cases.
|
| 25 |
+
#
|
| 26 |
+
# TODO: support cnp.int64_t for this case
|
| 27 |
+
# See: https://github.com/scikit-learn/scikit-learn/issues/23653
|
| 28 |
+
ctypedef cnp.int32_t SPARSE_INDEX_TYPE_t
|
videollama2/lib/python3.10/site-packages/sklearn/utils/_vector_sentinel.pxd
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
cimport numpy as cnp
|
| 2 |
+
|
| 3 |
+
from libcpp.vector cimport vector
|
| 4 |
+
from ..utils._typedefs cimport ITYPE_t, DTYPE_t, INT32TYPE_t, INT64TYPE_t
|
| 5 |
+
|
| 6 |
+
ctypedef fused vector_typed:
|
| 7 |
+
vector[DTYPE_t]
|
| 8 |
+
vector[ITYPE_t]
|
| 9 |
+
vector[INT32TYPE_t]
|
| 10 |
+
vector[INT64TYPE_t]
|
| 11 |
+
|
| 12 |
+
cdef cnp.ndarray vector_to_nd_array(vector_typed * vect_ptr)
|
videollama2/lib/python3.10/site-packages/sklearn/utils/discovery.py
ADDED
|
@@ -0,0 +1,217 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import pkgutil
|
| 2 |
+
import inspect
|
| 3 |
+
from importlib import import_module
|
| 4 |
+
from operator import itemgetter
|
| 5 |
+
from pathlib import Path
|
| 6 |
+
|
| 7 |
+
_MODULE_TO_IGNORE = {
|
| 8 |
+
"tests",
|
| 9 |
+
"externals",
|
| 10 |
+
"setup",
|
| 11 |
+
"conftest",
|
| 12 |
+
"experimental",
|
| 13 |
+
"estimator_checks",
|
| 14 |
+
}
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
def all_estimators(type_filter=None):
|
| 18 |
+
"""Get a list of all estimators from `sklearn`.
|
| 19 |
+
|
| 20 |
+
This function crawls the module and gets all classes that inherit
|
| 21 |
+
from BaseEstimator. Classes that are defined in test-modules are not
|
| 22 |
+
included.
|
| 23 |
+
|
| 24 |
+
Parameters
|
| 25 |
+
----------
|
| 26 |
+
type_filter : {"classifier", "regressor", "cluster", "transformer"} \
|
| 27 |
+
or list of such str, default=None
|
| 28 |
+
Which kind of estimators should be returned. If None, no filter is
|
| 29 |
+
applied and all estimators are returned. Possible values are
|
| 30 |
+
'classifier', 'regressor', 'cluster' and 'transformer' to get
|
| 31 |
+
estimators only of these specific types, or a list of these to
|
| 32 |
+
get the estimators that fit at least one of the types.
|
| 33 |
+
|
| 34 |
+
Returns
|
| 35 |
+
-------
|
| 36 |
+
estimators : list of tuples
|
| 37 |
+
List of (name, class), where ``name`` is the class name as string
|
| 38 |
+
and ``class`` is the actual type of the class.
|
| 39 |
+
"""
|
| 40 |
+
# lazy import to avoid circular imports from sklearn.base
|
| 41 |
+
from . import IS_PYPY
|
| 42 |
+
from ._testing import ignore_warnings
|
| 43 |
+
from ..base import (
|
| 44 |
+
BaseEstimator,
|
| 45 |
+
ClassifierMixin,
|
| 46 |
+
RegressorMixin,
|
| 47 |
+
TransformerMixin,
|
| 48 |
+
ClusterMixin,
|
| 49 |
+
)
|
| 50 |
+
|
| 51 |
+
def is_abstract(c):
|
| 52 |
+
if not (hasattr(c, "__abstractmethods__")):
|
| 53 |
+
return False
|
| 54 |
+
if not len(c.__abstractmethods__):
|
| 55 |
+
return False
|
| 56 |
+
return True
|
| 57 |
+
|
| 58 |
+
all_classes = []
|
| 59 |
+
root = str(Path(__file__).parent.parent) # sklearn package
|
| 60 |
+
# Ignore deprecation warnings triggered at import time and from walking
|
| 61 |
+
# packages
|
| 62 |
+
with ignore_warnings(category=FutureWarning):
|
| 63 |
+
for _, module_name, _ in pkgutil.walk_packages(path=[root], prefix="sklearn."):
|
| 64 |
+
module_parts = module_name.split(".")
|
| 65 |
+
if (
|
| 66 |
+
any(part in _MODULE_TO_IGNORE for part in module_parts)
|
| 67 |
+
or "._" in module_name
|
| 68 |
+
):
|
| 69 |
+
continue
|
| 70 |
+
module = import_module(module_name)
|
| 71 |
+
classes = inspect.getmembers(module, inspect.isclass)
|
| 72 |
+
classes = [
|
| 73 |
+
(name, est_cls) for name, est_cls in classes if not name.startswith("_")
|
| 74 |
+
]
|
| 75 |
+
|
| 76 |
+
# TODO: Remove when FeatureHasher is implemented in PYPY
|
| 77 |
+
# Skips FeatureHasher for PYPY
|
| 78 |
+
if IS_PYPY and "feature_extraction" in module_name:
|
| 79 |
+
classes = [
|
| 80 |
+
(name, est_cls)
|
| 81 |
+
for name, est_cls in classes
|
| 82 |
+
if name == "FeatureHasher"
|
| 83 |
+
]
|
| 84 |
+
|
| 85 |
+
all_classes.extend(classes)
|
| 86 |
+
|
| 87 |
+
all_classes = set(all_classes)
|
| 88 |
+
|
| 89 |
+
estimators = [
|
| 90 |
+
c
|
| 91 |
+
for c in all_classes
|
| 92 |
+
if (issubclass(c[1], BaseEstimator) and c[0] != "BaseEstimator")
|
| 93 |
+
]
|
| 94 |
+
# get rid of abstract base classes
|
| 95 |
+
estimators = [c for c in estimators if not is_abstract(c[1])]
|
| 96 |
+
|
| 97 |
+
if type_filter is not None:
|
| 98 |
+
if not isinstance(type_filter, list):
|
| 99 |
+
type_filter = [type_filter]
|
| 100 |
+
else:
|
| 101 |
+
type_filter = list(type_filter) # copy
|
| 102 |
+
filtered_estimators = []
|
| 103 |
+
filters = {
|
| 104 |
+
"classifier": ClassifierMixin,
|
| 105 |
+
"regressor": RegressorMixin,
|
| 106 |
+
"transformer": TransformerMixin,
|
| 107 |
+
"cluster": ClusterMixin,
|
| 108 |
+
}
|
| 109 |
+
for name, mixin in filters.items():
|
| 110 |
+
if name in type_filter:
|
| 111 |
+
type_filter.remove(name)
|
| 112 |
+
filtered_estimators.extend(
|
| 113 |
+
[est for est in estimators if issubclass(est[1], mixin)]
|
| 114 |
+
)
|
| 115 |
+
estimators = filtered_estimators
|
| 116 |
+
if type_filter:
|
| 117 |
+
raise ValueError(
|
| 118 |
+
"Parameter type_filter must be 'classifier', "
|
| 119 |
+
"'regressor', 'transformer', 'cluster' or "
|
| 120 |
+
"None, got"
|
| 121 |
+
f" {repr(type_filter)}."
|
| 122 |
+
)
|
| 123 |
+
|
| 124 |
+
# drop duplicates, sort for reproducibility
|
| 125 |
+
# itemgetter is used to ensure the sort does not extend to the 2nd item of
|
| 126 |
+
# the tuple
|
| 127 |
+
return sorted(set(estimators), key=itemgetter(0))
|
| 128 |
+
|
| 129 |
+
|
| 130 |
+
def all_displays():
|
| 131 |
+
"""Get a list of all displays from `sklearn`.
|
| 132 |
+
|
| 133 |
+
Returns
|
| 134 |
+
-------
|
| 135 |
+
displays : list of tuples
|
| 136 |
+
List of (name, class), where ``name`` is the display class name as
|
| 137 |
+
string and ``class`` is the actual type of the class.
|
| 138 |
+
"""
|
| 139 |
+
# lazy import to avoid circular imports from sklearn.base
|
| 140 |
+
from ._testing import ignore_warnings
|
| 141 |
+
|
| 142 |
+
all_classes = []
|
| 143 |
+
root = str(Path(__file__).parent.parent) # sklearn package
|
| 144 |
+
# Ignore deprecation warnings triggered at import time and from walking
|
| 145 |
+
# packages
|
| 146 |
+
with ignore_warnings(category=FutureWarning):
|
| 147 |
+
for _, module_name, _ in pkgutil.walk_packages(path=[root], prefix="sklearn."):
|
| 148 |
+
module_parts = module_name.split(".")
|
| 149 |
+
if (
|
| 150 |
+
any(part in _MODULE_TO_IGNORE for part in module_parts)
|
| 151 |
+
or "._" in module_name
|
| 152 |
+
):
|
| 153 |
+
continue
|
| 154 |
+
module = import_module(module_name)
|
| 155 |
+
classes = inspect.getmembers(module, inspect.isclass)
|
| 156 |
+
classes = [
|
| 157 |
+
(name, display_class)
|
| 158 |
+
for name, display_class in classes
|
| 159 |
+
if not name.startswith("_") and name.endswith("Display")
|
| 160 |
+
]
|
| 161 |
+
all_classes.extend(classes)
|
| 162 |
+
|
| 163 |
+
return sorted(set(all_classes), key=itemgetter(0))
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
def _is_checked_function(item):
|
| 167 |
+
if not inspect.isfunction(item):
|
| 168 |
+
return False
|
| 169 |
+
|
| 170 |
+
if item.__name__.startswith("_"):
|
| 171 |
+
return False
|
| 172 |
+
|
| 173 |
+
mod = item.__module__
|
| 174 |
+
if not mod.startswith("sklearn.") or mod.endswith("estimator_checks"):
|
| 175 |
+
return False
|
| 176 |
+
|
| 177 |
+
return True
|
| 178 |
+
|
| 179 |
+
|
| 180 |
+
def all_functions():
|
| 181 |
+
"""Get a list of all functions from `sklearn`.
|
| 182 |
+
|
| 183 |
+
Returns
|
| 184 |
+
-------
|
| 185 |
+
functions : list of tuples
|
| 186 |
+
List of (name, function), where ``name`` is the function name as
|
| 187 |
+
string and ``function`` is the actual function.
|
| 188 |
+
"""
|
| 189 |
+
# lazy import to avoid circular imports from sklearn.base
|
| 190 |
+
from ._testing import ignore_warnings
|
| 191 |
+
|
| 192 |
+
all_functions = []
|
| 193 |
+
root = str(Path(__file__).parent.parent) # sklearn package
|
| 194 |
+
# Ignore deprecation warnings triggered at import time and from walking
|
| 195 |
+
# packages
|
| 196 |
+
with ignore_warnings(category=FutureWarning):
|
| 197 |
+
for _, module_name, _ in pkgutil.walk_packages(path=[root], prefix="sklearn."):
|
| 198 |
+
module_parts = module_name.split(".")
|
| 199 |
+
if (
|
| 200 |
+
any(part in _MODULE_TO_IGNORE for part in module_parts)
|
| 201 |
+
or "._" in module_name
|
| 202 |
+
):
|
| 203 |
+
continue
|
| 204 |
+
|
| 205 |
+
module = import_module(module_name)
|
| 206 |
+
functions = inspect.getmembers(module, _is_checked_function)
|
| 207 |
+
functions = [
|
| 208 |
+
(func.__name__, func)
|
| 209 |
+
for name, func in functions
|
| 210 |
+
if not name.startswith("_")
|
| 211 |
+
]
|
| 212 |
+
all_functions.extend(functions)
|
| 213 |
+
|
| 214 |
+
# drop duplicates, sort for reproducibility
|
| 215 |
+
# itemgetter is used to ensure the sort does not extend to the 2nd item of
|
| 216 |
+
# the tuple
|
| 217 |
+
return sorted(set(all_functions), key=itemgetter(0))
|
videollama2/lib/python3.10/site-packages/sklearn/utils/estimator_checks.py
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
videollama2/lib/python3.10/site-packages/sklearn/utils/murmurhash.cpython-310-x86_64-linux-gnu.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:451f5dae12ddc37a669aa27561bb664bbee60a44ffb4d688d5507e9ac89f5f84
|
| 3 |
+
size 229345
|
videollama2/lib/python3.10/site-packages/sklearn/utils/parallel.py
ADDED
|
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Module that customize joblib tools for scikit-learn usage."""
|
| 2 |
+
|
| 3 |
+
import functools
|
| 4 |
+
import warnings
|
| 5 |
+
from functools import update_wrapper
|
| 6 |
+
|
| 7 |
+
import joblib
|
| 8 |
+
|
| 9 |
+
from .._config import config_context, get_config
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
def _with_config(delayed_func, config):
|
| 13 |
+
"""Helper function that intends to attach a config to a delayed function."""
|
| 14 |
+
if hasattr(delayed_func, "with_config"):
|
| 15 |
+
return delayed_func.with_config(config)
|
| 16 |
+
else:
|
| 17 |
+
warnings.warn(
|
| 18 |
+
"`sklearn.utils.parallel.Parallel` needs to be used in "
|
| 19 |
+
"conjunction with `sklearn.utils.parallel.delayed` instead of "
|
| 20 |
+
"`joblib.delayed` to correctly propagate the scikit-learn "
|
| 21 |
+
"configuration to the joblib workers.",
|
| 22 |
+
UserWarning,
|
| 23 |
+
)
|
| 24 |
+
return delayed_func
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
class Parallel(joblib.Parallel):
|
| 28 |
+
"""Tweak of :class:`joblib.Parallel` that propagates the scikit-learn configuration.
|
| 29 |
+
|
| 30 |
+
This subclass of :class:`joblib.Parallel` ensures that the active configuration
|
| 31 |
+
(thread-local) of scikit-learn is propagated to the parallel workers for the
|
| 32 |
+
duration of the execution of the parallel tasks.
|
| 33 |
+
|
| 34 |
+
The API does not change and you can refer to :class:`joblib.Parallel`
|
| 35 |
+
documentation for more details.
|
| 36 |
+
|
| 37 |
+
.. versionadded:: 1.3
|
| 38 |
+
"""
|
| 39 |
+
|
| 40 |
+
def __call__(self, iterable):
|
| 41 |
+
"""Dispatch the tasks and return the results.
|
| 42 |
+
|
| 43 |
+
Parameters
|
| 44 |
+
----------
|
| 45 |
+
iterable : iterable
|
| 46 |
+
Iterable containing tuples of (delayed_function, args, kwargs) that should
|
| 47 |
+
be consumed.
|
| 48 |
+
|
| 49 |
+
Returns
|
| 50 |
+
-------
|
| 51 |
+
results : list
|
| 52 |
+
List of results of the tasks.
|
| 53 |
+
"""
|
| 54 |
+
# Capture the thread-local scikit-learn configuration at the time
|
| 55 |
+
# Parallel.__call__ is issued since the tasks can be dispatched
|
| 56 |
+
# in a different thread depending on the backend and on the value of
|
| 57 |
+
# pre_dispatch and n_jobs.
|
| 58 |
+
config = get_config()
|
| 59 |
+
iterable_with_config = (
|
| 60 |
+
(_with_config(delayed_func, config), args, kwargs)
|
| 61 |
+
for delayed_func, args, kwargs in iterable
|
| 62 |
+
)
|
| 63 |
+
return super().__call__(iterable_with_config)
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
# remove when https://github.com/joblib/joblib/issues/1071 is fixed
|
| 67 |
+
def delayed(function):
|
| 68 |
+
"""Decorator used to capture the arguments of a function.
|
| 69 |
+
|
| 70 |
+
This alternative to `joblib.delayed` is meant to be used in conjunction
|
| 71 |
+
with `sklearn.utils.parallel.Parallel`. The latter captures the the scikit-
|
| 72 |
+
learn configuration by calling `sklearn.get_config()` in the current
|
| 73 |
+
thread, prior to dispatching the first task. The captured configuration is
|
| 74 |
+
then propagated and enabled for the duration of the execution of the
|
| 75 |
+
delayed function in the joblib workers.
|
| 76 |
+
|
| 77 |
+
.. versionchanged:: 1.3
|
| 78 |
+
`delayed` was moved from `sklearn.utils.fixes` to `sklearn.utils.parallel`
|
| 79 |
+
in scikit-learn 1.3.
|
| 80 |
+
|
| 81 |
+
Parameters
|
| 82 |
+
----------
|
| 83 |
+
function : callable
|
| 84 |
+
The function to be delayed.
|
| 85 |
+
|
| 86 |
+
Returns
|
| 87 |
+
-------
|
| 88 |
+
output: tuple
|
| 89 |
+
Tuple containing the delayed function, the positional arguments, and the
|
| 90 |
+
keyword arguments.
|
| 91 |
+
"""
|
| 92 |
+
|
| 93 |
+
@functools.wraps(function)
|
| 94 |
+
def delayed_function(*args, **kwargs):
|
| 95 |
+
return _FuncWrapper(function), args, kwargs
|
| 96 |
+
|
| 97 |
+
return delayed_function
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
class _FuncWrapper:
|
| 101 |
+
"""Load the global configuration before calling the function."""
|
| 102 |
+
|
| 103 |
+
def __init__(self, function):
|
| 104 |
+
self.function = function
|
| 105 |
+
update_wrapper(self, self.function)
|
| 106 |
+
|
| 107 |
+
def with_config(self, config):
|
| 108 |
+
self.config = config
|
| 109 |
+
return self
|
| 110 |
+
|
| 111 |
+
def __call__(self, *args, **kwargs):
|
| 112 |
+
config = getattr(self, "config", None)
|
| 113 |
+
if config is None:
|
| 114 |
+
warnings.warn(
|
| 115 |
+
"`sklearn.utils.parallel.delayed` should be used with "
|
| 116 |
+
"`sklearn.utils.parallel.Parallel` to make it possible to propagate "
|
| 117 |
+
"the scikit-learn configuration of the current thread to the "
|
| 118 |
+
"joblib workers.",
|
| 119 |
+
UserWarning,
|
| 120 |
+
)
|
| 121 |
+
config = {}
|
| 122 |
+
with config_context(**config):
|
| 123 |
+
return self.function(*args, **kwargs)
|