Dask: PCA

Orange/projection/pca.py

@@ -1,11 +1,13 @@
+import warnings
 import numbers
 import six
 import numpy as np
 import scipy.sparse as sp
+import dask.array as da
 from scipy.linalg import lu, qr, svd
 
 from sklearn import decomposition as skl_decomposition
-from sklearn.utils import check_array, check_random_state
+from sklearn.utils import check_random_state
 from sklearn.utils.extmath import svd_flip, safe_sparse_dot
 from sklearn.utils.validation import check_is_fitted
 
@@ -261,12 +263,39 @@
         self.params = vars()
 
     def fit(self, X, Y=None):
+        proj = self._initialize_wrapped(X, Y)
+        if isinstance(X, da.Array):
+            X = X.rechunk({0: "auto", 1: -1})
+        return proj.fit(X, Y)
+
+    def _initialize_wrapped(self, X=None, Y=None):
         params = self.params.copy()
         if params["n_components"] is not None:
-            params["n_components"] = min(min(X.shape), params["n_components"])
-        proj = self.__wraps__(**params)
-        proj = proj.fit(X, Y)
-        return PCAModel(proj, self.domain, len(proj.components_))
+            params["n_components"] = min(*X.shape, params["n_components"])
+
+        if isinstance(X, da.Array) or isinstance(Y, da.Array):
+            try:
+                import dask_ml.decomposition as dask_decomposition
+
+                if params["iterated_power"] == "auto":
+                    params["iterated_power"] = 0
+                del params["tol"]
+
+                # use IPCA instead of PCA due to memory issues
+                return dask_decomposition.IncrementalPCA(**params)
+
+            except ImportError:
+                warnings.warn("dask_ml is not installed. Using sklearn instead.")
+
+        return self.__wraps__(**params)
+
+    def __call__(self, data):
+        data = self.preprocess(data)
+        proj = self.fit(data.X, data.Y)
+        model = PCAModel(proj, data.domain, len(proj.components_))
+        model.pre_domain = data.domain
+        model.name = self.name
+        return model
 
 
 class SparsePCA(SklProjector):

Orange/tests/test_pca.py

@@ -12,18 +12,19 @@
 from Orange.preprocess import Continuize, Normalize
 from Orange.projection import pca, PCA, SparsePCA, IncrementalPCA, TruncatedSVD
 from Orange.tests import test_filename
+from Orange.tests.test_dasktable import with_dasktable
 
 
 class TestPCA(unittest.TestCase):
-    @classmethod
-    def setUpClass(cls):
-        cls.ionosphere = Table(test_filename('datasets/ionosphere.tab'))
-        cls.iris = Table('iris')
-        cls.zoo = Table('zoo')
-
-    def test_pca(self):
-        data = self.ionosphere
-        self.__pca_test_helper(data, n_com=3, min_xpl_var=0.5)
+    def setUp(self):
+        self.ionosphere = Table(test_filename('datasets/ionosphere.tab'))
+        self.iris = Table('iris')
+        self.zoo = Table('zoo')
+
+    @with_dasktable
+    def test_pca(self, prepare_table):
+        data = prepare_table(self.ionosphere)
+        self.__pca_test_helper(data, n_com=3, min_xpl_var=0.49)
         self.__pca_test_helper(data, n_com=10, min_xpl_var=0.7)
         self.__pca_test_helper(data, n_com=32, min_xpl_var=1)
 
@@ -35,7 +36,7 @@ def __pca_test_helper(self, data, n_com, min_xpl_var):
         self.assertEqual(n_com, pca_model.n_components)
         self.assertEqual((n_com, data.X.shape[1]), pca_model.components_.shape)
         proj = np.dot(data.X - pca_model.mean_, pca_model.components_.T)
-        np.testing.assert_almost_equal(pca_model(data).X, proj)
+        self.assertTrue(np.allclose(pca_model(data).X, proj))
 
     def test_sparse_pca(self):
         data = self.ionosphere[:100]
@@ -50,9 +51,10 @@ def __sparse_pca_test_helper(self, data, n_com, max_err):
         self.assertEqual((n_com, data.X.shape[1]), pca_model.components_.shape)
         self.assertLessEqual(pca_model.error_[-1], max_err)
 
-    def test_randomized_pca(self):
-        data = self.ionosphere
-        self.__rnd_pca_test_helper(data, n_com=3, min_xpl_var=0.5)
+    @with_dasktable
+    def test_randomized_pca(self, prepare_table):
+        data = prepare_table(self.ionosphere)
+        self.__rnd_pca_test_helper(data, n_com=3, min_xpl_var=0.47)
         self.__rnd_pca_test_helper(data, n_com=10, min_xpl_var=0.7)
         self.__rnd_pca_test_helper(data, n_com=32, min_xpl_var=0.98)
 
@@ -64,7 +66,7 @@ def __rnd_pca_test_helper(self, data, n_com, min_xpl_var):
         self.assertEqual(n_com, pca_model.n_components)
         self.assertEqual((n_com, data.X.shape[1]), pca_model.components_.shape)
         proj = np.dot(data.X - pca_model.mean_, pca_model.components_.T)
-        np.testing.assert_almost_equal(pca_model(data).X, proj)
+        self.assertTrue(np.allclose(pca_model(data).X, proj))
 
     def test_improved_randomized_pca_properly_called(self):
         # It doesn't matter what we put into the matrix
@@ -215,17 +217,20 @@ def test_transformed_domain_does_not_pickle_data(self):
         pca_iris2 = pickle.loads(pickle.dumps(pca_iris))
         self.assertIsNone(pca_iris2.domain[0].compute_value.transformed)
 
-    def test_chain(self):
-        zoo_c = Continuize()(self.zoo)
-        pca = PCA(n_components=3)(zoo_c)(self.zoo)
+    @with_dasktable
+    def test_chain(self, prepare_table):
+        zoo = prepare_table(self.zoo)
+        zoo_c = Continuize()(zoo)
+        pca = PCA(n_components=3)(zoo_c)(zoo)
         pca2 = PCA(n_components=3)(zoo_c)(zoo_c)
         pp = [Continuize()]
-        pca3 = PCA(n_components=3, preprocessors=pp)(self.zoo)(self.zoo)
-        np.testing.assert_almost_equal(pca.X, pca2.X)
-        np.testing.assert_almost_equal(pca.X, pca3.X)
+        pca3 = PCA(n_components=3, preprocessors=pp)(zoo)(zoo)
+        self.assertTrue(np.allclose(pca.X, pca2.X))
+        self.assertTrue(np.allclose(pca.X, pca3.X))
 
-    def test_PCA_scorer(self):
-        data = self.iris
+    @with_dasktable
+    def test_PCA_scorer(self, prepare_table):
+        data = prepare_table(self.iris)
         pca = PCA(preprocessors=[Normalize()])
         pca.component = 1
         scores = pca.score_data(data)
@@ -236,23 +241,28 @@ def test_PCA_scorer(self):
         self.assertEqual([round(s, 4) for s in scores[0]],
                          [0.5224, 0.2634, 0.5813, 0.5656])
 
-    def test_PCA_scorer_component(self):
+    @with_dasktable
+    def test_PCA_scorer_component(self, prepare_table):
         pca = PCA()
-        for i in range(1, len(self.zoo.domain.attributes) + 1):
+        zoo = prepare_table(self.zoo)
+        for i in range(1, len(zoo.domain.attributes) + 1):
             pca.component = i
-            scores = pca.score_data(self.zoo)
+            scores = pca.score_data(zoo)
             self.assertEqual(scores.shape,
-                             (pca.component, len(self.zoo.domain.attributes)))
+                             (pca.component, len(zoo.domain.attributes)))
 
-    def test_PCA_scorer_all_components(self):
-        n_attr = len(self.iris.domain.attributes)
+    @with_dasktable
+    def test_PCA_scorer_all_components(self, prepare_table):
+        iris = prepare_table(self.iris)
+        n_attr = len(iris.domain.attributes)
         pca = PCA()
-        scores = pca.score_data(self.iris)
+        scores = pca.score_data(iris)
         self.assertEqual(scores.shape, (n_attr, n_attr))
 
-    def test_max_components(self):
+    @with_dasktable
+    def test_max_components(self, prepare_table):
         d = np.random.RandomState(0).rand(20, 20)
-        data = Table.from_numpy(None, d)
+        data = prepare_table(Table.from_numpy(None, d))
         pca = PCA()(data)
         self.assertEqual(len(pca.explained_variance_ratio_), 20)
         pca = PCA(n_components=10)(data)

Orange/widgets/unsupervised/owpca.py

@@ -14,6 +14,7 @@
 from Orange.widgets import widget, gui, settings
 from Orange.widgets.utils.slidergraph import SliderGraph
 from Orange.widgets.utils.widgetpreview import WidgetPreview
+from Orange.widgets.utils.annotated_data import add_columns
 from Orange.widgets.widget import Input, Output
 
 # Maximum number of PCA components that we can set in the widget
@@ -329,14 +330,8 @@ def commit(self):
                                metas=metas)
             components.name = 'components'
 
-            data_dom = Domain(
-                self.data.domain.attributes,
-                self.data.domain.class_vars,
-                self.data.domain.metas + domain.attributes)
-            data = Table.from_numpy(
-                data_dom, self.data.X, self.data.Y,
-                numpy.hstack((self.data.metas, transformed.X)),
-                ids=self.data.ids)
+            data = self.data.transform(add_columns(self.data.domain,
+                                                   metas=domain.attributes))
 
         self._pca_projector.component = self.ncomponents
         self.Outputs.transformed_data.send(transformed)

Orange/widgets/unsupervised/tests/test_owpca.py

@@ -4,6 +4,8 @@
 from unittest.mock import patch, Mock
 
 import numpy as np
+from sklearn.utils import check_random_state
+from sklearn.utils.extmath import svd_flip
 
 from Orange.data import Table, Domain, ContinuousVariable, TimeVariable
 from Orange.preprocess import preprocess
@@ -12,22 +14,24 @@
 from Orange.widgets.tests.utils import table_dense_sparse, possible_duplicate_table
 from Orange.widgets.unsupervised.owpca import OWPCA
 from Orange.tests import test_filename
-from sklearn.utils import check_random_state
-from sklearn.utils.extmath import svd_flip
+from Orange.tests.test_dasktable import with_dasktable
 
 
 class TestOWPCA(WidgetTest):
     def setUp(self):
         self.widget = self.create_widget(OWPCA)  # type: OWPCA
         self.iris = Table("iris")  # type: Table
 
-    def test_set_variance100(self):
-        self.widget.set_data(self.iris)
+    @with_dasktable
+    def test_set_variance100(self, prepare_table):
+        data = prepare_table(self.iris)
+        self.widget.set_data(data)
         self.widget.variance_covered = 100
         self.widget._update_selection_variance_spin()
 
-    def test_constant_data(self):
-        data = self.iris[::5].copy()
+    @with_dasktable
+    def test_constant_data(self, prepare_table):
+        data = prepare_table(self.iris[::5].copy())
         with data.unlocked():
             data.X[:, :] = 1.0
         # Ignore the warning: the test checks whether the widget shows
@@ -38,23 +42,26 @@ def test_constant_data(self):
         self.assertIsNone(self.get_output(self.widget.Outputs.transformed_data))
         self.assertIsNone(self.get_output(self.widget.Outputs.components))
 
-    def test_empty_data(self):
+    @with_dasktable
+    def test_empty_data(self, prepare_table):
         """ Check widget for dataset with no rows and for dataset with no attributes """
-        self.send_signal(self.widget.Inputs.data, self.iris[:0])
+        data = prepare_table(self.iris)
+        self.send_signal(self.widget.Inputs.data, data[:0])
         self.assertTrue(self.widget.Error.no_instances.is_shown())
 
-        domain = Domain([], None, self.iris.domain.variables)
-        new_data = Table.from_table(domain, self.iris)
+        domain = Domain([], None, data.domain.variables)
+        new_data = data.transform(domain)
         self.send_signal(self.widget.Inputs.data, new_data)
         self.assertTrue(self.widget.Error.no_features.is_shown())
         self.assertFalse(self.widget.Error.no_instances.is_shown())
 
         self.send_signal(self.widget.Inputs.data, None)
         self.assertFalse(self.widget.Error.no_features.is_shown())
 
-    def test_limit_components(self):
+    @with_dasktable
+    def test_limit_components(self, prepare_table):
         X = np.random.RandomState(0).rand(101, 101)
-        data = Table.from_numpy(None, X)
+        data = prepare_table(Table.from_numpy(None, X))
         self.widget.ncomponents = 100
         self.send_signal(self.widget.Inputs.data, data)
         tran = self.get_output(self.widget.Outputs.transformed_data)
@@ -80,8 +87,10 @@ def test_migrate_settings_changes_variance_covered_to_int(self):
         OWPCA.migrate_settings(settings, 0)
         self.assertEqual(settings["variance_covered"], 100)
 
-    def test_variance_shown(self):
-        self.send_signal(self.widget.Inputs.data, self.iris)
+    @with_dasktable
+    def test_variance_shown(self, prepare_table):
+        data = prepare_table(self.iris)
+        self.send_signal(self.widget.Inputs.data, data)
         self.widget.maxp = 2
         self.widget._setup_plot()
         var2 = self.widget.variance_covered
@@ -90,15 +99,18 @@ def test_variance_shown(self):
         var3 = self.widget.variance_covered
         self.assertGreater(var3, var2)
 
-    def test_unique_domain_components(self):
-        table = possible_duplicate_table('components')
+    @with_dasktable
+    def test_unique_domain_components(self, prepare_table):
+        table = prepare_table(possible_duplicate_table('components'))
         self.send_signal(self.widget.Inputs.data, table)
         out = self.get_output(self.widget.Outputs.components)
         self.assertEqual(out.domain.metas[0].name, 'components (1)')
 
-    def test_variance_attr(self):
+    @with_dasktable
+    def test_variance_attr(self, prepare_table):
+        data = prepare_table(self.iris)
         self.widget.ncomponents = 2
-        self.send_signal(self.widget.Inputs.data, self.iris)
+        self.send_signal(self.widget.Inputs.data, data)
         self.wait_until_stop_blocking()
         self.widget._variance_ratio = np.array([0.5, 0.25, 0.2, 0.05])
         self.widget.commit.now()
@@ -225,12 +237,13 @@ def test_normalized_gives_correct_result(self, prepare_table):
 
         np.testing.assert_almost_equal(widget_result.X, pca_embedding)
 
-    def test_do_not_mask_features(self):
+    @with_dasktable
+    def test_do_not_mask_features(self, prepare_table):
         # the widget used to replace cached variables when creating the
         # components output (until 20170726)
-        data = Table("iris.tab")
+        data = prepare_table(self.iris)
+        ndata = data.copy()
         self.widget.set_data(data)
-        ndata = Table("iris.tab")
         self.assertEqual(data.domain[0], ndata.domain[0])
 
     def test_on_cut_changed(self):
@@ -248,27 +261,29 @@ def test_on_cut_changed(self):
         invalidate.assert_not_called()
         self.assertEqual(widget.ncomponents, 0)
 
-    def test_output_data(self):
+    @with_dasktable
+    def test_output_data(self, prepare_table):
         widget = self.widget
         widget.ncomponents = 2
-        domain = Domain(self.iris.domain.attributes[:3],
-                        self.iris.domain.class_var,
-                        self.iris.domain.attributes[3:])
-        iris = self.iris.transform(domain)
+        data = prepare_table(self.iris)
+        domain = Domain(data.domain.attributes[:3],
+                        data.domain.class_var,
+                        data.domain.attributes[3:])
+        iris = data.transform(domain)
         self.send_signal(widget.Inputs.data, iris)
         output = self.get_output(widget.Outputs.data)
         outdom = output.domain
         self.assertEqual(domain.attributes, outdom.attributes)
         self.assertEqual(domain.class_var, outdom.class_var)
         self.assertEqual(domain.metas, outdom.metas[:1])
         self.assertEqual(len(outdom.metas), 3)
-        np.testing.assert_equal(iris.X, output.X)
-        np.testing.assert_equal(iris.Y, output.Y)
+        self.assertTrue(np.all(iris.X == output.X))
+        self.assertTrue(np.all(iris.Y == output.Y))
         np.testing.assert_equal(iris.metas[:, 0], output.metas[:, 0])
 
         trans = self.get_output(widget.Outputs.transformed_data)
         self.assertEqual(trans.domain.attributes, outdom.metas[1:])
-        np.testing.assert_equal(trans.X, output.metas[:, 1:])
+        np.testing.assert_equal(trans.X, output.metas[:, 1:])  # dask
 
         self.send_signal(widget.Inputs.data, None)
         output = self.get_output(widget.Outputs.data)