Merge pull request #1974 from janezd/fix-kmeans-fail

[FIX] KMeans: Fix crashes when underlying algorithm fails
(cherry picked from commit c80741b)

Author: astaric

Orange/widgets/unsupervised/owkmeans.py

@@ -2,7 +2,7 @@
 
 import numpy as np
 from AnyQt.QtWidgets import QGridLayout, QSizePolicy, QTableView
-from AnyQt.QtGui import QStandardItemModel, QStandardItem, QIntValidator
+from AnyQt.QtGui import QStandardItemModel, QStandardItem, QIntValidator, QBrush
 from AnyQt.QtCore import Qt, QTimer
 
 from Orange.clustering import KMeans
@@ -24,15 +24,18 @@ class OWKMeans(widget.OWWidget):
     outputs = [("Annotated Data", Table, widget.Default),
                ("Centroids", Table)]
 
+    class Error(widget.OWWidget.Error):
+        failed = widget.Msg("Clustering failed\nError: {}")
+
     INIT_KMEANS, INIT_RANDOM = range(2)
     INIT_METHODS = "Initialize with KMeans++", "Random initialization"
 
     SILHOUETTE, INTERCLUSTER, DISTANCES = range(3)
-    SCORING_METHODS = [("Silhouette", lambda km: km.silhouette, False),
+    SCORING_METHODS = [("Silhouette", lambda km: km.silhouette, False, True),
                        ("Inter-cluster distance",
-                        lambda km: km.inter_cluster, True),
+                        lambda km: km.inter_cluster, True, False),
                        ("Distance to centroids",
-                        lambda km: km.inertia, True)]
+                        lambda km: km.inertia, True, False)]
 
     OUTPUT_CLASS, OUTPUT_ATTRIBUTE, OUTPUT_META = range(3)
     OUTPUT_METHODS = ("Class", "Feature", "Meta")
@@ -150,7 +153,8 @@ def __init__(self):
         table.setSelectionMode(QTableView.SingleSelection)
         table.setSelectionBehavior(QTableView.SelectRows)
         table.verticalHeader().hide()
-        table.setItemDelegateForColumn(1, gui.TableBarItem(self))
+        self.bar_delegate = gui.ColoredBarItemDelegate(self, color=Qt.cyan)
+        table.setItemDelegateForColumn(1, self.bar_delegate)
         table.setModel(self.table_model)
         table.selectionModel().selectionChanged.connect(
             self.table_item_selected)
@@ -219,6 +223,7 @@ def run_optimization(self):
         try:
             self.controlArea.setDisabled(True)
             self.optimization_runs = []
+            error = ""
             if not self.check_data_size(self.k_from, self.Error):
                 return
             self.check_data_size(self.k_to, self.Warning)
@@ -231,7 +236,15 @@ def run_optimization(self):
                 for k in range(self.k_from, k_to + 1):
                     progress.advance()
                     kmeans.params["n_clusters"] = k
-                    self.optimization_runs.append((k, kmeans(self.data)))
+                    try:
+                        self.optimization_runs.append((k, kmeans(self.data)))
+                    except BaseException as exc:
+                        error = str(exc)
+                        self.optimization_runs.append((k, error))
+            if all(isinstance(score, str)
+                   for _, score in self.optimization_runs):
+                self.Error.failed(error)  # Report just the last error
+                self.optimization_runs = []
         finally:
             self.controlArea.setDisabled(False)
         self.show_results()
@@ -240,11 +253,15 @@ def run_optimization(self):
     def cluster(self):
         if not self.check_data_size(self.k, self.Error):
             return
-        self.km = KMeans(
-            n_clusters=self.k,
-            init=['random', 'k-means++'][self.smart_init],
-            n_init=self.n_init,
-            max_iter=self.max_iterations)(self.data)
+        try:
+            self.km = KMeans(
+                n_clusters=self.k,
+                init=['random', 'k-means++'][self.smart_init],
+                n_init=self.n_init,
+                max_iter=self.max_iterations)(self.data)
+        except BaseException as exc:
+            self.Error.failed(str(exc))
+            self.km = None
         self.send_data()
 
     def run(self):
@@ -260,40 +277,55 @@ def commit(self):
         self.run()
 
     def show_results(self):
-        minimize = self.SCORING_METHODS[self.scoring][2]
-        k_scores = [(k, self.SCORING_METHODS[self.scoring][1](run)) for
-                    k, run in self.optimization_runs]
-        scores = list(zip(*k_scores))[1]
-        if minimize:
-            best_score, worst_score = min(scores), max(scores)
+        _, scoring_method, minimize, normal = self.SCORING_METHODS[self.scoring]
+        k_scores = [(k,
+                     scoring_method(run) if not isinstance(run, str) else run)
+                    for k, run in self.optimization_runs]
+        scores = [score for _, score in k_scores if not isinstance(score, str)]
+
+        min_score, max_score = min(scores, default=0), max(scores, default=1)
+        best_score = min_score if minimize else max_score
+        if normal:
+            min_score, max_score = 0, 1
+            nplaces = 3
         else:
-            best_score, worst_score = max(scores), min(scores)
+            nplaces = min(5, np.floor(abs(math.log(max(max_score, 1e-10)))) + 2)
+        score_span = (max_score - min_score) or 1
+        self.bar_delegate.scale = (min_score, max_score)
+        self.bar_delegate.float_fmt = "%%.%if" % int(nplaces)
 
-        best_run = scores.index(best_score)
-        score_span = (best_score - worst_score) or 1
-        max_score = max(scores)
-        nplaces = min(5, np.floor(abs(math.log(max(max_score, 1e-10)))) + 2)
-        fmt = "{{:.{}f}}".format(int(nplaces))
         model = self.table_model
         model.setRowCount(len(k_scores))
+        no_selection = True
         for i, (k, score) in enumerate(k_scores):
-            item = model.item(i, 0)
-            if item is None:
-                item = QStandardItem()
-            item.setData(k, Qt.DisplayRole)
-            item.setTextAlignment(Qt.AlignCenter)
-            model.setItem(i, 0, item)
-            item = model.item(i, 1)
-            if item is None:
-                item = QStandardItem()
-            item.setData(fmt.format(score) if not np.isnan(score) else 'out-of-memory error',
-                         Qt.DisplayRole)
-            bar_ratio = 0.95 * (score - worst_score) / score_span
-            item.setData(bar_ratio, gui.TableBarItem.BarRole)
+            item0 = model.item(i, 0) or QStandardItem()
+            item0.setData(k, Qt.DisplayRole)
+            item0.setTextAlignment(Qt.AlignCenter)
+            model.setItem(i, 0, item0)
+            item = model.item(i, 1) or QStandardItem()
+            if not isinstance(score, str):
+                item.setData(score, Qt.DisplayRole)
+                item.setData(None, Qt.ToolTipRole)
+                bar_ratio = 0.95 * (score - min_score) / score_span
+                item.setData(bar_ratio, gui.BarRatioRole)
+                if no_selection and score == best_score:
+                    self.table_view.selectRow(i)
+                    no_selection = False
+                color = Qt.black
+                flags = Qt.ItemIsEnabled | Qt.ItemIsSelectable
+            else:
+                item.setData("clustering failed", Qt.DisplayRole)
+                item.setData(score, Qt.ToolTipRole)
+                item.setData(None, gui.BarRatioRole)
+                color = Qt.gray
+                flags = Qt.NoItemFlags
+            item0.setData(QBrush(color), Qt.ForegroundRole)
+            item0.setFlags(flags)
+            item.setData(QBrush(color), Qt.ForegroundRole)
+            item.setFlags(flags)
             model.setItem(i, 1, item)
         self.table_view.resizeRowsToContents()
 
-        self.table_view.selectRow(best_run)
         self.table_view.show()
         if minimize:
             self.table_box.setTitle("Scoring (smaller is better)")
@@ -314,13 +346,12 @@ def selected_row(self):
     def table_item_selected(self):
         row = self.selected_row()
         if row is not None:
-            self.send_data(row)
+            self.send_data()
 
-    def send_data(self, row=None):
+    def send_data(self):
         if self.optimize_k:
-            if row is None:
-                row = self.selected_row()
-            km = self.optimization_runs[row][1]
+            row = self.selected_row() if self.optimization_runs else None
+            km = self.optimization_runs[row][1] if row is not None else None
         else:
             km = self.km
         if not self.data or not km:
@@ -356,6 +387,8 @@ def send_data(self, row=None):
     def set_data(self, data):
         self.data = data
         if data is None:
+            self.Error.clear()
+            self.Warning.clear()
             self.table_model.setRowCount(0)
             self.send("Annotated Data", None)
             self.send("Centroids", None)

Orange/widgets/unsupervised/tests/test_owkmeans.py

@@ -1,14 +1,17 @@
+from unittest.mock import patch
+
 from AnyQt.QtWidgets import QRadioButton
 
 from Orange.widgets.tests.base import WidgetTest
 from Orange.widgets.unsupervised.owkmeans import OWKMeans
+import Orange.clustering
 
 from Orange.data import Table
 class TestOWKMeans(WidgetTest):
 
     def setUp(self):
-        self.widget = self.create_widget(OWKMeans,
-                                         stored_settings={"auto_apply": False})
+        self.widget = self.create_widget(
+            OWKMeans, stored_settings={"auto_apply": False})  # type: OWKMeans
         self.iris = Table("iris")
 
     def test_optimization_report_display(self):
@@ -32,3 +35,59 @@ def test_data_on_output(self):
         self.send_signal("Data", None)
         # removing data should have cleared the output
         self.assertEqual(self.widget.data, None)
+
+    class KMeansFail(Orange.clustering.KMeans):
+        fail_on = set()
+
+        def fit(self, *args):
+            # when not optimizing, params is empty?!
+            k = self.params.get("n_clusters", 3)
+            if k in self.fail_on:
+                raise ValueError("k={} fails".format(k))
+            return super().fit(*args)
+
+    @patch("Orange.widgets.unsupervised.owkmeans.KMeans", new=KMeansFail)
+    def test_optimization_fails(self):
+        widget = self.widget
+        widget.k_from = 3
+        widget.k_to = 8
+        widget.scoring = 0
+        widget.optimize_k = True
+
+        self.KMeansFail.fail_on = {3, 5, 7}
+        self.send_signal("Data", self.iris)
+        self.assertIsInstance(widget.optimization_runs[0][1], str)
+        self.assertIsInstance(widget.optimization_runs[2][1], str)
+        self.assertIsInstance(widget.optimization_runs[4][1], str)
+        self.assertNotIsInstance(widget.optimization_runs[1][1], str)
+        self.assertNotIsInstance(widget.optimization_runs[3][1], str)
+        self.assertNotIsInstance(widget.optimization_runs[5][1], str)
+        self.assertFalse(widget.Error.failed.is_shown())
+        self.assertEqual(widget.selected_row(), 1)
+        self.assertIsNotNone(self.get_output("Annotated Data"))
+
+        self.KMeansFail.fail_on = set(range(3, 9))
+        widget.run()
+        self.assertTrue(widget.Error.failed.is_shown())
+        self.assertEqual(widget.optimization_runs, [])
+        self.assertIsNone(self.get_output("Annotated Data"))
+
+        self.KMeansFail.fail_on = set()
+        widget.run()
+        self.assertFalse(widget.Error.failed.is_shown())
+        self.assertEqual(widget.selected_row(), 0)
+        self.assertIsNotNone(self.get_output("Annotated Data"))
+
+    @patch("Orange.widgets.unsupervised.owkmeans.KMeans", new=KMeansFail)
+    def test_run_fails(self):
+        self.widget.k = 3
+        self.widget.optimize_k = False
+        self.KMeansFail.fail_on = {3}
+        self.send_signal("Data", self.iris)
+        self.assertTrue(self.widget.Error.failed.is_shown())
+        self.assertIsNone(self.get_output("Annotated Data"))
+
+        self.KMeansFail.fail_on = set()
+        self.widget.run()
+        self.assertFalse(self.widget.Error.failed.is_shown())
+        self.assertIsNotNone(self.get_output("Annotated Data"))