Merge pull request biolab#2936 from cemsbr/oversampling

Allow oversampling in Data Sampler widget

Orange/widgets/data/owdatasampler.py

@@ -7,7 +7,7 @@
 import numpy as np
 import sklearn.model_selection as skl
 
-from Orange.widgets import widget, gui
+from Orange.widgets import gui
 from Orange.widgets.settings import Setting
 from Orange.data import Table
 from Orange.data.sql.table import SqlTable
@@ -24,6 +24,8 @@ class OWDataSampler(OWWidget):
     category = "Data"
     keywords = ["data", "sample"]
 
+    _MAX_SAMPLE_SIZE = 2 ** 31 - 1
+
     class Inputs:
         data = Input("Data", Table)
 
@@ -52,6 +54,7 @@ class Outputs:
 
     class Warning(OWWidget.Warning):
         could_not_stratify = Msg("Stratification failed\n{}")
+        bigger_sample = Msg('Sample is bigger than input')
 
     class Error(OWWidget.Error):
         too_many_folds = Msg("Number of folds exceeds data size")
@@ -74,10 +77,10 @@ def __init__(self):
                                     callback=self.sampling_type_changed)
 
         def set_sampling_type(i):
-            def f():
+            def set_sampling_type_i():
                 self.sampling_type = i
                 self.sampling_type_changed()
-            return f
+            return set_sampling_type_i
 
         gui.appendRadioButton(sampling, "Fixed proportion of data:")
         self.sampleSizePercentageSlider = gui.hSlider(
@@ -91,8 +94,9 @@ def f():
         ibox = gui.indentedBox(sampling)
         self.sampleSizeSpin = gui.spin(
             ibox, self, "sampleSizeNumber", label="Instances: ",
-            minv=1, maxv=2 ** 31 - 1,
-            callback=set_sampling_type(self.FixedSize))
+            minv=1, maxv=self._MAX_SAMPLE_SIZE,
+            callback=set_sampling_type(self.FixedSize),
+            controlWidth=90)
         gui.checkBox(
             ibox, self, "replacement", "Sample with replacement",
             callback=set_sampling_type(self.FixedSize),
@@ -132,7 +136,6 @@ def f():
         spin.setSuffix(" %")
         self.sql_box.setVisible(False)
 
-
         self.options_box = gui.vBox(self.controlArea, "Options")
         self.cb_seed = gui.checkBox(
             self.options_box, self, "use_seed",
@@ -162,6 +165,7 @@ def fold_changed(self):
         self.sampling_type = self.CrossValidation
 
     def settings_changed(self):
+        self._update_sample_max_size()
         self.indices = None
 
     @Inputs.data
@@ -179,7 +183,7 @@ def set_data(self, dataset):
                     ('~', dataset.approx_len()) if sql else
                     ('', len(dataset)))))
             if not sql:
-                self.sampleSizeSpin.setMaximum(len(dataset))
+                self._update_sample_max_size()
                 self.updateindices()
         else:
             self.dataInfoLabel.setText('No data on input.')
@@ -188,6 +192,13 @@ def set_data(self, dataset):
             self.clear_messages()
         self.commit()
 
+    def _update_sample_max_size(self):
+        """Limit number of instances to input size unless using replacement."""
+        if not self.data or self.replacement:
+            self.sampleSizeSpin.setMaximum(self._MAX_SAMPLE_SIZE)
+        else:
+            self.sampleSizeSpin.setMaximum(len(self.data))
+
     def commit(self):
         if self.data is None:
             sample = other = None
@@ -231,6 +242,7 @@ def commit(self):
 
     def updateindices(self):
         self.Error.clear()
+        self.Warning.clear()
         repl = True
         data_length = len(self.data)
         num_classes = len(self.data.domain.class_var.values) \
@@ -260,8 +272,14 @@ def updateindices(self):
             self.indices = None
             return
 
+        # By the above, we can safely assume there is data
+        if self.sampling_type == self.FixedSize and repl and size and \
+                size > len(self.data):
+            # This should only be possible when using replacement
+            self.Warning.bigger_sample()
+
         stratified = (self.stratify and
-                      type(self.data) == Table and
+                      isinstance(self.data, Table) and
                       self.data.domain.has_discrete_class)
         try:
             self.indices = self.sample(data_length, size, stratified)
@@ -277,7 +295,8 @@ def sample(self, data_length, size, stratified):
                 random_state=rnd)
         elif self.sampling_type == self.FixedProportion:
             self.indice_gen = SampleRandomP(
-                self.sampleSizePercentage / 100, stratified=stratified, random_state=rnd)
+                self.sampleSizePercentage / 100, stratified=stratified,
+                random_state=rnd)
         elif self.sampling_type == self.Bootstrap:
             self.indice_gen = SampleBootstrap(data_length, random_state=rnd)
         else:
@@ -323,7 +342,8 @@ def __init__(self, folds=10, stratified=False, random_state=None):
         Args:
             folds (int): Number of folds
             stratified (bool): Return stratified indices (if applicable).
-            random_state (Random): An initial state for replicable random behavior
+            random_state (Random): An initial state for replicable random
+            behavior
 
         Returns:
             tuple-of-arrays: A tuple of array indices one for each fold.
@@ -349,7 +369,7 @@ def __call__(self, table):
 
 class SampleRandomN(Reprable):
     def __init__(self, n=0, stratified=False, replace=False,
-                    random_state=None):
+                 random_state=None):
         self.n = n
         self.stratified = stratified
         self.replace = replace

Orange/widgets/data/tests/test_owdatasampler.py

@@ -81,6 +81,38 @@ def test_no_intersection_in_outputs(self):
                     self.assertEqual(len(self.zoo), len(sample) + len(other))
                     self.assertNoIntersection(sample, other)
 
+    def test_bigger_size_with_replacement(self):
+        """Allow bigger output without replacement."""
+        self.send_signal('Data', self.iris[:2])
+        sample_size = self.set_fixed_sample_size(3, with_replacement=True)
+        self.assertEqual(3, sample_size, 'Should be able to set a bigger size '
+                         'with replacement')
+
+    def test_bigger_size_without_replacement(self):
+        """Lower output samples to match input's without replacement."""
+        self.send_signal('Data', self.iris[:2])
+        sample_size = self.set_fixed_sample_size(3)
+        self.assertEqual(2, sample_size)
+
+    def test_bigger_output_warning(self):
+        """Should warn when sample size is bigger than input."""
+        self.send_signal('Data', self.iris[:2])
+        self.set_fixed_sample_size(3, with_replacement=True)
+        self.assertTrue(self.widget.Warning.bigger_sample.is_shown())
+
+    def set_fixed_sample_size(self, sample_size, with_replacement=False):
+        """Set fixed sample size and return the number of gui spin.
+
+        Return the actual number in gui so we can check whether it is different
+        from sample_size. The number can be changed depending on the spin
+        maximum value.
+        """
+        self.select_sampling_type(self.widget.FixedSize)
+        self.widget.controls.replacement.setChecked(with_replacement)
+        self.widget.sampleSizeSpin.setValue(sample_size)
+        self.widget.commit()
+        return self.widget.sampleSizeSpin.value()
+
     def assertNoIntersection(self, sample, other):
         for inst in sample:
             self.assertNotIn(inst, other)

doc/visual-programming/source/widgets/data/datasampler.rst

@@ -30,7 +30,8 @@ provided and *Sample Data* is pressed.
    -  **Fixed sample size** returns a selected number of data instances
       with a chance to set *Sample with replacement*, which always samples
       from the entire dataset (does not subtract instances already in
-      the subset)
+      the subset). With replacement, you can generate more instances than
+      available in the input dataset.
    -  `Cross Validation <https://en.wikipedia.org/wiki/Cross-validation_(statistics)>`_
       partitions data instances into complementary subsets, where you can
       select the number of folds (subsets) and which fold you want to