Merge pull request #246 from corochann/features_indexer_with_length0

support features indexer with length 0

Author: mottodora

chainer_chemistry/dataset/indexer.py

@@ -159,16 +159,24 @@ def _extract_feature(self, data_index, j):
                 res = [self.extract_feature(i, j) for i in
                        six.moves.range(current, stop, step)]
         elif isinstance(data_index, (list, numpy.ndarray)):
-            if isinstance(data_index[0], (bool, numpy.bool, numpy.bool_)):
-                # Access by bool flag list
-                if len(data_index) != self.dataset_length:
-                    raise ValueError('Feature index wrong length {} instead of'
-                                     ' {}'.format(len(data_index),
-                                                  self.dataset_length))
-                data_index = numpy.argwhere(data_index).ravel()
-
-            res = [self.extract_feature(i, j) for i in data_index]
+            if len(data_index) == 0:
+                try:
+                    # HACKING
+                    return self.extract_feature_by_slice(slice(0, 0, 1), j)
+                except ExtractBySliceNotSupportedError:
+                    res = []
+            else:
+                if isinstance(data_index[0], (bool, numpy.bool, numpy.bool_)):
+                    # Access by bool flag list
+                    if len(data_index) != self.dataset_length:
+                        raise ValueError('Feature index wrong length {} instead of'
+                                         ' {}'.format(len(data_index),
+                                                      self.dataset_length))
+                    data_index = numpy.argwhere(data_index).ravel()
+
+                res = [self.extract_feature(i, j) for i in data_index]
         else:
+            # `data_index` is expected to be `int`
             return self.extract_feature(data_index, j)
         try:
             feature = numpy.asarray(res)

tests/dataset_tests/test_numpy_tuple_feature_indexer.py

@@ -26,20 +26,32 @@ class TestNumpyTupleDatasetFeatureIndexer(object):
     def test_feature_length(self, indexer):
         assert indexer.features_length() == 3
 
-    @pytest.mark.parametrize('slice_index', [0, 1, slice(0, 2, None)])
+    @pytest.mark.parametrize('slice_index', [
+        0, 1, slice(0, 2, None), slice(0, 0, None)])
     @pytest.mark.parametrize('j', [0, 1])
     def test_extract_feature_by_slice(self, indexer, data, slice_index, j):
         numpy.testing.assert_array_equal(
             indexer.extract_feature_by_slice(slice_index, j),
             data[j][slice_index])
+        # indexer's __getitem__ should call `extract_feature_by_slice` method,
+        # result should be same with above.
+        numpy.testing.assert_array_equal(
+            indexer[slice_index, j],
+            data[j][slice_index])
 
-    @pytest.mark.parametrize('ndarray_index', [numpy.asarray([0, 1]),
-                                               numpy.asarray([1])])
+    @pytest.mark.parametrize('ndarray_index', [
+        numpy.asarray([0, 1]), numpy.asarray([1]),
+        numpy.asarray([], dtype=numpy.int32)])
     @pytest.mark.parametrize('j', [0, 1])
     def test_extract_feature_by_ndarray(self, indexer, data, ndarray_index, j):
         numpy.testing.assert_array_equal(
             indexer.extract_feature_by_slice(ndarray_index, j),
             data[j][ndarray_index])
+        # indexer's __getitem__ should call `extract_feature_by_slice` method,
+        # result should be same with above.
+        numpy.testing.assert_array_equal(
+            indexer[ndarray_index, j],
+            data[j][ndarray_index])
 
 
 if __name__ == '__main__':

tests/datasets_tests/test_numpy_tuple_dataset.py

@@ -55,8 +55,9 @@ def test_get_item_slice_index(self, data, index):
             for a, e in six.moves.zip(tuple_a, tuple_e):
                 numpy.testing.assert_array_equal(a, e)
 
-    @pytest.mark.parametrize('index', [numpy.asarray([2, 0]),
-                                       numpy.asarray([1])])
+    @pytest.mark.parametrize('index', [
+        numpy.asarray([2, 0]), numpy.asarray([1]),
+        numpy.asarray([], dtype=numpy.int32)])
     def test_get_item_ndarray_index(self, long_data, index):
         dataset = NumpyTupleDataset(*long_data)
         actual = dataset[index]