ENH: Allow genfromtxt to unpack structured arrays

Previously, `genfromtxt` failed to transpose its output when `unpack=True` and `dtype` was structured (or `dtype=None` and the inferred `dtype` was structured). This patch resolves the issue by returning a list of arrays, as in `loadtxt`. Includes tests and updates the docstring to match `loadtxt`. Fixes #4341
numpy · rossbar · Sep 11, 2020 · Jun 20, 2020 · Jul 14, 2020 · Jul 14, 2020
commit 189ad50bb5b3f3dbcd2eed269e916f3a02b9aee3
diff --git a/numpy/lib/npyio.py b/numpy/lib/npyio.py
@@ -837,8 +837,9 @@ def loadtxt(fname, dtype=float, comments='#', delimiter=None,
             fourth column the same way as ``usecols = (3,)`` would.
     unpack : bool, optional
         If True, the returned array is transposed, so that arguments may be
-        unpacked using ``x, y, z = loadtxt(...)``.  When used with a structured
-            data-type, arrays are returned for each field.  Default is False.
+        unpacked using ``x, y, z = loadtxt(...)``.  When used with a
+        structured data-type, arrays are returned for each field.
+        Default is False.
     ndmin : int, optional
         The returned array will have at least `ndmin` dimensions.
         Otherwise mono-dimensional axes will be squeezed.
@@ -1631,7 +1632,9 @@ def genfromtxt(fname, dtype=float, comments='#', delimiter=None,
         If 'lower', field names are converted to lower case.
     unpack : bool, optional
         If True, the returned array is transposed, so that arguments may be
-        unpacked using ``x, y, z = loadtxt(...)``
+        unpacked using ``x, y, z = genfromtxt(...)``.  When used with a
+        structured data-type, arrays are returned for each field.
+        Default is False.
     usemask : bool, optional
         If True, return a masked array.
         If False, return a regular array.
@@ -2242,9 +2245,18 @@ def encode_unicode_cols(row_tup):
     if usemask:
         output = output.view(MaskedArray)
         output._mask = outputmask
+    output = np.squeeze(output)
     if unpack:
-        return output.squeeze().T
-    return output.squeeze()
+        if names is None:
+            return output.T
+        elif len(names) == 1:
+            # squeeze single-name dtypes too
+            return output[names[0]]
+        else:
+            # For structured arrays with multiple fields,
+            # return an array for each field.
+            return [output[field] for field in names]
+    return output
 
 
 def ndfromtxt(fname, **kwargs):

diff --git a/numpy/lib/tests/test_io.py b/numpy/lib/tests/test_io.py
@@ -1011,7 +1011,7 @@ def test_empty_field_after_tab(self):
         a = np.array([b'start ', b'  ', b''])
         assert_array_equal(x['comment'], a)
 
-    def test_structure_unpack(self):
+    def test_unpack_structured(self):
         txt = TextIO("M 21 72\nF 35 58")
         dt = {'names': ('a', 'b', 'c'), 'formats': ('|S1', '<i4', '<f4')}
         a, b, c = np.loadtxt(txt, dtype=dt, unpack=True)
@@ -2343,6 +2343,51 @@ def test_auto_dtype_largeint(self):
         assert_equal(test['f1'], 17179869184)
         assert_equal(test['f2'], 1024)
 
+    def test_unpack_structured(self):
+        # Regression test for gh-4341
+        # Unpacking should work on structured arrays
+        txt = TextIO("M 21 72\nF 35 58")
+        dt = {'names': ('a', 'b', 'c'), 'formats': ('|S1', '<i4', '<f4')}
+        a, b, c = np.genfromtxt(txt, dtype=dt, unpack=True)
+        assert_(a.dtype.str == '|S1')
+        assert_(b.dtype.str == '<i4')
+        assert_(c.dtype.str == '<f4')
+        assert_array_equal(a, np.array([b'M', b'F']))
+        assert_array_equal(b, np.array([21, 35]))
+        assert_array_equal(c, np.array([72.,  58.]))
+
+    def test_unpack_auto_dtype(self):
+        # Regression test for gh-4341
+        # Unpacking should work when dtype=None
+        txt = TextIO("M 21 72.\nF 35 58.")
+        expected = (np.array(["M", "F"]), np.array([21, 35]), np.array([72., 58.]))
+        test = np.genfromtxt(txt, dtype=None, unpack=True, encoding=None)
+        for arr, result in zip(expected, test):
+            assert_array_equal(arr, result)
+            assert_equal(arr.dtype, result.dtype)
+
+    def test_unpack_single_name(self):
+        # Regression test for gh-4341
+        # Unpacking should work when structured dtype has only one field
+        txt = TextIO("21\n35")
+        dt = {'names': ('a',), 'formats': ('<i4',)}
+        expected = np.array([21, 35], dtype=np.int32)
+        test = np.genfromtxt(txt, dtype=dt, unpack=True)
+        assert_array_equal(expected, test)
+        assert_equal(expected.dtype, test.dtype)
+
+    def test_squeeze_scalar(self):
+        # Regression test for gh-4341
+        # Unpacking a scalar should give zero-dim output,
+        # even if dtype is structured
+        txt = TextIO("1")
+        dt = {'names': ('a',), 'formats': ('<i4',)}
+        expected = np.array((1,), dtype=np.int32)
+        test = np.genfromtxt(txt, dtype=dt, unpack=True)
+        assert_array_equal(expected, test)
+        assert_equal((), test.shape)
+        assert_equal(expected.dtype, test.dtype)
+
 
 class TestPathUsage:
     # Test that pathlib.Path can be used