Fix runtime crash with the local_subtensor_of_alloc optimization when the input…

NEWS.txt

@@ -168,6 +168,7 @@ Crashes fixed:
  * Support for OSX Enthought Python Distribution 7.x. (Graham Taylor, Olivier)
  * When the subtensor inputs had 0 dimensions and the outputs 0 dimensions. (Frederic)
  * Crash when the step to subtensor was not 1 in conjunction with some optimization. (Frederic, reported by Olivier Chapelle)
+ * Runtime crash related to an optimization with subtensor of alloc (reported by Razvan, fixed by Frederic)
  * Fix dot22scalar cast of integer scalars (Justin Bayer, Frédéric, Olivier)
  * Fix runtime crash in gemm, dot22. FB
  * Fix on 32bits computer: make sure all shape are int64.(Olivier)

theano/tensor/opt.py

@@ -1842,7 +1842,13 @@ def local_subtensor_of_alloc(node):
         # If val was not copied over that dim,
         # we need to take the appropriate subtensor on it.
         if i >= n_added_dims:
-            val_slices.append(sl)
+            # We check that the corresponding val dimensions was
+            # not a broadcasted dimensions.
+            if (val.type.ndim > (i - n_added_dims) and
+                val.type.broadcastable[i - n_added_dims]):
+                val_slices.append(slice(None))
+            else:
+                val_slices.append(sl)
 
         csl, _ = T.get_canonical_form_slice(sl, dim)
         if type(csl) is not slice:

theano/tensor/tests/test_opt.py

@@ -2073,18 +2073,6 @@ class Test_alloc_zero(unittest.TestCase):
 
 def test_local_subtensor_of_alloc():
     x = tensor.matrix('x')
-    y = tensor.vector('y')
-
-    # The rows of yx are copies of y
-    yx = tensor.alloc(y, x.shape[0], x.shape[1])
-
-    # Slice of each row
-    z_mat = yx[:, 3:]
-    assert z_mat.ndim == 2
-
-    # Only one column
-    z_vec = yx[:, 3]
-    assert z_vec.ndim == 1
 
     # DebugMode should detect if something goes wrong.
     # test shape combination of odd and event shape.
@@ -2093,22 +2081,35 @@ def test_local_subtensor_of_alloc():
         xval = numpy.zeros(shape, dtype=config.floatX)
         yval = numpy.arange(shape[1], dtype=config.floatX)
 
-        for slices in [
-            # results are vector
-            (slice(None), 3),
-            (2, slice(None)),
-            # results are matrix
-            (slice(None), slice(3, None)),
-            (slice(3, None), ),
-            (slice(3, None), slice(3, None)),
-            (slice(1, 3), slice(None, -1)),
-            (slice(None, None, 2)),
-            (slice(1, None, 2)),
-            ]:
-            z = yx.__getitem__(slices)
-            f = theano.function([x, y], z)
-            val = f(xval, yval)
-            assert xval.__getitem__(slices).shape == val.shape
+        for y in [theano.shared(yval), tensor.constant([1.])]:
+
+            # The rows of yx are copies of y
+            yx = tensor.alloc(y, x.shape[0], x.shape[1])
+
+            # Slice of each row
+            z_mat = yx[:, 3:]
+            assert z_mat.ndim == 2
+
+            # Only one column
+            z_vec = yx[:, 3]
+            assert z_vec.ndim == 1
+
+            for slices in [
+                # results are vector
+                (slice(None), 3),
+                (2, slice(None)),
+                # results are matrix
+                (slice(None), slice(3, None)),
+                (slice(3, None), ),
+                (slice(3, None), slice(3, None)),
+                (slice(1, 3), slice(None, -1)),
+                (slice(None, None, 2)),
+                (slice(1, None, 2)),
+                ]:
+                z = yx.__getitem__(slices)
+                f = theano.function([x], z)
+                val = f(xval)
+                assert xval.__getitem__(slices).shape == val.shape
 
 
 def test_local_fill_useless():