sparse.Dot() test grad with mixed sparse/dense.

theano/gradient.py

@@ -1088,14 +1088,17 @@ def _populate_grad_dict(var_to_app_to_idx,
                 if len(input_grads) != len(inputs):
                     raise ValueError(("%s returned the wrong number of" +
                                       " gradient terms.") % str(node.op))
-
-                for ig, i in zip(input_grads, inputs):
-                    if (not isinstance(ig.type, DisconnectedType) and
-                        type(ig.type) != type(i.type)):
-                        raise ValueError(
-                            "%s returned the wrong type for gradient terms."
-                            " Sparse inputs must have sparse grads and dense"
-                            " inputs must have dense grad" % str(node.op))
+# We can not enforce this, as AdvancedSubtensor1 have an option to
+# return the sparse grad for optimization reason.
+
+                    #                for ig, i in zip(input_grads, inputs):
+#                    if (not isinstance(ig.type, (DisconnectedType, NullType)) and
+#                        type(ig.type) != type(i.type)):
+#                        raise ValueError(
+#                            "%s returned the wrong type for gradient terms."
+#                            " Sparse inputs must have sparse grads and dense"
+#                            " inputs must have dense grad. Got %s, expected %s" % (
+#                                str(node.op), ig.type, i.type))
 
             # must convert to list in case the op returns a tuple
             # we won't be able to post-process out the Nones if it does that

theano/sparse/basic.py

@@ -3709,11 +3709,15 @@ class Dot(gof.op.Op):
         raise NotImplementedError()
 
     def make_node(self, x, y):
-        dtype_out = scalar.upcast(x.type.dtype, y.type.dtype)
+        dtype_out = scalar.upcast(x.dtype, y.dtype)
 
         # Sparse dot product should have at least one sparse variable
         # as input. If the other one is not sparse, it has to be converted
         # into a tensor.
+        if isinstance(x, scipy.sparse.spmatrix):
+            x = as_sparse_variable(x)
+        if isinstance(y, scipy.sparse.spmatrix):
+            y = as_sparse_variable(y)
         x_is_sparse_var = _is_sparse_variable(x)
         y_is_sparse_var = _is_sparse_variable(y)
 

theano/sparse/tests/test_basic.py

@@ -1337,6 +1337,20 @@ class DotTests(utt.InferShapeTester):
                           dtype=intX)
         f(i, a)
 
+    def test_csr_dense_grad(self):
+
+        #shortcut: testing csc in float32, testing csr in float64
+
+        # allocate a random sparse matrix
+        spmat = sp.csr_matrix(random_lil((4, 3), 'float64', 3))
+
+        mat = numpy.asarray(numpy.random.randn(2, 4), 'float64')
+
+        def buildgraph_T(mat):
+            return Dot()(mat, spmat)
+
+        theano.tests.unittest_tools.verify_grad(buildgraph_T, [mat])
+
 
 class UsmmTests(unittest.TestCase):
     """ Test the Usmm and UsmmCscDense class and related optimization """