Add regular grad to sp_sum.

theano/sparse/basic.py

@@ -1371,8 +1371,10 @@ class SpSum(gof.op.Op):
     :note:
     - The grad implementation is controlled with the `sparse_grad`
       parameter. `True` will provide a structured grad and `False`
-      will provide a regular grad.
-    - This op does not return a sparse matrix.
+      will provide a regular grad. For both choice, the grad
+      return a sparse matrix having the same format as `x`.
+    - This op does not return a sparse matrix, but a dense tensor
+      matrix.
     """
 
     def __init__(self, axis=None, sparse_grad=False):
@@ -1414,6 +1416,9 @@ class SpSum(gof.op.Op):
             z[0] = numpy.asarray(x.sum(self.axis)).ravel()
 
     def grad(self, (x,), (gz,)):
+        if x.dtype not in continuous_dtypes:
+            return [None]
+
         if self.structured:
             if self.axis is None:
                 r = gz * theano.sparse.sp_ones_like(x)
@@ -1424,8 +1429,21 @@ class SpSum(gof.op.Op):
             else:
                 raise ValueError('Illegal value for self.axis.')
         else:
-            # TODO
-            raise NotImplementedError()
+            o_format = x.format
+            x = dense_from_sparse(x)
+            if _is_sparse_variable(gz):
+                gz = dense_from_sparse(gz)
+            if self.axis is None:
+                r = tensor.second(x, gz)
+            else:
+                ones = tensor.ones_like(x)
+                if self.axis == 0:
+                    r = tensor.addbroadcast(gz.dimshuffle('x', 0), 0) * ones
+                elif self.axis == 1:
+                    r = tensor.addbroadcast(gz.dimshuffle(0, 'x'), 1) * ones
+                else:
+                    raise ValueError('Illegal value for self.axis.')
+            r = SparseFromDense(o_format)(r)
         return [r]
 
     def infer_shape(self, node, shapes):

theano/sparse/tests/test_basic.py

@@ -96,9 +96,6 @@ def sparse_random_inputs(format, shape, n=1, out_dtype=None, p=0.5):
     assert len(shape) == 2
     assert out_dtype in sparse.all_dtypes
 
-    variable = [getattr(theano.sparse, format + '_matrix')(dtype=out_dtype)
-                for k in range(n)]
-
     def _rand():
         where = numpy.random.binomial(1, p, size=shape).astype('int8')
 
@@ -106,9 +103,10 @@ def sparse_random_inputs(format, shape, n=1, out_dtype=None, p=0.5):
             value = numpy.random.randint(20, size=shape).astype(out_dtype)
         else:
             value = numpy.random.random(shape)
-
         return where * value
 
+    variable = [getattr(theano.sparse, format + '_matrix')(dtype=out_dtype)
+                for k in range(n)]
     data = [getattr(scipy.sparse, format + '_matrix')(_rand())
             for k in range(n)]
 
@@ -1410,7 +1408,7 @@ class SpSumTester(utt.InferShapeTester):
     def test_grad(self):
         for format in sparse.sparse_formats:
             for axis in self.possible_axis:
-                for struct in [True]:
+                for struct in [True, False]:
                     variable, data = sparse_random_inputs(format,
                                                           shape=(10, 10))
                     verify_grad_sparse(
@@ -1811,8 +1809,8 @@ def _hv_switch(op, expected_function):
     :Parameters:
     - `op`: HStack or VStack class.
     - `expected_function`: function from scipy for comparaison.
-
     """
+
     class XStackTester(_HVStackTester):
         op_class = op