Merge pull request #1691 from SuperElectric/sparse_stack_fix

theano/sparse/basic.py

@@ -1820,7 +1820,7 @@ class AddSD(gof.op.Op):
                     pos = row * Yi + col * Yj;
                     zdata[pos] = ydata[pos] + data[ind];
                   }
-                 } 
+                 }
                 }
              """ % dict(locals(), **sub)
         return code
@@ -2302,7 +2302,7 @@ def hstack(blocks, format=None, dtype=None):
 
     blocks = [as_sparse_variable(i) for i in blocks]
     if dtype is None:
-        dtype = theano.scalar.upcast([i.dtype for i in blocks])
+        dtype = theano.scalar.upcast(*[i.dtype for i in blocks])
     return HStack(format=format, dtype=dtype)(*blocks)
 
 
@@ -2378,7 +2378,7 @@ def vstack(blocks, format=None, dtype=None):
 
     blocks = [as_sparse_variable(i) for i in blocks]
     if dtype is None:
-        dtype = theano.scalar.upcast([i.dtype for i in blocks])
+        dtype = theano.scalar.upcast(*[i.dtype for i in blocks])
     return VStack(format=format, dtype=dtype)(*blocks)
 
 

theano/sparse/tests/test_basic.py

@@ -1950,7 +1950,7 @@ class Test_getitem(unittest.TestCase):
             assert r1.shape == t1.shape
             assert numpy.all(t1.toarray() == r1.toarray())
 
-            """"
+            """
             Important: based on a discussion with both Fred and James
             The following indexing methods is not supported because the rval
             would be a sparse matrix rather than a sparse vector, which is a
@@ -2463,6 +2463,35 @@ def elemwise_checker(op, expected_f, gap=None, test_dtypes=None,
 
     return Tester
 
+def test_hstack_vstack():
+    """
+    Tests sparse.hstack and sparse.vstack (as opposed to the HStack and VStack
+    classes that they wrap).
+    """
+
+    def make_block(dtype):
+        return theano.sparse.csr_matrix(name="%s block" % dtype,
+                                        dtype=dtype)
+
+    def get_expected_dtype(blocks, to_dtype):
+        if to_dtype is None:
+            block_dtypes = tuple(b.dtype for b in blocks)
+            return theano.scalar.upcast(*block_dtypes)
+        else:
+            return to_dtype
+
+    # a deliberately weird mix of dtypes to stack
+    dtypes = ('complex128', theano.config.floatX)
+
+    blocks = [make_block(dtype) for dtype in dtypes]
+
+    for stack_dimension, stack_function in enumerate((theano.sparse.vstack,
+                                                      theano.sparse.hstack)):
+
+        for to_dtype in (None, ) + dtypes:
+            stacked_blocks = stack_function(blocks, dtype=to_dtype)
+            expected_dtype = get_expected_dtype(blocks, to_dtype)
+            assert stacked_blocks.dtype == expected_dtype
 
 def structure_function(f, index=0):
     """Decorator to structure a function wich