[BUG] Important bug fix in an optimization.

theano/configdefaults.py

@@ -422,6 +422,15 @@ AddConfigVar('warn.signal_conv2d_interface',
              BoolParam(warn_default('0.7')),
              in_c_key=False)
 
+AddConfigVar('warn.reduce_join',
+             ('Your current code is fine, but Theano versions '
+              'prior to 0.7 (or this development version) '
+              'might have given an incorrect result. '
+              'To disable this warning, set the Theano flag '
+              'warn.reduce_join to False.'),
+             BoolParam(warn_default('0.7')),
+             in_c_key=False)
+
 AddConfigVar('compute_test_value',
         ("If 'True', Theano will run each op at graph build time, using "
          "Constants, SharedVariables and the tag 'test_value' as inputs "

theano/tensor/opt.py

@@ -3469,10 +3469,14 @@ ALL_REDUCE = [T.elemwise.CAReduce, T.elemwise.All, T.elemwise.Any,
 @register_uncanonicalize  # Needed for MaxAndArgmax -> CAReduce
 @gof.local_optimizer(ALL_REDUCE)
 def local_reduce_join(node):
-    """Reduce{scalar.op}(Join(a, b), axis=0) -> Elemwise{scalar.op}(a, b)
+    """Reduce{scalar.op}(Join(axis=0, a, b), axis=0) -> Elemwise{scalar.op}(a, b)
 
     :note: supported scalar.op are Maximum, Mimimum in some cases and
-    Add and Mul in all cases.
+        Add and Mul in all cases.
+
+    :note: Currently we must reduce on axis 0. It is probably
+        extensible to the case where we join and reduce on the same
+        set of axis.
 
     """
     if (isinstance(node.op, T.CAReduce) and
@@ -3483,6 +3487,21 @@ def local_reduce_join(node):
         if T.extract_constant(join.inputs[0]) != 0:
             return
 
+        if len(node.op.axis) != 1 or 0 not in node.op.axis:
+            if theano.config.warn.reduce_join:
+                _logger.warn((
+                    'Your current code is fine, but Theano versions '
+                    'prior to 0.7 (or this development version Sept 2014) '
+                    'might have given an incorrect result. '
+                    'To disable this warning, set the Theano flag '
+                    'warn.reduce_join to False.'))
+            return
+        try:
+            join_axis = get_scalar_constant_value(join.inputs[0])
+            if join_axis != node.op.axis[0]:
+                return
+        except NotScalarConstantError:
+            return
         if isinstance(node.op.scalar_op, (scalar.Maximum, scalar.Minimum)):
             #Support only 2 inputs for now
             if len(join.inputs) != 3:

theano/tensor/tests/test_opt.py

@@ -3858,7 +3858,7 @@ class T_local_reduce(unittest.TestCase):
         x = numpy.asarray([[1, 0], [3, 4]], dtype=config.floatX)
         y = numpy.asarray([[4, 0], [2, 1]], dtype=config.floatX)
         z = numpy.asarray([[5, 0], [1, 2]], dtype=config.floatX)
-
+        # Test different reduction scalar operation
         for out, res in [
             (T.max((vx, vy), 0), numpy.max((x, y), 0)),
             (T.min((vx, vy), 0), numpy.min((x, y), 0)),
@@ -3873,6 +3873,35 @@ class T_local_reduce(unittest.TestCase):
             assert len(topo) <= 2, out
             assert isinstance(topo[-1].op, T.Elemwise), out
 
+        # Test different axis for the join and the reduction
+        A = theano.shared(numpy.array([1, 2, 3, 4, 5]))
+
+        f = theano.function([], T.sum(T.stack(A, A), axis=0), mode=self.mode)
+        assert numpy.allclose(f(), [2, 4, 6, 8, 10])
+        topo = f.maker.fgraph.toposort()
+        assert isinstance(topo[-1].op, T.Elemwise)
+
+        # Test a case that was bugged in a old Theano bug
+        f = theano.function([], T.sum(T.stack(A, A), axis=1), mode=self.mode)
+        assert numpy.allclose(f(), [15, 15])
+        topo = f.maker.fgraph.toposort()
+        assert not isinstance(topo[-1].op, T.Elemwise)
+
+        # This case could be optimized
+        A = theano.shared(numpy.array([1, 2, 3, 4, 5]).reshape(5, 1))
+        f = theano.function([], T.sum(T.concatenate((A, A), axis=1), axis=1),
+                            mode=self.mode)
+        assert numpy.allclose(f(), [2, 4, 6, 8, 10])
+        topo = f.maker.fgraph.toposort()
+        assert not isinstance(topo[-1].op, T.Elemwise)
+
+        A = theano.shared(numpy.array([1, 2, 3, 4, 5]).reshape(5, 1))
+        f = theano.function([], T.sum(T.concatenate((A, A), axis=1), axis=0),
+                            mode=self.mode)
+        assert numpy.allclose(f(), [15, 15])
+        topo = f.maker.fgraph.toposort()
+        assert not isinstance(topo[-1].op, T.Elemwise)
+
 
 class T_local_sum_dimshuffle(unittest.TestCase):
     def setUp(self):