Merge pull request #791 from nouiz/fix_test3

theano/tensor/opt.py

@@ -426,8 +426,11 @@ def dimshuffle_as_view(node):
     new_op = DimShuffle(op.input_broadcastable, op.new_order, inplace=True)
     return [new_op(*node.inputs)]
 
-
-register_specialize(dimshuffle_as_view, 'inplace')
+#Step 60 is the inplace optimization stage.
+compile.optdb.register('dimshuffle_as_view',
+                       TopoOptimizer(dimshuffle_as_view,
+    failure_callback=TopoOptimizer.warn_inplace), 60,
+                       'fast_run', 'inplace')
 register_canonicalize(local_dimshuffle_lift)
 register_specialize(local_dimshuffle_lift)
 
@@ -3142,9 +3145,14 @@ def local_cut_useless_reduce(node):
             return [summed]
 
 
-@register_canonicalize
+#Enabling this optimization at canonicalization step break this test:
+#theano/tensor/tests/test_opt.py:T_local_reduce.test_local_reduce_broadcast_some_0
+# see gh-790 issue.
+#
+#@register_canonicalize
+@register_specialize
 @gof.local_optimizer([])
-def local_sum_broadcastable(node):
+def local_reduce_broadcastable(node):
     """Remove reduction over broadcastable dimensions"""
     if isinstance(node.op, T.CAReduce):
         reduced, = node.inputs
@@ -3169,12 +3177,18 @@ def local_sum_broadcastable(node):
                         ii += 1
                 new_reduced = reduced.dimshuffle(*pattern)
                 if new_axis:
-                    new_op = node.op.__class__(axis=new_axis)
+                    if type(node.op) == theano.tensor.elemwise.CAReduce:
+                        # This happen for tensor.max(), tensor.min()
+                        new_op = node.op.__class__(node.op.scalar_op,
+                                                   axis=new_axis)
+                    else:
+                        new_op = node.op.__class__(axis=new_axis)
                     return [new_op(new_reduced)]
                 else:
                     # -- in this case we can remove the reduction completely
                     return [new_reduced.astype(odtype)]
 
+
 @register_specialize
 @gof.local_optimizer([])
 def local_sum_alloc(node):

theano/tensor/tests/test_opt.py

@@ -3188,7 +3188,8 @@ class test_local_remove_switch_const_cond(unittest.TestCase):
 
 class T_local_sum(unittest.TestCase):
     def setUp(self):
-        self.mode = theano.compile.get_default_mode().including('canonicalize')
+        self.mode = theano.compile.get_default_mode().including('canonicalize',
+                                                                'specialize')
 
     def test_local_sum_all_to_none(self):
         a = T.tensor3()
@@ -3312,54 +3313,59 @@ class T_local_sum(unittest.TestCase):
         finally:
             config.on_opt_error = backup
 
-    def test_local_sum_broadcast_all_0(self):
-        optimizer = optdb.query(self.mode._optimizer)
-
-        x = T.TensorType('int64', (True, True, True))()
-        g = FunctionGraph([x], [x.sum()])
-        optimizer.optimize(g)
-        assert not any([
-            isinstance(node.op, T.CAReduce)
-            for node in g.toposort()])
-
-    def test_local_sum_broadcast_all_1(self):
-        optimizer = optdb.query(self.mode._optimizer)
-
-        x = T.TensorType('int64', (True, True))()
-        g = FunctionGraph([x], [x.sum(axis=[0, 1])])
-        optimizer.optimize(g)
-        assert not any([
-            isinstance(node.op, T.CAReduce)
-            for node in g.toposort()])
-
-    def test_local_sum_broadcast_some_0(self):
-        optimizer = optdb.query(self.mode._optimizer)
-
-        x = T.TensorType('int64', (True, False, True))()
-        g = FunctionGraph([x], [x.sum(axis=[0, 1])])
-        optimizer.optimize(g)
-        order = g.toposort()
-        assert 1 == sum([isinstance(node.op, T.CAReduce) for node in order])
-        if config.mode == 'FAST_COMPILE':
-            node = order[-1]
-        else:
-            node = order[-2]
-        op = node.op
-        assert isinstance(op, T.CAReduce)
-        # -- the leading broadcastable dimension has been dropped
-        #   by the local_sum_broadcastable optimization
-        #   now summation is over the original x's dimension 1.
-        assert node.inputs[0].ndim == 2, node
-        assert op.axis == (0,), op.axis
-
-    def test_local_sum_broadcast_some_1(self):
-        optimizer = optdb.query(self.mode._optimizer)
-
-        x = T.TensorType('int64', (True, False, True))()
-        g = FunctionGraph([x], [x.sum(axis=[0, 2])])
-        optimizer.optimize(g)
-        order = g.toposort()
-        assert 0 == sum([isinstance(node.op, T.CAReduce) for node in order])
+
+class T_local_reduce(unittest.TestCase):
+    def setUp(self):
+        self.mode = theano.compile.get_default_mode().including('canonicalize',
+                                                                'specialize')
+
+    def test_local_reduce_broadcast_all_0(self):
+        for fct in [tensor.sum, tensor.all, tensor.any, tensor.prod,
+                    tensor.max, tensor.min]:
+            x = T.TensorType('int64', (True, True, True))()
+            f = theano.function([x], [fct(x)], mode=self.mode)
+            assert not any([
+                isinstance(node.op, T.CAReduce)
+                for node in f.maker.fgraph.toposort()])
+
+    def test_local_reduce_broadcast_all_1(self):
+        for fct in [tensor.sum, tensor.all, tensor.any, tensor.prod,
+                    tensor.max, tensor.min]:
+            x = T.TensorType('int64', (True, True))()
+            f = theano.function([x], [fct(x, axis=[0, 1])], mode=self.mode)
+            assert not any([
+                isinstance(node.op, T.CAReduce)
+                for node in f.maker.fgraph.toposort()])
+
+    def test_local_reduce_broadcast_some_0(self):
+        for fct in [tensor.sum, tensor.all, tensor.any, tensor.prod,
+                    tensor.max, tensor.min]:
+            x = T.TensorType('int64', (True, False, True))()
+            f = theano.function([x], [fct(x, axis=[0, 1])], mode=self.mode)
+
+            order = f.maker.fgraph.toposort()
+            assert 1 == sum([isinstance(node.op, T.CAReduce)
+                             for node in order])
+
+            node = [node for node in order if isinstance(node.op,
+                                                         tensor.CAReduce)][0]
+
+            op = node.op
+            assert isinstance(op, T.CAReduce)
+            # -- the leading broadcastable dimension has been dropped
+            #   by the local_reduce_broadcastable optimization
+            #   now summation is over the original x's dimension 1.
+            assert node.inputs[0].ndim == 2, node
+            assert op.axis == (0,), op.axis
+
+    def test_local_reduce_broadcast_some_1(self):
+        for fct in [tensor.sum, tensor.all, tensor.any, tensor.prod,
+                    tensor.max, tensor.min]:
+            x = T.TensorType('int64', (True, True, True))()
+            f = theano.function([x], [fct(x, axis=[0, 2])], mode=self.mode)
+            assert not any([
+                isinstance(node.op, T.CAReduce)
+                for node in f.maker.fgraph.toposort()])
 
 
 class T_local_sum_dimshuffle(unittest.TestCase):