Re-enable the grad on the switch condition as it was working before the grad change.

theano/scalar/basic.py

@@ -1051,7 +1051,12 @@ class Switch(ScalarOp):
         else:
             second_part = None
 
-        return (None, first_part, second_part)
+        # cond does affect the elements of the output so it is connected.
+        # For the sake of making the gradient convenient we assume that
+        # condition + epsilon always triggers the same branch as condition
+        condition_grad = cond.zeros_like().astype(theano.config.floatX)
+
+        return (condition_grad, first_part, second_part)
 
     def output_types(self, (cond_t, ift_t, iff_t)):
         return upcast_out(ift_t, iff_t)

theano/tensor/tests/test_basic.py

@@ -39,7 +39,7 @@ from theano.tensor import (_shared, wvector, bvector, autocast_float_as,
         tile, patternbroadcast, Eye, Shape, Default, Dot, PermuteRowElements,
         ScalarFromTensor, TensorFromScalar, dtensor4, Rebroadcast, Alloc,
         dtensor3, SpecifyShape, Mean, IncSubtensor, AdvancedIncSubtensor1,
-        itensor3, Tile, AdvancedIncSubtensor)
+        itensor3, Tile, AdvancedIncSubtensor, switch)
 from theano.tests import unittest_tools as utt
 from theano.printing import debugprint
 
@@ -618,6 +618,36 @@ SubInplaceTester = makeBroadcastTester(op=inplace.sub_inplace,
                                          grad = _grad_broadcast_binary_normal,
                                          inplace = True)
 
+
+SwitchTester = makeBroadcastTester(
+    op=switch,
+    expected=numpy.where,
+    good=dict(all_true=(numpy.asarray(1, dtype=config.floatX),
+                        rand(4, 5), rand(4, 5)),
+              false_true=(numpy.asarray(0, dtype=config.floatX),
+                          rand(4, 5), rand(4, 5)),
+              mixed=(randint_ranged(0, 1, (4, 5)),
+                     rand(4, 5), rand(4, 5))
+          ),
+    bad_build=dict(all_true=(numpy.asarray(1, dtype=config.floatX),
+                             rand(4, 5))),
+    bad_runtime=dict(all_true=(numpy.asarray(1, dtype=config.floatX),
+                               rand(3, 5), rand(4, 5)),
+                     false_true=(numpy.asarray(0, dtype=config.floatX),
+                                 rand(4, 6), rand(4, 5)),
+                 ),
+    # We suppose that cond+eps do not switch branch in switch.grad()
+    # So we can't call verify_grad with cond 0.
+    grad=dict(all_true=(numpy.asarray(1, dtype=config.floatX),
+                        rand(4, 5), rand(4, 5)),
+#              false_true=(numpy.asarray(0, dtype=config.floatX),
+#                          rand(4, 5), rand(4, 5)),
+#              mixed=(randint_ranged(0, 1, (4, 5)).astype(config.floatX),
+#                     rand(4, 5), rand(4, 5))
+          ),
+)
+
+
 MaximumTester = makeBroadcastTester(op=maximum,
                                   expected = lambda *inputs: check_floatX(inputs, numpy.maximum(*inputs)),
                                   good = _good_broadcast_binary_normal,