Fixed issue with pow grad function trying to evalutate log at 0. Issue 1780

theano/scalar/basic.py

@@ -1701,6 +1701,7 @@ class Pow(BinaryScalarOp):
         first_part = gz * y * x ** (y - 1)
 
         second_part = gz * log(x) * x ** y
+        second_part = switch(eq(x, 0), 0, second_part)
 
         return (first_part, second_part)
 

theano/tensor/tests/test_basic.py

@@ -926,6 +926,10 @@ _grad_broadcast_pow_normal = dict(same_shapes = (rand_ranged(1, 5, (2, 3)), rand
 #empty2 case is not supported by numpy.
 _good_broadcast_pow_normal_float_pow = copy(_good_broadcast_pow_normal_float)
 del _good_broadcast_pow_normal_float_pow["empty2"]
+_grad_broadcast_pow_normal["x_eq_zero"] = (
+    numpy.asarray([0.], dtype=config.floatX),
+    numpy.asarray([2.], dtype=config.floatX)
+)
 
 PowTester = makeBroadcastTester(
         op=pow,