Use gradient_undefined instead of None in MRG uniform

theano/sandbox/rng_mrg.py

@@ -10,6 +10,7 @@ import warnings
 import numpy
 
 from theano import Op, Apply, shared, config, Variable
+from theano import gradient
 from theano import tensor
 from theano.tensor import (raw_random, TensorType, as_tensor_variable,
                            get_vector_length, cast, opt, scal)
@@ -175,7 +176,10 @@ class mrg_uniform_base(Op):
                      [rstate.type(), self.output_type()])
 
     def grad(self, inputs, ograd):
-        return [None for i in inputs]
+        return [gradient.grad_undefined(
+                    self, k, inp,
+                    'No gradient defined through random sampling op')
+                for k, inp in enumerate(inputs)]
 
     def R_op(self, inputs, eval_points):
         return [None for i in eval_points]

theano/sandbox/test_rng_mrg.py

@@ -752,3 +752,17 @@ def test_random_state_transfer():
         su2[0].set_value(su1[0].get_value())
 
     numpy.testing.assert_array_almost_equal(f1(), f2(), decimal=6)
+
+
+def test_gradient_scan():
+    # Test for a crash when using MRG inside scan and taking the gradient
+    # See https://groups.google.com/d/msg/theano-dev/UbcYyU5m-M8/UO9UgXqnQP0J
+    theano_rng = MRG_RandomStreams(10)
+    w = theano.shared(numpy.ones(1, dtype='float32'))
+
+    def one_step(x):
+        return x + theano_rng.uniform((1,), dtype='float32') * w
+
+    x = tensor.vector(dtype='float32')
+    values, updates = theano.scan(one_step, outputs_info=x, n_steps=10)
+    theano.grad(tensor.sum(values[-1]), w)