initial commit for gradient for MRG_RandomStreams distribution variables should…

theano/gradient.py

@@ -1994,6 +1994,38 @@ def zero_grad(x):
     return zero_grad_(x)
 
 
+class UndefinedGrad(ViewOp):
+    def grad(self, args, g_outs):
+        return [grad_undefined(self, i, inp[i]) for i in xrange(args)]
+
+    def R_op(self, inputs, eval_points):
+        return [None]
+
+    def connection_pattern(self, node):
+        return [[False]]
+
+
+undefined_grad_ = UndefinedGrad()
+
+
+def undefined_grad(x):
+    """
+    Consider the gradient of this variable undefined and
+    generate an error message if its gradient is taken.
+
+    The expression itself is unaffected, but when its gradient is
+    computed, or the gradient of another expression that this
+    expression is a subexpression of, an error message will be generated
+    specifying such gradient is not defined.
+
+    :param x: A Theano expression whose gradient should be undefined.
+
+    :return: The expression is returned unmodified, but its gradient
+        is now undefined.
+    """
+    return undefined_grad_(x)
+
+
 class DisconnectedGrad(ViewOp):
     def grad(self, args, g_outs):
         return [disconnected_type() for g_out in g_outs]

theano/sandbox/rng_mrg.py

@@ -15,6 +15,7 @@ from six.moves import xrange
 import theano
 from theano import Op, Apply, shared, config, Variable
 from theano import gradient, function
+from theano.gradient import undefined_grad
 from theano import tensor
 from theano.tensor import (TensorType, as_tensor_variable, get_vector_length,
                            cast, opt, scal)
@@ -773,7 +774,9 @@ class MRG_RandomStreams(object):
 
         """
         low = as_tensor_variable(low)
+        low = undefined_grad(low)
         high = as_tensor_variable(high)
+        high = undefined_grad(high)
         if dtype is None:
             dtype = scal.upcast(config.floatX, low.dtype, high.dtype)
 
@@ -821,6 +824,7 @@ class MRG_RandomStreams(object):
                  nstreams=None):
         # TODO : need description for method, parameter and return
         if n == 1:
+            p = undefined_grad(p)
             x = self.uniform(size=size, nstreams=nstreams)
             return cast(x < p, dtype)
         else:
@@ -852,6 +856,7 @@ class MRG_RandomStreams(object):
         if pvals is None:
             raise TypeError("You have to specify pvals")
         pvals = as_tensor_variable(pvals)
+        pvals = undefined_grad(pvals)
         if size is not None:
             if any([isinstance(i, integer_types) and i <= 0 for i in size]):
                 raise ValueError(
@@ -978,7 +983,9 @@ class MRG_RandomStreams(object):
         # second half our U2's. See Wikipedia page:
         # http://en.wikipedia.org/wiki/Box%E2%80%93Muller_transform
         avg = as_tensor_variable(avg)
+        avg = undefined_grad(avg)
         std = as_tensor_variable(std)
+        std = undefined_grad(std)
 
         if dtype is None:
             dtype = scal.upcast(config.floatX, avg.dtype, std.dtype)