Some changes from review.

theano/compile/tests/test_nanguardmode.py

@@ -14,11 +14,9 @@ import theano.tensor as T
 
 
 def test_NanGuardMode():
-    """
-    Tests if NanGuardMode is working by feeding in numpy.inf and numpy.nans
-    intentionally. A working implementation should be able to capture all
-    the abnormalties.
-    """
+    # Tests if NanGuardMode is working by feeding in numpy.inf and numpy.nans
+    # intentionally. A working implementation should be able to capture all
+    # the abnormalties.
     x = T.matrix()
     w = theano.shared(numpy.random.randn(5, 7).astype(theano.config.floatX))
     y = T.dot(x, w)

theano/configparser.py

@@ -90,10 +90,11 @@ theano_cfg.read(config_files)
 theano_raw_cfg = ConfigParser.RawConfigParser()
 theano_raw_cfg.read(config_files)
 
+
 class change_flags(object):
     """
     Use this as a decorator or context manager to change the value of
-    Theano config variable.
+    Theano config variables.
 
     Useful during tests.
     """
@@ -112,6 +113,7 @@ class change_flags(object):
         def res(*args, **kwargs):
             with self:
                 return f(*args, **kwargs)
+        return res
 
     def __enter__(self):
         self.old_vals = {}

theano/gof/type.py

@@ -670,6 +670,12 @@ class CDataType(Type):
         return data
 
     def _get_func(self):
+        """
+        Return a function that makes a value from an integer.
+
+        The integer value is assumed to be a valid pointer for the
+        type and no check is done to ensure that.
+        """
         from theano.scalar import get_scalar_type
 
         if self._fn is None:
@@ -680,6 +686,15 @@ class CDataType(Type):
         return self._fn
 
     def make_value(self, ptr):
+        """
+        Make a value of this type.
+
+        Parameters
+        ----------
+        ptr : int
+            Integer representation of a valid pointer value
+
+        """
         return self._get_func()(ptr)
 
     def c_declare(self, name, sub, check_input=True):

theano/gpuarray/dnn.py

@@ -58,11 +58,12 @@ def _dnn_lib():
         lib_name = ctypes.util.find_library('cudnn')
         if lib_name is None and sys.platform == 'win32':
             # Update these names when new versions of cudnn are supported.
-            lib_name = ctypes.util.find_library('cudnn64_5.dll')
-            if lib_name is None:
-                lib_name = ctypes.util.find_library('cudnn64_4.dll')
+            for name in ['cudnn64_5.dll', 'cudnn64_4.dll']:
+                lib_name = ctypes.util.find_library(name)
+                if lib_name:
+                    break
         if lib_name is None:
-            raise RuntimeError('Could not find cudnn library')
+            raise RuntimeError('Could not find cudnn library (maybe your are using a )')
         _dnn_lib.handle = ctypes.cdll.LoadLibrary(lib_name)
         cudnn = _dnn_lib.handle
         cudnn.cudnnCreate.argtypes = [ctypes.POINTER(ctypes.c_void_p)]
@@ -1978,8 +1979,10 @@ class _RNNSplitParams(DnnBase):
 
     def make_node(self, w, desc, layer, isize, typecode):
         w = as_gpuarray_variable(w, infer_context_name(w))
+        assert w.ndim == 1
         layer = as_scalar(layer).astype('int32')
         isize = as_tensor_variable(isize).astype('uint64')
+        assert isize.ndim == 2
         typecode = as_scalar(typecode).astype('int32')
         _1d = GpuArrayType(w.type.dtype, [False],
                            context_name=w.type.context_name)
@@ -2087,6 +2090,7 @@ class _RNNSplitParams(DnnBase):
   nshp[0] = PyGpuArray_DIM(%(w)s, 0);
   nshp[1] = 1;
         """ % kw
+
         def get_params(id, m, b):
             kw2 = kw.copy()
             kw2['id'] = id
@@ -2152,7 +2156,7 @@ class _RNNSplitParams(DnnBase):
   %(m)s->ga.strides[1] = %(m)s->ga.dimensions[0] * gpuarray_get_elsize(%(m)s->ga.typecode);
             """ % kw2
 
-        for i in range(len(outputs)//2):
+        for i in range(len(outputs) // 2):
             code += get_params(i, outputs[2 * i], outputs[(2 * i) + 1])
 
         code += """
@@ -2189,16 +2193,16 @@ class GpuDnnRNNOp(DnnBase):
         elif direction_mode == 'unidirectional':
             self.num_dirs = 1
         else:
-            raise ValueError('direction_mode')
+            raise ValueError('direction_mode is invalid (got %s)' % (direction_mode,))
 
     def dnn_context(self, node):
         return node.outputs[1].type.context_name
 
     def make_node(self, desc, w, x, hx, cx=None):
         if cx is None:
-            context_name = infer_context_name(w, x, hx, cx)
-        else:
             context_name = infer_context_name(w, x, hx)
+        else:
+            context_name = infer_context_name(w, x, hx, cx)
 
         w = as_gpuarray_variable(w, context_name)
         x = as_gpuarray_variable(x, context_name)
@@ -2232,8 +2236,15 @@ class GpuDnnRNNOp(DnnBase):
         reserve, y, hy = outputs[:3]
         _, dy, dhy = output_grads[:3]
         dcy = output_grads[3] if len(output_grads) == 4 else None
-        # If both dy and dhy are disconnected, then this will error
-        # out, but it is indeed an error.
+        # Since the op return two outputs which contain essentially
+        # the same information, the user will most likely only use one
+        # of them.  This leads to the situation that the other is
+        # considered "disconnected" by theano in the gradient.
+        # However we know that this isn't really the case so we fix it
+        # up here.
+
+        # If both dy and dhy are disconnected the fixup will fail, but
+        # that's ok as in that case we really are disconnected.
         if isinstance(dy.type, DisconnectedType):
             dy = as_gpuarray_variable(dhy[-1],
                                       context_name=dhy.type.context_name)

theano/gpuarray/tests/rnn_support.py

-from __future__ import absolute_import, print_function
+from __future__ import absolute_import, print_function, division
 
 import theano
 import theano.tensor as T

theano/gpuarray/tests/test_dnn.py

@@ -1448,11 +1448,8 @@ def test_dnn_rnn_gru():
 
     # test code
     X = T.tensor3('X')
-    X.tag.test_value = numpy.zeros((timesteps, batch_size, input_dim), dtype=theano.config.floatX)
     Y = T.tensor3('Y')
-    Y.tag.test_value = numpy.zeros((timesteps, batch_size, hidden_dim), dtype=theano.config.floatX)
     h0 = T.tensor3('h0')
-    h0.tag.test_value = numpy.zeros((depth, batch_size, hidden_dim), dtype=theano.config.floatX)
 
     rnnb = dnn.RNNBlock(theano.config.floatX, hidden_dim, depth, 'gru')
     psize = rnnb.get_param_size([batch_size, input_dim])

theano/gpuarray/type.py

@@ -20,7 +20,6 @@ except ImportError:
     pygpu = None
 
 _context_reg = {}
-_props_map = {}
 
 
 def move_to_gpu(data):
@@ -91,6 +90,20 @@ def get_context(name):
     return _context_reg[name]
 
 
+def list_contexts():
+    """
+    Return an iterable of all the registered context names.
+
+    """
+    return _context_reg.keys()
+
+# Mappings of properties to contexts.  Please never use this if you
+# can avoid it.
+
+# This is basically a way to store "global" variables that depend on
+# the context.
+_props_map = {}
+
 def _get_props(name):
     ctx = get_context(name)
     return _props_map[ctx]
@@ -104,14 +117,6 @@ def set_prop(name, k, v):
     _get_props(name)[k] = v
 
 
-def list_contexts():
-    """
-    Return an iterable of all the registered context names.
-
-    """
-    return _context_reg.keys()
-
-
 # Private method
 def _name_for_ctx(ctx):
     for k, v in iteritems(_context_reg):