Changes following review.

theano/configdefaults.py

@@ -79,13 +79,13 @@ class DeviceParam(ConfigParam):
             else:
                 raise ValueError(('Invalid value ("%s") for configuration '
                                   'variable "%s". Valid options start with '
-                                  'one of "gpu", "opencl", "cuda"'
-                                  % (val, self.fullname)))
+                                  'one of "%s", "gpu", "opencl", "cuda"'
+                                  % (self.default, val, self.fullname)))
         over = kwargs.get("allow_override", True)
         super(DeviceParam, self).__init__(default, filter, over)
 
     def __str__(self):
-        return '%s (gpu*, opencl*, cuda*) ' % (self.fullname, self.default)
+        return '%s (%s, gpu*, opencl*, cuda*) ' % (self.fullname, self.default)
 
 AddConfigVar(
     'device',
@@ -208,6 +208,7 @@ AddConfigVar('dnn.conv.algo_bwd',
                      'time_on_shape_change'),
              in_c_key=False)
 
+
 def default_dnn_path(suffix):
     def f(suffix=suffix):
         if config.cuda.root == '':

theano/sandbox/cuda/__init__.py

@@ -536,6 +536,9 @@ def handle_shared_float32(tf):
 if config.device.startswith('gpu'):
     use(device=config.device, force=config.force_device, test_driver=False)
 elif config.init_gpu_device.startswith('gpu'):
+    assert config.device == "cpu", (
+        "We can use the Theano flag init_gpu_device"
+        " only when the Theano flag device=='cpu'")
     _logger.warning(("GPU device %s will be initialized, and used if a GPU is "
           "needed. "
           "However, no computation, nor shared variables, will be implicitly "

theano/sandbox/gpuarray/conv.py

@@ -232,7 +232,7 @@ class GpuConv(GpuKernelBase, gof.Op):
 
     def c_code_cache_version(self):
         # raise this whenever modifying any of the support_code_files
-        return (0, 21)
+        return (0, 22)
 
     def c_code(self, node, nodename, inp, out_, sub):
         img, kern = inp

theano/sandbox/gpuarray/dnn.py

@@ -29,7 +29,6 @@ from .opt import gpu_seqopt, register_opt, conv_groupopt, op_lifter
 from .opt_util import alpha_merge, output_merge, inplace_allocempty
 
 
-
 def dnn_available():
     if dnn_available.avail is not None:
         return dnn_available.avail

theano/sandbox/gpuarray/elemwise.py

@@ -437,7 +437,7 @@ class GpuElemwise(GpuKernelBase, HideC, Elemwise):
     def c_code_cache_version(self):
         ver = self.scalar_op.c_code_cache_version()
         if ver:
-            return (3, ver)
+            return (4, ver)
         else:
             return ver
 
@@ -542,7 +542,7 @@ class GpuDimShuffle(HideC, DimShuffle):
         return process
 
     def c_code_cache_version(self):
-        return (4,)
+        return (5,)
 
 
 class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
@@ -791,7 +791,7 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
             Py_XDECREF(%(z)s);
             %(z)s = pygpu_empty(%(nd_out)s, new_dims,
                                 %(out_typecode)s, GA_C_ORDER,
-                                %(ctx)s, Py_None);
+                                pygpu_default_context(), Py_None);
             if (NULL == %(z)s)
             {
                 PyErr_Format(PyExc_RuntimeError, "Failed to allocate output");
@@ -1911,7 +1911,7 @@ class GpuCAReduceCuda(GpuKernelBase, HideC, CAReduceDtype):
         """ % locals(), file=sio)
 
     def c_code_cache_version_apply(self, node):
-        version = [16]  # the version corresponding to the c code in this Op
+        version = [17]  # the version corresponding to the c code in this Op
 
         # now we insert versions for the ops on which we depend...
         version.extend(self.scalar_op.c_code_cache_version())
@@ -2884,7 +2884,7 @@ class GpuCAReduceCPY(GpuKernelBase, HideC, CAReduceDtype):
         return code
 
     def c_code_cache_version(self):
-        return (1, self.GpuKernelBase_version)
+        return (2, self.GpuKernelBase_version)
 
     def generate_kernel(self, node, odtype, redux):
         if isinstance(self.scalar_op, scalar.basic.Add):

theano/sandbox/gpuarray/neighbours.py

@@ -40,7 +40,7 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
                                    dtype=ten4.type.dtype)()])
 
     def c_code_cache_version(self):
-        return (10, 1)
+        return (11,)
 
     def c_headers(self):
         return ['<numpy_compat.h>', '<gpuarray/types.h>']

theano/sandbox/gpuarray/nnet.py

@@ -293,7 +293,7 @@ class GpuCrossentropySoftmaxArgmax1HotWithBias(GpuKernelBase, Op):
         return sio.getvalue()
 
     def c_code_cache_version(self):
-        return (7,)
+        return (8,)
 
 
 gpu_crossentropy_softmax_argmax_1hot_with_bias = GpuCrossentropySoftmaxArgmax1HotWithBias()
@@ -319,7 +319,7 @@ class GpuCrossentropySoftmax1HotWithBiasDx(GpuKernelBase, Op):
         return Apply(self, [dnll, sm, y_idx], [sm.type()])
 
     def c_code_cache_version(self):
-        return (10,)
+        return (11,)
 
     def c_headers(self):
         return ['<numpy_compat.h>', '<gpuarray/types.h>']
@@ -524,7 +524,7 @@ class GpuSoftmax(GpuKernelBase, Op):
         return shape
 
     def c_code_cache_version(self):
-        return (14,) + inline_softmax.code_version
+        return (15,) + inline_softmax.code_version
 
     def c_headers(self):
         return ['<numpy_compat.h>', '<gpuarray/types.h>']
@@ -718,7 +718,7 @@ class GpuSoftmaxWithBias (GpuKernelBase, Op):
         return [shape[0]]
 
     def c_code_cache_version(self):
-        return (13,) + inline_softmax.code_version
+        return (14,) + inline_softmax.code_version
 
     def c_headers(self):
         return ['<numpy_compat.h>', '<gpuarray/types.h>']

theano/sandbox/gpuarray/subtensor.py

@@ -394,7 +394,7 @@ class GpuIncSubtensor(GpuKernelBase, IncSubtensor):
         elemwise_version = self.iadd_node.c_code_cache_version()
         if not parent_version or not elemwise_version:
             return
-        return parent_version + elemwise_version + (2,)
+        return parent_version + elemwise_version + (3,)
 
 
 class GpuAdvancedSubtensor1(HideC, tensor.AdvancedSubtensor1):

theano/sandbox/gpuarray/type.py

@@ -202,10 +202,10 @@ class GpuArrayType(Type):
     def convert_variable(self, var):
         vt = var.type
         if (type(self) == type(vt) and
-                self.typecode == var.type.typecode and
-                self.ndim == var.type.ndim and
+                self.typecode == vt.typecode and
+                self.ndim == vt.ndim and
                 all(sb == ob or ob for sb, ob in zip(self.broadcastable,
-                                                     var.type.broadcastable))):
+                                                     vt.broadcastable))):
             return theano.tensor.patternbroadcast(var, self.broadcastable)
 
     def __hash__(self):