Generalized filter dialation parameter and reverted a change

theano/gpuarray/basic_ops.py

@@ -973,22 +973,7 @@ def empty_like(var):
     return GpuAllocEmpty(var.type.dtype, var.type.context_name)(*var.shape)
 
 
-def gpu_alloc_empty(ctx, **kwargs):
-    '''
-    This is the cache method of GpuAllocEmpty class.
-    This takes the parameters of context name and props_dict
-    and retrieves the dtype key from the dictionary
-
-    Parameters
-    ----------
-    ctx : String
-        The context name.
-
-    kwargs : Dict
-        The props_dict of the Op
-
-    '''
-    dtype = kwargs.get('dtype')
+def gpu_alloc_empty(ctx, dtype):
     key = (dtype, ctx)
     if key not in gpu_alloc_empty.cache:
         gpu_alloc_empty.cache[key] = GpuAllocEmpty(dtype, ctx)

theano/gpuarray/dnn.py

@@ -956,8 +956,8 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
         shape2 = shape_i(img, 2, fgraph) + shape_i(kerns, 2, fgraph) - 1
         shape3 = shape_i(img, 3, fgraph) + shape_i(kerns, 3, fgraph) - 1
         out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(shape_i(img, 0, fgraph),
-                                                   shape_i(kerns, 1, fgraph),
-                                                   shape2, shape3)
+                                                         shape_i(kerns, 1, fgraph),
+                                                         shape2, shape3)
         desc = GpuDnnConvDesc(border_mode='valid', subsample=(1, 1),
                               conv_mode=conv_mode, precision=precision)(kerns.shape)
         return gpu_dnn_conv_gradI()(kerns, img, out, desc)

theano/gpuarray/extra_ops.py

@@ -2,7 +2,6 @@ from __future__ import absolute_import, print_function, division
 import os
 from theano import Apply, Op
 from theano.tensor.extra_ops import CumsumOp
-from .type import GpuArrayType
 from .basic_ops import infer_context_name
 try:
     from pygpu import gpuarray
@@ -42,6 +41,7 @@ class GpuCumsum(GpuKernelBase, Op):
         assert x.type.dtype == 'float32', "Only float32 supported for GpuCumSum"
 
         context_name = infer_context_name(x)
+
         x = as_gpuarray_variable(x, context_name)
 
         if x.ndim > GpuCumsum.SUPPORTED_NDIMS:

theano/gpuarray/opt.py

@@ -591,8 +591,7 @@ def local_gpuaalloc(op, context_name, inputs, outputs):
 def local_gpuaallocempty(op, context_name, inputs, outputs):
     # We use _props_dict() to make sure that the GPU op know all the
     # CPU op props.
-    dtype = op._props_dict().get('dtype')
-    return gpu_alloc_empty(context_name, dtype=dtype)(*inputs)
+    return gpu_alloc_empty(context_name, **op._props_dict())(*inputs)
 
 
 @register_opt()
@@ -960,7 +959,7 @@ def local_gpua_incsubtensor(op, context_name, inputs, outputs):
                          op.set_instead_of_inc,
                          op.destroyhandler_tolerate_aliased)
     ret = op(*inputs)
-    val = getattr(op.make_node(*inputs).outputs[0].tag, 'nan_guard_mode_check', True)
+    val = getattr(outputs[0].tag, 'nan_guard_mode_check', True)
     ret.tag.nan_guard_mode_check = val
     return ret
 

theano/gpuarray/tests/test_abstractconv.py

@@ -26,9 +26,11 @@ class TestDnnConv2d(test_abstract_conv.BaseTestConv2d):
         if not dnn_available(test_ctx_name):
             raise SkipTest(dnn_available.msg)
         mode = mode_with_gpu
+
         if fd != (1, 1):
             raise SkipTest("Doesn't have CUDNN implementation")
         o = self.get_output_shape(i, f, s, b, fd)
+
         self.run_fwd(inputs_shape=i, filters_shape=f, subsample=s,
                      verify_grad=True, mode=mode,
                      provide_shape=provide_shape, border_mode=b,