Revert some changed I didn't wanted to commit.

theano/sandbox/cuda/tests/test_abstractconv.py

@@ -191,35 +191,26 @@ class TestConv2d(unittest.TestCase):
         mode = mode_with_gpu
         # provide_shape is not used by the CuDNN impementation
         provide_shape = False
-        #((6, 1, 10, 11), (1, 1, 2, 5), (2, 4), 'full', True) fail with cudnn v3 and v4
-        #((6, 1, 10, 11), (1, 1, 2, 5), (2, 4), 'full', False) fail with cudnn v3 and v4
-        
         for (i, f), s, b, flip in itertools.product(
                 zip(self.inputs_shapes, self.filters_shapes),
                 self.subsamples,
                 self.border_modes,
                 self.filter_flip):
             o = self.get_output_shape(i, f, s, b)
-#            self.run_fwd(inputs_shape=i, filters_shape=f, subsample=s,
-#                         verify_grad=True, mode=mode, device='gpu',
-#                         provide_shape=provide_shape, border_mode=b,
-            #                         filter_flip=flip)
-            print(i, f, s, b, flip)
-            if s == (2, 2) or s == (1, 1):
-                continue
-#            if i == (6, 1, 10, 11) and f == (1, 1, 2, 5) and s == (2, 4) and b == 'full':
-#                print "skip"
-#                continue
+            self.run_fwd(inputs_shape=i, filters_shape=f, subsample=s,
+                         verify_grad=True, mode=mode, device='gpu',
+                         provide_shape=provide_shape, border_mode=b,
+                         filter_flip=flip)
             self.run_gradweight(inputs_shape=i, filters_shape=f,
                                 output_shape=o, subsample=s,
                                 verify_grad=True, mode=mode, device='gpu',
                                 provide_shape=provide_shape, border_mode=b,
                                 filter_flip=flip)
-#            self.run_gradinput(inputs_shape=i, filters_shape=f,
-#                               output_shape=o, subsample=s,
-#                               verify_grad=True, mode=mode, device='gpu',
-#                               provide_shape=provide_shape, border_mode=b,
-#                               filter_flip=flip)
+            self.run_gradinput(inputs_shape=i, filters_shape=f,
+                               output_shape=o, subsample=s,
+                               verify_grad=True, mode=mode, device='gpu',
+                               provide_shape=provide_shape, border_mode=b,
+                               filter_flip=flip)
 
     def test_gpucorrmm_conv(self):
         if not dnn_available():