test optimization in grouped convolution

theano/gpuarray/tests/test_dnn.py

@@ -2288,27 +2288,17 @@ def dconvgi(border_mode, subsample, filter_dilation, num_groups):
 
 
 class Cudnn_grouped_conv(Grouped_conv_noOptim):
-    mode = mode_with_gpu
-    conv = staticmethod(dconvfwd)
-    conv_gradw = staticmethod(dconvgw)
-    conv_gradi = staticmethod(dconvgi)
+    mode = mode_with_gpu.excluding('conv_gemm')
     conv_op = dnn.GpuDnnConv
     conv_gradw_op = dnn.GpuDnnConvGradW
     conv_gradi_op = dnn.GpuDnnConvGradI
-    flip_filter = False
-    is_dnn = True
 
 
 class Cudnn_grouped_conv3d(Grouped_conv3d_noOptim):
-    mode = mode_with_gpu
-    conv = staticmethod(dconvfwd)
-    conv_gradw = staticmethod(dconvgw)
-    conv_gradi = staticmethod(dconvgi)
+    mode = mode_with_gpu.excluding('conv_gemm')
     conv_op = dnn.GpuDnnConv
     conv_gradw_op = dnn.GpuDnnConvGradW
     conv_gradi_op = dnn.GpuDnnConvGradI
-    flip_filter = False
-    is_dnn = True
 
 
 def test_dnn_spatialtf():

theano/gpuarray/tests/test_gemmcorr.py

@@ -223,12 +223,7 @@ class TestCorrMM(unittest.TestCase):
 
 
 class TestGroupGpuCorr2d(Grouped_conv_noOptim):
-    mode = theano.compile.get_mode("FAST_RUN")
-    conv = GpuCorrMM
-    conv_gradw = GpuCorrMM_gradWeights
-    conv_gradi = GpuCorrMM_gradInputs
+    mode = mode_with_gpu.excluding('cudnn')
     conv_op = GpuCorrMM
     conv_gradw_op = GpuCorrMM_gradWeights
     conv_gradi_op = GpuCorrMM_gradInputs
-    flip_filter = True
-    is_dnn = False

theano/gpuarray/tests/test_gemmcorr3d.py

@@ -222,12 +222,7 @@ class TestCorr3dMM(unittest.TestCase):
 
 
 class TestGroupGpuCorr3d(Grouped_conv3d_noOptim):
-    mode = theano.compile.get_mode("FAST_RUN")
-    conv = GpuCorr3dMM
-    conv_gradw = GpuCorr3dMM_gradWeights
-    conv_gradi = GpuCorr3dMM_gradInputs
+    mode = mode_with_gpu.excluding('cudnn')
     conv_op = GpuCorr3dMM
     conv_gradw_op = GpuCorr3dMM_gradWeights
     conv_gradi_op = GpuCorr3dMM_gradInputs
-    flip_filter = True
-    is_dnn = False

theano/tensor/nnet/tests/test_abstract_conv.py

@@ -1439,7 +1439,6 @@ class Grouped_conv_noOptim(unittest.TestCase):
     conv_gradw_op = theano.tensor.nnet.abstract_conv.AbstractConv2d_gradWeights
     conv_gradi_op = theano.tensor.nnet.abstract_conv.AbstractConv2d_gradInputs
     mode = theano.Mode(optimizer=None)
-    flip_filter = False
     is_dnn = False
 
     def setUp(self):
@@ -1476,13 +1475,7 @@ class Grouped_conv_noOptim(unittest.TestCase):
                                         subsample=self.subsample,
                                         filter_dilation=self.filter_dilation,
                                         num_groups=groups)
-            if self.flip_filter:
-                if self.convdim == 2:
-                    grouped_conv_output = grouped_conv_op(img_sym, kern_sym[:, :, ::-1, ::-1])
-                else:
-                    grouped_conv_output = grouped_conv_op(img_sym, kern_sym[:, :, ::-1, ::-1, ::-1])
-            else:
-                grouped_conv_output = grouped_conv_op(img_sym, kern_sym)
+            grouped_conv_output = grouped_conv_op(img_sym, kern_sym)
 
             grouped_func = theano.function([img_sym, kern_sym], grouped_conv_output, mode=self.mode)
             assert any([isinstance(node.op, self.conv_op)
@@ -1527,13 +1520,7 @@ class Grouped_conv_noOptim(unittest.TestCase):
             grouped_conv_output = grouped_convgrad_op(img_sym,
                                                       top_sym,
                                                       tensor.as_tensor_variable(
-                                                          kshp if self.is_dnn
-                                                          else kshp[-self.convdim:]))
-            if self.flip_filter:
-                if self.convdim == 2:
-                    grouped_conv_output = grouped_conv_output[:, :, ::-1, ::-1]
-                else:
-                    grouped_conv_output = grouped_conv_output[:, :, ::-1, ::-1, ::-1]
+                                                             kshp[-self.convdim:]))
             grouped_func = theano.function([img_sym, top_sym], grouped_conv_output, mode=self.mode)
             assert any([isinstance(node.op, self.conv_gradw_op)
                        for node in grouped_func.maker.fgraph.toposort()])
@@ -1556,8 +1543,7 @@ class Grouped_conv_noOptim(unittest.TestCase):
             def conv_gradweight(inputs_val, output_val):
                 return grouped_convgrad_op(inputs_val, output_val,
                                            tensor.as_tensor_variable(
-                                               kshp if self.is_dnn
-                                               else kshp[-self.convdim:]))
+                                                    kshp[-self.convdim:]))
 
             utt.verify_grad(conv_gradweight,
                             [img, top],
@@ -1580,19 +1566,10 @@ class Grouped_conv_noOptim(unittest.TestCase):
                                                   subsample=self.subsample,
                                                   filter_dilation=self.filter_dilation,
                                                   num_groups=groups)
-            if self.flip_filter:
-                if self.convdim == 2:
-                    grouped_conv_output = grouped_convgrad_op(kern_sym[:, :, ::-1, ::-1], top_sym,
-                                                              tensor.as_tensor_variable(imshp[-self.convdim:]))
-                else:
-                    grouped_conv_output = grouped_convgrad_op(kern_sym[:, :, ::-1, ::-1, ::-1], top_sym,
-                                                              tensor.as_tensor_variable(imshp[-self.convdim:]))
-            else:
-                grouped_conv_output = grouped_convgrad_op(kern_sym,
-                                                          top_sym,
-                                                          tensor.as_tensor_variable(
-                                                              imshp if self.is_dnn
-                                                              else imshp[-self.convdim:]))
+            grouped_conv_output = grouped_convgrad_op(kern_sym,
+                                                      top_sym,
+                                                      tensor.as_tensor_variable(
+                                                               imshp[-self.convdim:]))
             grouped_func = theano.function([kern_sym, top_sym], grouped_conv_output, mode=self.mode)
             assert any([isinstance(node.op, self.conv_gradi_op)
                        for node in grouped_func.maker.fgraph.toposort()])
@@ -1615,8 +1592,7 @@ class Grouped_conv_noOptim(unittest.TestCase):
             def conv_gradinputs(filters_val, output_val):
                 return grouped_convgrad_op(filters_val, output_val,
                                            tensor.as_tensor_variable(
-                                               imshp if self.is_dnn
-                                               else imshp[-self.convdim:]))
+                                                    imshp[-self.convdim:]))
 
             utt.verify_grad(conv_gradinputs,
                             [kern, top],
@@ -1631,8 +1607,6 @@ class Grouped_conv3d_noOptim(Grouped_conv_noOptim):
     conv_gradw_op = theano.tensor.nnet.abstract_conv.AbstractConv3d_gradWeights
     conv_gradi_op = theano.tensor.nnet.abstract_conv.AbstractConv3d_gradInputs
     mode = theano.Mode(optimizer=None)
-    flip_filter = False
-    is_dnn = False
 
     def setUp(self):
         self.num_groups = [3, 2, 4, 4]

theano/tensor/nnet/tests/test_corr.py

@@ -418,18 +418,10 @@ class TestCorr2D(utt.InferShapeTester):
 
 
 class TestGroupCorr2d(Grouped_conv_noOptim):
-    if theano.config.mode == "FAST_COMPILE":
-        mode = theano.compile.get_mode("FAST_RUN")
-    else:
-        mode = None
-    conv = corr.CorrMM
-    conv_gradw = corr.CorrMM_gradWeights
-    conv_gradi = corr.CorrMM_gradInputs
+    mode = theano.compile.get_mode("FAST_RUN").excluding('gpuarray')
     conv_op = corr.CorrMM
     conv_gradw_op = corr.CorrMM_gradWeights
     conv_gradi_op = corr.CorrMM_gradInputs
-    flip_filter = True
-    is_dnn = False
 
     def test_graph(self):
         # define common values  first

theano/tensor/nnet/tests/test_corr3d.py

@@ -420,13 +420,7 @@ class TestCorr3D(utt.InferShapeTester):
 
 
 class TestGroupCorr3d(Grouped_conv3d_noOptim):
-    if theano.config.mode == "FAST_COMPILE":
-        mode = theano.compile.get_mode("FAST_RUN")
-    else:
-        mode = None
-    conv = corr3d.Corr3dMM
-    conv_gradw = corr3d.Corr3dMM_gradWeights
-    conv_gradi = corr3d.Corr3dMM_gradInputs
+    mode = theano.compile.get_mode("FAST_RUN")
     conv_op = corr3d.Corr3dMM
     conv_gradw_op = corr3d.Corr3dMM_gradWeights
     conv_gradi_op = corr3d.Corr3dMM_gradInputs