Merge pull request #4874 from gvtulder/f-abstractconv-dilationtest

theano/tensor/nnet/abstract_conv.py

@@ -966,14 +966,15 @@ class AbstractConv2d_gradWeights(BaseAbstractConv2d):
                 '"valid", "full", "half", an integer or a pair of'
                 ' integers'.format(mode))
 
+        dil_shape = ((shape[0] - 1) * self.filter_dilation[0] + 1,
+                     (shape[1] - 1) * self.filter_dilation[1] + 1)
+
         if mode == "full":
-            mode = (shape[0] - 1, shape[1] - 1)
+            mode = (dil_shape[0] - 1, dil_shape[1] - 1)
         elif mode == "half":
-            mode = (shape[0] // 2, shape[1] // 2)
+            mode = (dil_shape[0] // 2, dil_shape[1] // 2)
         if isinstance(mode, tuple):
             pad_h, pad_w = map(int, mode)
-            pad_h = (pad_h - 1) * self.filter_dilation[0] + 1
-            pad_w = (pad_w - 1) * self.filter_dilation[1] + 1
 
             mode = "valid"
             new_img = numpy.zeros((img.shape[0], img.shape[1],
@@ -983,8 +984,6 @@ class AbstractConv2d_gradWeights(BaseAbstractConv2d):
             img = new_img
 
         if self.subsample[0] > 1 or self.subsample[1] > 1:
-            dil_shape = ((shape[0] - 1) * self.filter_dilation[0] + 1,
-                         (shape[1] - 1) * self.filter_dilation[1] + 1)
             new_shape = (topgrad.shape[0], topgrad.shape[1],
                          img.shape[2] - dil_shape[0] + 1,
                          img.shape[3] - dil_shape[1] + 1)
@@ -1114,9 +1113,9 @@ class AbstractConv2d_gradInputs(BaseAbstractConv2d):
                        (kern.shape[3] - 1) * self.filter_dilation[1] + 1)
         pad_h, pad_w = 0, 0
         if mode == "full":
-            pad_h, pad_w = (dil_kernshp[0] - 1, dil_kernshp[0] - 1)
+            pad_h, pad_w = (dil_kernshp[0] - 1, dil_kernshp[1] - 1)
         elif mode == "half":
-            pad_h, pad_w = (dil_kernshp[1] // 2, dil_kernshp[1] // 2)
+            pad_h, pad_w = (dil_kernshp[0] // 2, dil_kernshp[1] // 2)
         elif isinstance(mode, tuple):
             pad_h, pad_w = map(int, self.border_mode)
         if self.subsample[0] > 1 or self.subsample[1] > 1:

theano/tensor/nnet/tests/test_abstract_conv.py

@@ -89,7 +89,7 @@ class BaseTestConv2d:
                               (1, 1, 2, 3), (4, 1, 1, 3), (4, 5, 3, 2)]
         cls.subsamples = [(1, 1), (2, 2), (2, 4)]
         cls.filters_dilations = [(1, 1), (1, 2), (2, 1)]
-        cls.border_modes = ["valid", "full", (0, 0), (1, 1), (5, 5), (5, 2)]
+        cls.border_modes = ["valid", "half", "full", (0, 0), (1, 1), (5, 5), (5, 2)]
         cls.filter_flip = [True, False]
         cls.provide_shape = [True, False]
         cls.shared = staticmethod(theano.compile.shared)
@@ -100,6 +100,9 @@ class BaseTestConv2d:
                        (filters_shape[3] - 1) * filter_dilation[1] + 1)
         if border_mode == "valid":
             border_mode = (0, 0)
+        if border_mode == "half":
+            border_mode = (dil_filters[0] // 2,
+                           dil_filters[1] // 2)
         if border_mode == "full":
             border_mode = (dil_filters[0] - 1,
                            dil_filters[1] - 1)
@@ -292,7 +295,7 @@ class BaseTestConv2d:
             for fd in self.filters_dilations:
                 for s in self.subsamples:
                     for b in self.border_modes:
-                        yield (self.tcase, i, f, s, db, dflip,
+                        yield (self.tcase, i, f, s, b, dflip,
                                dprovide_shape, fd)
             for flip in self.filter_flip:
                 yield (self.tcase, i, f, ds, db, flip, dprovide_shape)