Add attrs `deformable_groups` for deformable_conv API (#31335)

* add attrs deformable_groups

python/paddle/fluid/tests/unittests/test_deform_conv2d.py

@@ -32,6 +32,7 @@ class TestDeformConv2D(TestCase):
         self.padding = [0, 0]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = True
 
@@ -67,11 +68,11 @@ class TestDeformConv2D(TestCase):
         self.input_shape = (self.batch_size, self.in_channels
                             ) + self.spatial_shape
 
-        self.offset_shape = (self.batch_size, 2 * filter_shape[0] *
-                             filter_shape[1]) + out_shape
+        self.offset_shape = (self.batch_size, self.deformable_groups * 2 *
+                             filter_shape[0] * filter_shape[1]) + out_shape
 
-        self.mask_shape = (self.batch_size, filter_shape[0] * filter_shape[1]
-                           ) + out_shape
+        self.mask_shape = (self.batch_size, self.deformable_groups *
+                           filter_shape[0] * filter_shape[1]) + out_shape
 
         self.input = np.random.uniform(-1, 1,
                                        self.input_shape).astype(self.dtype)
@@ -89,12 +90,12 @@ class TestDeformConv2D(TestCase):
             x = paddle.static.data(
                 "input", (-1, self.in_channels, -1, -1), dtype=self.dtype)
             offset = paddle.static.data(
-                "offset",
-                (-1, 2 * self.filter_shape[0] * self.filter_shape[1], -1, -1),
+                "offset", (-1, self.deformable_groups * 2 *
+                           self.filter_shape[0] * self.filter_shape[1], -1, -1),
                 dtype=self.dtype)
             mask = paddle.static.data(
-                "mask",
-                (-1, self.filter_shape[0] * self.filter_shape[1], -1, -1),
+                "mask", (-1, self.deformable_groups * self.filter_shape[0] *
+                         self.filter_shape[1], -1, -1),
                 dtype=self.dtype)
 
             y_v1 = paddle.fluid.layers.deformable_conv(
@@ -107,7 +108,7 @@ class TestDeformConv2D(TestCase):
                 padding=self.padding,
                 dilation=self.dilation,
                 groups=self.groups,
-                deformable_groups=1,
+                deformable_groups=self.deformable_groups,
                 im2col_step=1,
                 param_attr=I.Assign(self.weight),
                 bias_attr=False if self.no_bias else I.Assign(self.bias),
@@ -123,7 +124,7 @@ class TestDeformConv2D(TestCase):
                 padding=self.padding,
                 dilation=self.dilation,
                 groups=self.groups,
-                deformable_groups=1,
+                deformable_groups=self.deformable_groups,
                 im2col_step=1,
                 param_attr=I.Assign(self.weight),
                 bias_attr=False if self.no_bias else I.Assign(self.bias))
@@ -154,6 +155,7 @@ class TestDeformConv2D(TestCase):
             stride=self.stride,
             padding=self.padding,
             dilation=self.dilation,
+            deformable_groups=self.deformable_groups,
             groups=self.groups,
             weight_attr=I.Assign(self.weight),
             bias_attr=False if self.no_bias else I.Assign(self.bias))
@@ -194,6 +196,7 @@ class TestDeformConv2DFunctional(TestCase):
         self.padding = [0, 0]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = True
 
@@ -229,11 +232,11 @@ class TestDeformConv2DFunctional(TestCase):
         self.input_shape = (self.batch_size, self.in_channels
                             ) + self.spatial_shape
 
-        self.offset_shape = (self.batch_size, 2 * filter_shape[0] *
-                             filter_shape[1]) + out_shape
+        self.offset_shape = (self.batch_size, self.deformable_groups * 2 *
+                             filter_shape[0] * filter_shape[1]) + out_shape
 
-        self.mask_shape = (self.batch_size, filter_shape[0] * filter_shape[1]
-                           ) + out_shape
+        self.mask_shape = (self.batch_size, self.deformable_groups *
+                           filter_shape[0] * filter_shape[1]) + out_shape
 
         self.input = np.random.uniform(-1, 1,
                                        self.input_shape).astype(self.dtype)
@@ -251,12 +254,12 @@ class TestDeformConv2DFunctional(TestCase):
             x = paddle.static.data(
                 "input", (-1, self.in_channels, -1, -1), dtype=self.dtype)
             offset = paddle.static.data(
-                "offset",
-                (-1, 2 * self.filter_shape[0] * self.filter_shape[1], -1, -1),
+                "offset", (-1, self.deformable_groups * 2 *
+                           self.filter_shape[0] * self.filter_shape[1], -1, -1),
                 dtype=self.dtype)
             mask = paddle.static.data(
-                "mask",
-                (-1, self.filter_shape[0] * self.filter_shape[1], -1, -1),
+                "mask", (-1, self.deformable_groups * self.filter_shape[0] *
+                         self.filter_shape[1], -1, -1),
                 dtype=self.dtype)
 
             y_v1 = paddle.fluid.layers.deformable_conv(
@@ -269,7 +272,7 @@ class TestDeformConv2DFunctional(TestCase):
                 padding=self.padding,
                 dilation=self.dilation,
                 groups=self.groups,
-                deformable_groups=1,
+                deformable_groups=self.deformable_groups,
                 im2col_step=1,
                 param_attr=I.Assign(self.weight),
                 bias_attr=False if self.no_bias else I.Assign(self.bias),
@@ -285,7 +288,7 @@ class TestDeformConv2DFunctional(TestCase):
                 padding=self.padding,
                 dilation=self.dilation,
                 groups=self.groups,
-                deformable_groups=1,
+                deformable_groups=self.deformable_groups,
                 im2col_step=1,
                 param_attr=I.Assign(self.weight),
                 bias_attr=False if self.no_bias else I.Assign(self.bias))
@@ -317,6 +320,7 @@ class TestDeformConv2DFunctional(TestCase):
             stride=self.stride,
             padding=self.padding,
             dilation=self.dilation,
+            deformable_groups=self.deformable_groups,
             groups=self.groups, )
 
         y_v2 = paddle.vision.ops.deform_conv2d(
@@ -328,6 +332,7 @@ class TestDeformConv2DFunctional(TestCase):
             stride=self.stride,
             padding=self.padding,
             dilation=self.dilation,
+            deformable_groups=self.deformable_groups,
             groups=self.groups, )
 
         out_v1 = y_v1.numpy()
@@ -343,12 +348,12 @@ class TestDeformConv2DFunctional(TestCase):
             x = paddle.static.data(
                 "input", (-1, self.in_channels, -1, -1), dtype=self.dtype)
             offset = paddle.static.data(
-                "offset",
-                (-1, 2 * self.filter_shape[0] * self.filter_shape[1], -1, -1),
+                "offset", (-1, self.deformable_groups * 2 *
+                           self.filter_shape[0] * self.filter_shape[1], -1, -1),
                 dtype=self.dtype)
             mask = paddle.static.data(
-                "mask",
-                (-1, self.filter_shape[0] * self.filter_shape[1], -1, -1),
+                "mask", (-1, self.deformable_groups * self.filter_shape[0] *
+                         self.filter_shape[1], -1, -1),
                 dtype=self.dtype)
 
             weight = paddle.static.data(
@@ -365,6 +370,7 @@ class TestDeformConv2DFunctional(TestCase):
                 stride=self.stride,
                 padding=self.padding,
                 dilation=self.dilation,
+                deformable_groups=self.deformable_groups,
                 groups=self.groups, )
 
             y_v2 = paddle.vision.ops.deform_conv2d(
@@ -376,6 +382,7 @@ class TestDeformConv2DFunctional(TestCase):
                 stride=self.stride,
                 padding=self.padding,
                 dilation=self.dilation,
+                deformable_groups=self.deformable_groups,
                 groups=self.groups, )
 
         exe = paddle.static.Executor(self.place)
@@ -421,6 +428,7 @@ class TestDeformConv2DWithPadding(TestDeformConv2D):
         self.padding = [2, 2]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = True
 
@@ -433,6 +441,7 @@ class TestDeformConv2DWithBias(TestDeformConv2D):
         self.padding = [2, 2]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = False
 
@@ -445,6 +454,7 @@ class TestDeformConv2DWithAsynPadding(TestDeformConv2D):
         self.padding = [1, 2]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = False
 
@@ -457,6 +467,7 @@ class TestDeformConv2DWithDilation(TestDeformConv2D):
         self.padding = [1, 1]
         self.stride = [1, 1]
         self.dilation = [3, 3]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = False
 
@@ -469,6 +480,20 @@ class TestDeformConv2DWithStride(TestDeformConv2D):
         self.padding = [1, 1]
         self.stride = [2, 2]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
+        self.groups = 1
+        self.no_bias = False
+
+
+class TestDeformConv2DWithDeformable_Groups(TestDeformConv2D):
+    def setUp(self):
+        self.in_channels = 5
+        self.out_channels = 5
+        self.kernel_size = [3, 3]
+        self.padding = [1, 1]
+        self.stride = [1, 1]
+        self.dilation = [1, 1]
+        self.deformable_groups = 5
         self.groups = 1
         self.no_bias = False
 
@@ -481,6 +506,7 @@ class TestDeformConv2DWithGroups(TestDeformConv2D):
         self.padding = [1, 1]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 5
         self.no_bias = False
 
@@ -494,6 +520,7 @@ class TestDeformConv2DFunctionalWithPadding(TestDeformConv2DFunctional):
         self.padding = [2, 2]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = True
 
@@ -506,6 +533,7 @@ class TestDeformConv2DFunctionalWithBias(TestDeformConv2DFunctional):
         self.padding = [2, 2]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = False
 
@@ -518,6 +546,7 @@ class TestDeformConv2DFunctionalWithAsynPadding(TestDeformConv2DFunctional):
         self.padding = [1, 2]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = False
 
@@ -530,6 +559,7 @@ class TestDeformConv2DFunctionalWithDilation(TestDeformConv2DFunctional):
         self.padding = [1, 1]
         self.stride = [1, 1]
         self.dilation = [3, 3]
+        self.deformable_groups = 1
         self.groups = 1
         self.no_bias = False
 
@@ -542,6 +572,21 @@ class TestDeformConv2DFunctionalWithStride(TestDeformConv2DFunctional):
         self.padding = [1, 1]
         self.stride = [2, 2]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
+        self.groups = 1
+        self.no_bias = False
+
+
+class TestDeformConv2DFunctionalWithDeformable_Groups(
+        TestDeformConv2DFunctional):
+    def setUp(self):
+        self.in_channels = 5
+        self.out_channels = 5
+        self.kernel_size = [3, 3]
+        self.padding = [1, 1]
+        self.stride = [1, 1]
+        self.dilation = [1, 1]
+        self.deformable_groups = 5
         self.groups = 1
         self.no_bias = False
 
@@ -554,6 +599,7 @@ class TestDeformConv2DFunctionalWithGroups(TestDeformConv2DFunctional):
         self.padding = [1, 1]
         self.stride = [1, 1]
         self.dilation = [1, 1]
+        self.deformable_groups = 1
         self.groups = 5
         self.no_bias = False
 

python/paddle/vision/ops.py

@@ -398,6 +398,7 @@ def deform_conv2d(x,
                   stride=1,
                   padding=0,
                   dilation=1,
+                  deformable_groups=1,
                   groups=1,
                   mask=None,
                   name=None):
@@ -462,6 +463,8 @@ def deform_conv2d(x,
         dilation (int|list|tuple, optional): The dilation size. If dilation is a tuple, it must
             contain two integers, (dilation_H, dilation_W). Otherwise, the
             dilation_H = dilation_W = dilation. Default: dilation = 1.
+        deformable_groups (int): The number of deformable group partitions.
+            Default: deformable_groups = 1.
         groups (int, optonal): The groups number of the deformable conv layer. According to
             grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
             the first half of the filters is only connected to the first half
@@ -521,7 +524,8 @@ def deform_conv2d(x,
 
     if in_dygraph_mode():
         attrs = ('strides', stride, 'paddings', padding, 'dilations', dilation,
-                 'groups', groups, 'im2col_step', 1)
+                 'deformable_groups', deformable_groups, 'groups', groups,
+                 'im2col_step', 1)
         if use_deform_conv2d_v1:
             op_type = 'deformable_conv_v1'
             pre_bias = getattr(core.ops, op_type)(x, offset, weight, *attrs)
@@ -572,7 +576,7 @@ def deform_conv2d(x,
             'paddings': padding,
             'dilations': dilation,
             'groups': groups,
-            'deformable_groups': 1,
+            'deformable_groups': deformable_groups,
             'im2col_step': 1,
         }
         helper.append_op(
@@ -649,6 +653,8 @@ class DeformConv2D(Layer):
         dilation(int|list|tuple, optional): The dilation size. If dilation is a tuple, it must
             contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the
             dilation_D = dilation_H = dilation_W = dilation. The default value is 1.
+        deformable_groups (int): The number of deformable group partitions.
+            Default: deformable_groups = 1.
         groups(int, optional): The groups number of the Conv3D Layer. According to grouped
             convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
             the first half of the filters is only connected to the first half
@@ -726,6 +732,7 @@ class DeformConv2D(Layer):
                  stride=1,
                  padding=0,
                  dilation=1,
+                 deformable_groups=1,
                  groups=1,
                  weight_attr=None,
                  bias_attr=None):
@@ -733,6 +740,7 @@ class DeformConv2D(Layer):
         assert weight_attr is not False, "weight_attr should not be False in Conv."
         self._weight_attr = weight_attr
         self._bias_attr = bias_attr
+        self._deformable_groups = deformable_groups
         self._groups = groups
         self._in_channels = in_channels
         self._out_channels = out_channels
@@ -770,6 +778,7 @@ class DeformConv2D(Layer):
             stride=self._stride,
             padding=self._padding,
             dilation=self._dilation,
+            deformable_groups=self._deformable_groups,
             groups=self._groups,
             mask=mask)
         return out