!906 fix a bug that support dilation greater than 1 in conv2dTranspose ops

Merge pull request !906 from yangyongjie/master

mindspore/nn/layer/conv.py

@@ -358,6 +358,7 @@ class Conv2dTranspose(_Conv):
     def _deconv_output_length(self, input_length, filter_size, stride_size, dilation_size):
         """Calculate the width and height of output."""
         length = 0
+        filter_size = filter_size + (filter_size - 1) * (dilation_size - 1)
         if self.is_valid:
             if filter_size - stride_size > 0:
                 length = input_length * stride_size + filter_size - stride_size
@@ -366,8 +367,7 @@ class Conv2dTranspose(_Conv):
         elif self.is_same:
             length = input_length * stride_size
         elif self.is_pad:
-            length = input_length * stride_size - 2 * self.padding + filter_size + \
-                     (filter_size - 1) * (dilation_size - 1) - stride_size
+            length = input_length * stride_size - 2 * self.padding + filter_size - stride_size
 
         return length
 

mindspore/ops/operations/nn_ops.py

@@ -1186,16 +1186,18 @@ class Conv2DBackpropInput(PrimitiveWithInfer):
         kernel_w = self.kernel_size[1]
         stride_h = self.stride[0]
         stride_w = self.stride[1]
+        dilation_h = self.dilation[2]
+        dilation_w = self.dilation[3]
         # default pad mode is valid
         pad_list = (0, 0, 0, 0)
         if self.pad_list:
             pad_list = tuple(self.pad_list)
         elif self.pad_mode == "SAME":
-            pad_needed_h = max(0, (dout_shape[2] - 1) * stride_h + kernel_h - x_size_v[2])
+            pad_needed_h = max(0, (dout_shape[2] - 1) * stride_h + dilation_h * (kernel_h - 1) + 1 - x_size_v[2])
             pad_top = math.floor(pad_needed_h / 2)
             pad_bottom = pad_needed_h - pad_top
 
-            pad_needed_w = max(0, (dout_shape[3] - 1) * stride_w + kernel_w - x_size_v[3])
+            pad_needed_w = max(0, (dout_shape[3] - 1) * stride_w + dilation_w * (kernel_w - 1) + 1 - x_size_v[3])
             pad_left = math.floor(pad_needed_w / 2)
             pad_right = pad_needed_w - pad_left
             pad_list = (pad_top, pad_bottom, pad_left, pad_right)

tests/ut/python/nn/test_conv.py

@@ -190,3 +190,15 @@ def test_compile_transpose_stride2():
     net = NetConv2dTranspose(3, 64, 4, stride=2, weight_init='normal')
     input_data = Tensor(np.ones([1, 3, 16, 50], dtype=np.float32))
     net(input_data)
+
+
+def test_compile_transpose_dilation_2():
+    net = NetConv2dTranspose(3, 64, 4, stride=2, dilation=2, pad_mode='same', weight_init='normal')
+    input_data = Tensor(np.ones([1, 3, 16, 50], dtype=np.float32))
+    net(input_data)
+
+
+def test_compile_transpose_dilation_2_pad_mode_pad():
+    net = NetConv2dTranspose(3, 64, 4, stride=2, dilation=2, pad_mode='pad', weight_init='normal')
+    input_data = Tensor(np.ones([1, 3, 16, 50], dtype=np.float32))
+    net(input_data)