Don't add crop layer if output_padding is not used for deconvolution (#496)

Author: bhushan23

onnx_coreml/_operators.py

@@ -302,14 +302,6 @@ def _get_conv_params(builder, node, graph, err, params_dict, axis=None):
             pads = [0, pads[0], 0, pads[1]]
         params_dict['pads'] = pads
 
-    if params_dict['is_deconv']:
-        params_dict['crops'] = copy.copy(params_dict['pads'])
-        params_dict['pads'] = [0, 0, 0, 0]
-        if sum(params_dict['crops']) == 0:
-            params_dict['is_post_crop'] = False
-        else:
-            params_dict['is_post_crop'] = True
-
     if "kernel_shape" in node.attrs:
         params_dict['kernel_shape'] = node.attrs["kernel_shape"]
     else:
@@ -332,7 +324,6 @@ def _get_conv_params(builder, node, graph, err, params_dict, axis=None):
         params_dict['strides'].insert(0,1)
         params_dict['kernel_shape'].insert(0,1)
 
-        
     params_dict['out_shape'] = None
     params_dict['padding_type'] = 'valid'
     params_dict['same_padding_asymmetry_mode'] = 'BOTTOM_RIGHT_HEAVY'
@@ -358,6 +349,8 @@ def _get_conv_params(builder, node, graph, err, params_dict, axis=None):
             else:
                 params_dict['out_shape'] = (node.attrs['output_shape'][-2], node.attrs['output_shape'][-1])  # (Hout, wout)
         elif 'output_padding' in node.attrs:
+            params_dict['crops'] = copy.copy(params_dict['pads'])
+            params_dict['pads'] = [0, 0, 0, 0]
             post_pads = node.attrs['output_padding']
             if sum(post_pads) != 0:
                 t = l = b = r = 0
@@ -432,7 +425,6 @@ def _add_conv(input_names, output_names, **kwargs):
             output_name=input_names[0],
             value=0
         )
-    
     builder.add_convolution(
         name=node.name,
         kernel_channels=kc,

tests/pytorch_model_test.py

@@ -46,7 +46,6 @@ def _test_torch_model_single_io(torch_model, torch_input_shape, coreml_input_sha
 
     # convert to coreml and run
     coreml_model = convert(onnx_model, target_ios=target_ios)
-
     output_name = [o.name for o in onnx_model.graph.output][0]
     initializer_names = {t.name for t in onnx_model.graph.initializer}
     input_name = [i.name for i in onnx_model.graph.input if i.name not in initializer_names][0]
@@ -157,15 +156,29 @@ def test_conv2D_transpose(self): # type: () -> None
         class Net(nn.Module):
             def __init__(self):
                 super(Net, self).__init__()
-                self.convT = torch.nn.ConvTranspose2d(1, 1, kernel_size=3, stride=2, output_padding=1, padding=1, groups=1)
+                self.convT = torch.nn.ConvTranspose2d(1, 1, kernel_size=3, stride=2, output_padding=0, padding=3, groups=1)
+
+            def forward(self, x):
+                y = self.convT(x)
+                return y
+
+        torch_model = Net()  # type: ignore
+        torch_model.train(False)
+        _test_torch_model_single_io(torch_model, (1, 1, 64, 64), (1, 64, 64))  # type: ignore
+
+    def test_conv2D_transpose_output_padding(self): # type: () -> None
+        class Net(nn.Module):
+            def __init__(self):
+                super(Net, self).__init__()
+                self.convT = torch.nn.ConvTranspose2d(1, 1, kernel_size=3, stride=2, output_padding=1, padding=3, groups=1)
 
             def forward(self, x):
                 y = self.convT(x)
                 return y
 
         torch_model = Net()  # type: ignore
         torch_model.train(False)
-        _test_torch_model_single_io(torch_model, (1, 1, 2, 2), (1, 2, 2))  # type: ignore
+        _test_torch_model_single_io(torch_model, (1, 1, 64, 64), (1, 64, 64))  # type: ignore
 
     def test_conv2D_transpose_groups(self): # type: () -> None
         class Net(nn.Module):