Merge pull request #471 from PaddlePaddle/onnx

fix the onnx bug

x2paddle/decoder/onnx_decoder.py

@@ -205,7 +205,7 @@ class ONNXGraph(Graph):
                             shape = raw_input(
                                 "Shape of Input(e.g. -1,3,224,224), enter 'N' to skip: "
                             )
-                        except:
+                        except NameError:
                             shape = input(
                                 "Shape of Input(e.g. -1,3,224,224), enter 'N' to skip: "
                             )
@@ -302,6 +302,17 @@ class ONNXGraph(Graph):
                         if opt == in_node:
                             self.connect(nd.name, layer_name)
                             flag = 1
+                            if nd.name in node.which_child:
+                                for n_i, n_ipt in enumerate(node.inputs):
+                                    if first_i == n_i:
+                                        continue
+                                    if n_ipt == nd.name:
+                                        new_nd_name = "{}/{}".format(nd.name, n_i)
+                                        if new_nd_name not in node.which_child:
+                                            node.which_child[new_nd_name] = idx
+                                            break
+                            else:
+                                first_i = node.inputs.index(nd.name)
                                 node.which_child[nd.name] = idx
                             self.node_map[nd.name].index = 0
                             break
@@ -318,11 +329,15 @@ class ONNXGraph(Graph):
         if len(node.which_child) == 0:
             ipt_node = super(ONNXGraph, self).get_node(node.inputs[idx], copy)
             return ipt_node
-
         else:
             ipt_node = super(ONNXGraph, self).get_node(node.inputs[idx], copy)
+            new_ipt_name = "{}/{}".format(ipt_node.layer_name, idx)
+            if new_ipt_name in node.which_child:
+                ipt_node.index = node.which_child[new_ipt_name]
+            else:
                 if ipt_node.layer_name in node.which_child:
                     ipt_node.index = node.which_child[ipt_node.layer_name]
+                
             return ipt_node
         
 

x2paddle/op_mapper/dygraph/onnx2paddle/opset9/opset.py

@@ -250,15 +250,22 @@ class OpSet9():
     def _interpolate(self, node):
         val_x = self.graph.get_input_node(node, idx=0, copy=True)
         inputs = {'x': val_x.name}
+        attrs = dict()
         if node.layer_type == 'Resize':
             if len(node.layer.input) == 2:
                 # opset 10
                 val_scales = self.graph.get_input_node(node, idx=1, copy=True)
-                inputs['scale_factor'] = val_scales.name
+                # TODO(syf): paddle.nn.functional.interpolate will support the length  
+                # which is the same as the rank of input.
+#                 inputs['scale_factor'] = val_scales.name
+                attrs['scale_factor'] = self.weights[val_scales.name].tolist()[2:]
             elif len(node.layer.input) == 3:
                 # opset 11
                 val_scales = self.graph.get_input_node(node, idx=2, copy=True)
-                inputs['scale_factor'] = val_scales.name
+                # TODO(syf): paddle.nn.functional.interpolate will support the length  
+                # which is the same as the rank of input.
+#                 inputs['scale_factor'] = val_scales.name
+                attrs['scale_factor'] = self.weights[val_scales.name].tolist()[2:]
             elif len(node.layer.input) == 4:
                 # opset 11
                 val_sizes = self.graph.get_input_node(node, idx=3, copy=True)
@@ -281,7 +288,7 @@ class OpSet9():
                 ipt = inputs.pop("x")
                 inputs["input"] = ipt
                 mode = node.get_attr('mode', 'nearest')
-                attrs = {"align_corners": False}
+                attrs.update({"align_corners": False})
                 self.paddle_graph.add_layer(
                     kernel="fluid.layers.resize_nearest",
                     inputs=inputs,
@@ -290,12 +297,12 @@ class OpSet9():
                 return
         elif node.layer_type == 'Upsample':
             val_scales = self.graph.get_input_node(node, idx=1, copy=True)
-            inputs['scale'] = val_scales
+            inputs['scale_factor'] = val_scales
 
         mode = node.get_attr('mode', 'nearest')
-        attrs = {"align_corners": False,
+        attrs.update({"align_corners": False,
                       "mode": string(mode),
-                 "align_mode": 1}
+                      "align_mode": 1})
         self.paddle_graph.add_layer(
             kernel="paddle.nn.functional.interpolate",
             inputs=inputs,
@@ -926,16 +933,17 @@ class OpSet9():
                 'max': max_value,
                 'min': min_value,
             }
+            
             self.paddle_graph.add_layer(
                 'paddle.clip', 
                 inputs={"x": val_x.name}, 
                 outputs=[node.name], 
                 **layer_attrs)
         else:
-            max_ipt = self.graph.get_input_node(node, idx=1, copy=True)
-            min_ipt = self.graph.get_input_node(node, idx=2, copy=True)
-            max_value = _const_weight_or_none(max_ipt)
+            min_ipt = self.graph.get_input_node(node, idx=1, copy=True)
+            max_ipt = self.graph.get_input_node(node, idx=2, copy=True)
             min_value = _const_weight_or_none(min_ipt)
+            max_value = _const_weight_or_none(max_ipt)
             if max_value.shape == (1, ):
                 max_value = max_value[0]
             if min_value.shape == (1, ):

x2paddle/op_mapper/static/onnx2paddle/opset9/opset.py

@@ -240,15 +240,22 @@ class OpSet9():
     def _interpolate(self, node):
         val_x = self.graph.get_input_node(node, idx=0, copy=True)
         inputs = {'x': val_x.name}
+        attrs = dict()
         if node.layer_type == 'Resize':
             if len(node.layer.input) == 2:
                 # opset 10
                 val_scales = self.graph.get_input_node(node, idx=1, copy=True)
-                inputs['scale_factor'] = val_scales.name
+                # TODO(syf): paddle.nn.functional.interpolate will support the length  
+                # which is the same as the rank of input.
+#                 inputs['scale_factor'] = val_scales.name
+                attrs['scale_factor'] = self.params[val_scales.name].tolist()[2:]
             elif len(node.layer.input) == 3:
                 # opset 11
                 val_scales = self.graph.get_input_node(node, idx=2, copy=True)
-                inputs['scale_factor'] = val_scales.name
+                # TODO(syf): paddle.nn.functional.interpolate will support the length  
+                # which is the same as the rank of input.
+#                 inputs['scale_factor'] = val_scales.name
+                attrs['scale_factor'] = self.params[val_scales.name].tolist()[2:]
             elif len(node.layer.input) == 4:
                 # opset 11
                 val_sizes = self.graph.get_input_node(node, idx=3, copy=True)
@@ -271,7 +278,7 @@ class OpSet9():
                 ipt = inputs.pop("x")
                 inputs["input"] = ipt
                 mode = node.get_attr('mode', 'nearest')
-                attrs = {"align_corners": False}
+                attrs.update({"align_corners": False})
                 self.paddle_graph.add_layer(
                     kernel="fluid.layers.resize_nearest",
                     inputs=inputs,
@@ -283,9 +290,9 @@ class OpSet9():
             inputs['scale'] = val_scales
 
         mode = node.get_attr('mode', 'nearest')
-        attrs = {"align_corners": False,
+        attrs.update({"align_corners": False,
                  "mode": string(mode),
-                 "align_mode": 1}
+                 "align_mode": 1})
         self.paddle_graph.add_layer(
             kernel="paddle.nn.functional.interpolate",
             inputs=inputs,
@@ -917,10 +924,10 @@ class OpSet9():
                 outputs=[node.name], 
                 **layer_attrs)
         else:
-            max_ipt = self.graph.get_input_node(node, idx=1, copy=True)
-            min_ipt = self.graph.get_input_node(node, idx=2, copy=True)
-            max_value = _const_weight_or_none(max_ipt)
+            min_ipt = self.graph.get_input_node(node, idx=1, copy=True)
+            max_ipt = self.graph.get_input_node(node, idx=2, copy=True)
             min_value = _const_weight_or_none(min_ipt)
+            max_value = _const_weight_or_none(max_ipt)
             if max_value.shape == (1, ):
                 max_value = max_value[0]
             if min_value.shape == (1, ):