Support ONNX Roberta model (#793)

* add replication pad

* update op_list.md

* re-lint

* fixed aten::index

* rm useless code

* Support Wav2vec2

* Support onnx roberta

* update op_list.md

* deal with comments

* deal with comments

docs/inference_model_convertor/op_list.md

@@ -76,6 +76,7 @@
 | 81 | Add | 82 | Concat | 83 | Max | 84 | Min |
 | 85 | GreaterOrEqual | 86 | GatherND | 87 | And | 88 | cos |
 | 89 | Neg | 90 | SpaceToDepth | 91 | GatherElement | 92 | Sin |
+| 93 | CumSum |  |  |  |  |  |  |
 
 ## PyTorch
 

x2paddle/op_mapper/onnx2paddle/opset9/opset.py

@@ -457,6 +457,19 @@ class OpSet9():
             outputs=[node.name],
             **attrs)
 
+    @print_mapping_info
+    def CumSum(self, node):
+        val_x = self.graph.get_input_node(node, idx=0, copy=True)
+        axis = self.graph.get_input_node(node, idx=1, copy=True)
+        axis_values = _const_weight_or_none(axis)
+        assert axis_values is not None, 'Axis only support constant tensor!'
+        layer_attrs = {'axis': axis_values}
+        self.paddle_graph.add_layer(
+            'paddle.cumsum',
+            inputs={"x": val_x.name},
+            outputs=[node.name],
+            **layer_attrs)
+
     @print_mapping_info
     def HardSigmoid(self, node):
         val_x = self.graph.get_input_node(node, idx=0, copy=True)