Support YOLOv8 for dynamic shape

x2paddle/decoder/onnx_decoder.py

@@ -308,7 +308,14 @@ def build_connection(self, layer_name, node):
                                             break
                             else:
                                 first_i = node.inputs.index(nd.name)
-                                node.which_child[nd.name] = idx
+                                # deal with Multiple outputs correspond to one node
+                                if self.node_map[nd.name].outputs.count(
+                                        layer_name) > 1:
+                                    new_child_name = "{}/{}".format(nd.name,
+                                                                    idx)
+                                    node.which_child[new_child_name] = idx
+                                else:
+                                    node.which_child[nd.name] = idx
                             self.node_map[nd.name].index = 0
                             break
                     if flag == 1:

x2paddle/op_mapper/onnx2paddle/opset13.py

@@ -38,7 +38,8 @@ def Unsqueeze(self, node):
         val_x = self.graph.get_input_node(node, idx=0, copy=True)
         axes = self.graph.get_input_node(node, idx=1, copy=True)
         # deal with scalar(0D) tensor
-        if len(val_x.out_shapes[0]) == 0 and len(axes.out_shapes[0]) == 1:
+        if len(val_x.out_shapes[0]) == 0 and len(axes.out_shapes[
+                0]) == 1 and len(node.out_shapes[0]) == 1:
             self.paddle_graph.add_layer(
                 'paddle.reshape',
                 inputs={"x": val_x.name},

x2paddle/op_mapper/onnx2paddle/opset_legacy.py

@@ -2309,8 +2309,16 @@ def Conv(self, node):
                 rename_mapper=self.rename_mapper)
         else:
             layer_attrs["bias_attr"] = False
-        if reduce(lambda x, y: x * y,
-                  input_shape) in [1, -1] and 1 not in input_shape:
+        # deal with dynamic shape
+        if len(input_shape) == 0:
+            self.paddle_graph.add_layer(
+                "paddle.reshape",
+                inputs=layer_inputs,
+                outputs=[layer_inputs["x"]],
+                shape=[0, num_in_channels * num_groups, 0, -1])
+        if len(input_shape) != 0 and reduce(
+                lambda x, y: x * y,
+                input_shape) in [1, -1] and 1 not in input_shape:
             input_shape[1] = num_in_channels * num_groups
             input_shape[0] = 0
             input_shape[2] = 0