add setattr

x2paddle/core/program.py

@@ -101,6 +101,11 @@ class PaddleGraph(object):
         self.clear_edges()
         outputs_from_nodes = dict()
         for layer_id, layer in self.layers.items():
+            #             if "x5109" in layer.outputs or "x5110" in layer.outputs:
+            #                 print(layer.kernel)
+            #                 print(layer.inputs)
+            #                 print(layer.outputs)
+            #                 print(layer.attrs)
             for input_key, input_var in layer.inputs.items():
                 vs = input_var
                 if not isinstance(vs, list):

x2paddle/op_mapper/pytorch2paddle/prim.py

@@ -44,7 +44,7 @@ def prim_GetAttr(mapper, graph, node):
         %7 (Tensor): 输入Tensor。
         %27 (Tensor): 输入Tensor。
     """
-    output_name = mapper._get_outputs_name(node)[0]
+    current_node = node
     field_name_list = [node.s('name')]
     while True:
         input_node = list(node.inputs())[0].node()
@@ -53,18 +53,16 @@ def prim_GetAttr(mapper, graph, node):
             node = input_node
         except Exception:
             break
-    if ".".join(field_name_list) in mapper.pytorch_params:
-        mapper.pytorch_params[output_name] = mapper.pytorch_params[".".join(
-            field_name_list)]
-    else:
-        part_script = mapper.script
-        for field_name in field_name_list:
-            if hasattr(part_script, field_name):
-                param = getattr(part_script, field_name)
-                if isinstance(param, torch.Tensor):
-                    param = param.detach().numpy()
-                mapper.pytorch_params[output_name] = param
-                part_script = param
+    attr_name = ".".join(field_name_list)
+    output_name = mapper._get_outputs_name(current_node, attr_name)[0]
+    part_script = mapper.script
+    for field_name in field_name_list:
+        if hasattr(part_script, field_name):
+            param = getattr(part_script, field_name)
+            if isinstance(param, torch.Tensor):
+                param = param.detach().numpy()
+            mapper.pytorch_params[output_name] = param
+            part_script = param
     return [], [output_name]
 
 
@@ -295,8 +293,15 @@ def prim_SetAttr(mapper, graph, node):
     field_name_list.append(node.s('name'))
 
     inputs_name, inputs_node = mapper._get_inputs_name(node)
-    param = {"Tensor": inputs_name[1]}
+    param = {
+        "Tensor": "self." + ".".join(field_name_list).replace(".", "_"),
+        "parent_layer_id": graph.parent_layer.id
+    }
     mapper.pytorch_params[".".join(field_name_list)] = param
+    graph.add_layer(
+        "prim.set_attr",
+        inputs={"input": inputs_name[1]},
+        outputs=["self." + ".".join(field_name_list).replace(".", "_")])
     return [], [output_name]
 
 

x2paddle/op_mapper/pytorch2paddle/prim2code.py

@@ -208,6 +208,11 @@ def prim_select(layer, indent=1, init_func=[], forward_func=[]):
     forward_func.extend(gen_codes([line], indent=indent))
 
 
+def prim_set_attr(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {}".format(layer.outputs[0], get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
 def prim_shape(layer, indent=1, init_func=[], forward_func=[]):
     line = "{} = {}.shape".format(layer.outputs[0], get_value(layer, "input"))
     forward_func.extend(gen_codes([line], indent=indent))

x2paddle/op_mapper/pytorch2paddle/pytorch_op_mapper.py

@@ -113,15 +113,19 @@ class PyTorchOpMapper(OpMapper):
             graph.set_parameters(self.paddle_params)
         return graph, graph_inputs
 
-    def _get_outputs_name(self, node):
+    def _get_outputs_name(self, node, attr_name=None):
         outputs_name = []
         for output_ivalue in node.outputs():
-            output_name = 'x' + str(self.output_index)
             script_unique_id = output_ivalue.unique()
-            if script_unique_id in self.outputs_info:
-                output_name = self.outputs_info[script_unique_id]
+            if attr_name is None:
+                output_name = 'x' + str(self.output_index)
+                if script_unique_id in self.outputs_info:
+                    output_name = self.outputs_info[script_unique_id]
+            else:
+                output_name = attr_name.replace(".", "_")
             self.outputs_info[script_unique_id] = output_name
             self.output_index += 1
+
             outputs_name.append(output_name)
         # if或loop节点没有输出的情况
         if len(list(node.outputs())) == 0:
@@ -148,7 +152,33 @@ class PyTorchOpMapper(OpMapper):
                     outputs=[output_name],
                     value="params[{}]".format(string(output_name)))
             else:
-                if isinstance(param, dict) and "Tensor" in param:
+                if isinstance(param, dict) and "Tensor" in param and \
+                "parent_layer_id" in param:
+                    if graph.parent_layer is not None:
+                        # 当某个param被2个控制流（if-else）赋值时，else不可以引用if中的赋值结果
+                        id1 = param["parent_layer_id"]
+                        id2 = graph.parent_layer.id
+                        id1_part = id1.split(".")
+                        id2_part = id2.split(".")
+                        if len(id1_part) >= len(id2_part):
+                            for i in range(len(id1_part)):
+                                if id1_part[i] == id2_part[i]:
+                                    continue
+                                else:
+                                    if id1_part[i] == "0" and id2_part[
+                                            i] == "1":
+                                        if add_dim:
+                                            param = param[np.newaxis, :]
+                                        self.paddle_params[output_name] = param
+                                        graph.add_layer(
+                                            "fluid.dygraph.base.to_variable",
+                                            inputs={},
+                                            outputs=[output_name],
+                                            value="params[{}]".format(
+                                                string(output_name)))
+                                        node_outputs.append(output_name)
+                                        return
+                    # 若if-else外，则可直接引用if-else中的赋值结果
                     graph.add_layer(
                         "prim.constant",
                         inputs={},

x2paddle/optimizer/fusion/__init__.py

@@ -20,3 +20,5 @@ from .functional_adaptive_pool2d_fuser import FunctionalAdaptivePool2dFuser
 from .functional_adaptive_pool2d_fuse_pass import FunctionalAdaptivePool2dFusePass
 from .constant_fuser import ConstantFuser
 from .constant_fuse_pass import ConstantFusePass
+from .batchnorm2d_fuser import BatchNorm2dFuser
+from .batchnorm2d_fuse_pass import BatchNorm2dFusePass

x2paddle/optimizer/optimizer.py

@@ -20,7 +20,8 @@ class GraphOptimizer(object):
     def __init__(self):
         self.passes = [
             "fc_fuse_pass", "nn_adaptive_pool2d_fuse_pass",
-            "functional_adaptive_pool2d_fuse_pass", "constant_fuse_pass"
+            "functional_adaptive_pool2d_fuse_pass", "batchnorm2d_fuse_pass",
+            "constant_fuse_pass"
         ]
 
     def optimize(self, graph):