[PaddleV3] 修复 pytorch 中 aten_linear 的算子映射并修复相关模型

test_benchmark/PyTorch/MiniFasNet/convert.py

@@ -11,7 +11,7 @@
 import torch
 import numpy as np
 from torchvision.models import AlexNet
-from torchvision.models.utils import load_state_dict_from_url
+# from torchvision.models.utils import load_state_dict_from_url
 from x2paddle.convert import pytorch2paddle
 # 构建输入,pytorch-to-paddle
 input_data = np.random.rand(1, 3, 80, 80).astype("float32")

test_benchmark/PyTorch/MiniFasNet/deploy_infer.py

@@ -4,7 +4,6 @@
 
 import numpy as np
 import paddle
-import paddle.fluid as fluid
 from paddle.inference import Config
 from paddle.inference import create_predictor
 

test_benchmark/PyTorch/MiniFasNet/pd_infer.py

@@ -1,5 +1,4 @@
 from __future__ import print_function
-import paddle.fluid as fluid
 import paddle
 import sys
 import os
@@ -13,11 +12,8 @@
     # trace
     paddle.enable_static()
     exe = paddle.static.Executor(paddle.CPUPlace())
-    [prog, inputs, outputs] = fluid.io.load_inference_model(
-        dirname="pd_model_trace/inference_model/",
-        executor=exe,
-        model_filename="model.pdmodel",
-        params_filename="model.pdiparams")
+    [prog, inputs, outputs] = paddle.static.load_inference_model(
+        path_prefix="pd_model_trace/inference_model/model", executor=exe)
     result = exe.run(prog, feed={inputs[0]: img}, fetch_list=outputs)
     df = pytorch_output - result
     if np.max(np.fabs(df)) > 1e-03:

test_benchmark/PyTorch/SwinTransformer/convert_trace.py

@@ -16,7 +16,8 @@
 model_name = "swin_tiny_patch4_window7_224"
 torch_module = SwinTransformer(**swin_model_cfg_map[model_name])
 torch_state_dict = torch.load(
-    "../dataset/SwinTransformer/{}.pth".format(model_name))["model"]
+    "../dataset/SwinTransformer/{}.pth".format(model_name),
+    map_location=torch.device('cpu'))["model"]
 torch_module.load_state_dict(torch_state_dict)
 model_name = "pd_model_trace"
 # 设置为eval模式

test_benchmark/PyTorch/SwinTransformer/deploy_infer.py

@@ -4,7 +4,6 @@
 
 import numpy as np
 import paddle
-import paddle.fluid as fluid
 from paddle.inference import Config
 from paddle.inference import create_predictor
 

test_benchmark/PyTorch/SwinTransformer/pd_infer.py

@@ -1,5 +1,4 @@
 from __future__ import print_function
-import paddle.fluid as fluid
 import paddle
 import sys
 import os
@@ -14,11 +13,8 @@
     # trace
     paddle.enable_static()
     exe = paddle.static.Executor(paddle.CPUPlace())
-    [prog, inputs, outputs] = fluid.io.load_inference_model(
-        dirname="pd_model_trace/inference_model/",
-        executor=exe,
-        model_filename="model.pdmodel",
-        params_filename="model.pdiparams")
+    [prog, inputs, outputs] = paddle.static.load_inference_model(
+        path_prefix="pd_model_trace/inference_model/model", executor=exe)
     result = exe.run(prog, feed={inputs[0]: input_data}, fetch_list=outputs)
     df = pytorch_output - result
     if numpy.max(numpy.fabs(df)) > 1e-04:

test_benchmark/PyTorch/black.list

@@ -1,8 +1,6 @@
 EasyOCR_detector
 EasyOCR_recognizer
-MiniFasNet
 MockingBird
-SwinTransformer
 dataset
 tools
 output

x2paddle/op_mapper/pytorch2paddle/aten.py

@@ -3441,8 +3441,23 @@ def aten_linear(mapper, graph, node):
                     scope_name=scope_name,
                     **layer_attrs)
     if len(inputs_name) == 3:
-        mapper._check_input(graph, inputs_node[2], inputs_name[2],
-                            current_outputs, scope_name)
+        # make `bias` dtype like `weight`, and shape is `1` for broadcast
+        if (inputs_name[2] not in mapper.pytorch_params
+                and inputs_name[1] in mapper.pytorch_params):
+            param = mapper.pytorch_params[inputs_name[1]]
+            dtype = string(str(param.dtype))
+            shape = (1, )
+            mapper._check_input(graph,
+                                inputs_node[2],
+                                inputs_name[2],
+                                current_outputs,
+                                scope_name,
+                                dtype=dtype,
+                                shape=shape)
+        else:
+            mapper._check_input(graph, inputs_node[2], inputs_name[2],
+                                current_outputs, scope_name)
+
         graph.add_layer("paddle.add",
                         inputs={
                             "x": output_name,

x2paddle/op_mapper/pytorch2paddle/pytorch_op_mapper.py

@@ -173,7 +173,14 @@ def _get_outputs_name(self, node, attr_name=None):
             outputs_name.append(output_name)
         return outputs_name
 
-    def _check_input(self, graph, node, output_name, node_outputs, scope_name):
+    def _check_input(self,
+                     graph,
+                     node,
+                     output_name,
+                     node_outputs,
+                     scope_name,
+                     dtype=None,
+                     shape=None):
         if node.kind() == "prim::GetAttr":
             param = self.pytorch_params[output_name]
             if isinstance(param, np.ndarray):
@@ -231,19 +238,32 @@ def _check_input(self, graph, node, output_name, node_outputs, scope_name):
                                     value=string(param) if isinstance(
                                         param, str) else param)
             node_outputs.append(output_name)
-        elif node.kind(
-        ) == "prim::Constant" and output_name in self.pytorch_params:
-            param = self.pytorch_params[output_name]
-            self.paddle_params[output_name] = param
-            layer_id = graph.add_layer(
-                "self.create_parameter",
-                inputs={},
-                outputs=[output_name],
-                scope_name=scope_name,
-                dtype=string(str(param.dtype)),
-                shape=param.shape,
-                default_initializer="paddle.nn.initializer.Constant(value=0.0)")
-            self.output2id[output_name] = layer_id
+        elif node.kind() == "prim::Constant":
+            if output_name in self.pytorch_params:
+                param = self.pytorch_params[output_name]
+                self.paddle_params[output_name] = param
+                layer_id = graph.add_layer(
+                    "self.create_parameter",
+                    inputs={},
+                    outputs=[output_name],
+                    scope_name=scope_name,
+                    dtype=string(str(param.dtype)),
+                    shape=param.shape,
+                    default_initializer=
+                    "paddle.nn.initializer.Constant(value=0.0)")
+                self.output2id[output_name] = layer_id
+            else:
+                if dtype is not None and shape is not None:
+                    layer_id = graph.add_layer(
+                        "self.create_parameter",
+                        inputs={},
+                        outputs=[output_name],
+                        scope_name=scope_name,
+                        dtype=dtype,
+                        shape=shape,
+                        default_initializer=
+                        "paddle.nn.initializer.Constant(value=0.0)")
+                    self.output2id[output_name] = layer_id
 
     def _get_inputs_name(self, node):
         inputs_name = []