add tf static

x2paddle/convert.py

@@ -132,18 +132,9 @@ def tf2paddle(model_path,
         graph_opt = GraphOptimizer(source_frame="tf", paddle_type=paddle_type)
         graph_opt.optimize(mapper.paddle_graph)
     else:
-        from x2paddle.optimizer.tensorflow.bias import BiasOpt
-        from x2paddle.optimizer.tensorflow.transpose import TransposeOpt
-        from x2paddle.optimizer.tensorflow.batch_norm import BatchNormOpt
-        from x2paddle.optimizer.tensorflow.prelu import PReLUOpt
-        bias_opt = BiasOpt()
-        transpose_opt = TransposeOpt()
-        batch_norm_opt = BatchNormOpt()
-        prelu_opt = PReLUOpt()
-        bias_opt.run(mapper.paddle_graph)
-        batch_norm_opt.run(mapper.paddle_graph)
-        prelu_opt.run(mapper.paddle_graph)
-        transpose_opt.run(mapper.paddle_graph)
+        from x2paddle.optimizer.optimizer import GraphOptimizer
+        graph_opt = GraphOptimizer(source_frame="tf", paddle_type=paddle_type)
+        graph_opt.optimize(mapper.paddle_graph)
     mapper.paddle_graph.gen_model(save_dir)
         
 

x2paddle/op_mapper/dygraph/tf2paddle/tf_op_mapper.py

@@ -284,7 +284,6 @@ class TFOpMapper(OpMapper):
             inputs["shape"] = dims.name
         layer_attrs["dtype"] = string(input_value.dtype)
         layer_attrs["fill_value"] = input_value.value
-        
 
         self.paddle_graph.add_layer(
             "paddle.full",
@@ -578,6 +577,9 @@ class TFOpMapper(OpMapper):
                 inputs={"x": node.name},
                 outputs=[node.name],
                 perm=[0, 2, 3, 1])
+            
+    def FusedBatchNormV3(self, node):
+        self.FusedBatchNorm(node)
 
     def Mean(self, node):
         input = self.graph.get_input_node(node, 0)
@@ -930,6 +932,23 @@ class TFOpMapper(OpMapper):
             outputs=[node.name],
             axis=axis)
         
+    def Concat(self, node):
+        inputs_list = list()
+        for i in range(1, len(node.inputs)):
+            inputs_list.append(self.graph.get_input_node(node, i))
+        axis = self.graph.get_input_node(node, 0)
+        assert axis.layer_type == "Const", "axis for ConcatV2 must be type Const"
+        axis = axis.value
+        if axis < 0:
+            axis += len(inputs_list[0].out_shapes[0])
+            
+        input_names = [i.name for i in inputs_list]
+        self.paddle_graph.add_layer(
+            kernel="paddle.concat",
+            inputs={"x": input_names},
+            outputs=[node.name],
+            axis=axis)
+        
     def AddN(self, node):
         inputs_list = list()
         for i in range(len(node.inputs) - 1):
@@ -1400,6 +1419,7 @@ class TFOpMapper(OpMapper):
         inputs = {"x": x.name, "y": y.name}
         x_shape = x.out_shapes[0]
         y_shape = y.out_shapes[0]
+        # TODO(syf)
         layer_id = self.paddle_graph.add_layer(
             "fluid.layers.elementwise_sub", inputs=inputs, outputs=[node.name])
         self.paddle_graph.layers[layer_id].input_shapes = {"x": x_shape, "y": y_shape}

x2paddle/op_mapper/static/caffe2paddle/caffe_op_mapper.py

@@ -89,6 +89,7 @@ class CaffeOpMapper(OpMapper):
             input_shape.append(last_node.output_shape[idx])
 
         node.input_shape = input_shape
+
         func_name = 'shape_' + node.layer_type.lower()
         if is_fluid_op:
             node.output_shape = getattr(caffe_shape, func_name)(node.layer,
@@ -974,4 +975,4 @@ class CaffeOpMapper(OpMapper):
             kernel=op_info,
             inputs={"x": self.get_input_name(input)},
             outputs=[node.layer_name])
-        
+        
\ No newline at end of file

x2paddle/optimizer/code_optimizer/hierachical_tree.py

@@ -201,6 +201,7 @@ class HierarchicalTree(Tree):
         code_str = gen_layer_code(self.pd_graph, sub_layers, module_name, 
                                                      different_attrs=diff_attrs_column)
         
+#         print(code_str)
         self.codes.append(code_str)
         for sub_layers in sub_layers_list:
             inputs, outputs = get_inputs_outputs(self.pd_graph, sub_layers)
@@ -358,7 +359,7 @@ class HierarchicalTree(Tree):
                 run_func_list.append("    # {}: 形状为{}，类型为{}。".format(k, v[0], v[1]))
             run_func_list.extend(
                 ["    paddle.disable_static()",
-                 "    params = paddle.load('{}/model.pdparams')".format(osp.abspath(save_dir)),
+                 "    params, _ = fluid.load_dygraph('{}/model')".format(save_dir),
                  "    model = {}()".format(self.pd_graph.name),
                  "    model.set_dict(params)",
                  "    model.eval()",
@@ -370,12 +371,7 @@ class HierarchicalTree(Tree):
         self.update_parameters()
         import_list = ["import paddle",
                        "import paddle.fluid as fluid",
-                       "from paddle.fluid.initializer import Constant",
-                       "from paddle.fluid.param_attr import ParamAttr",
-                       "import math",
-                       "from x2paddle.op_mapper.dygraph.pytorch2paddle " + \
-                                 "import pytorch_custom_layer as x2paddle_nn"
-                       "\n",]
+                       "",]
         import_str = "\n".join(import_list)
         if not osp.exists(save_dir):
             os.makedirs(save_dir)

x2paddle/optimizer/code_optimizer/layer_code_generator.py

@@ -29,9 +29,9 @@ NN_KERNEL_NAME = {"paddle.nn.BatchNorm": "bn",
                   "paddle.nn.Tanh": "tanh",
                   "paddle.nn.AvgPool2D": "pool",
                   "paddle.nn.MaxPool2D": "pool",
-                  "paddle.nn.Pad1D": "pad",
-                  "paddle.nn.Pad2D": "pad",
-                  "paddle.nn.Pad3D": "pad",
+                  "paddle.nn.Pad1d": "pad",
+                  "paddle.nn.Pad2d": "pad",
+                  "paddle.nn.Pad3d": "pad",
                   "paddle.nn.Dropout": "dropout",
                   "paddle.nn.GELU": "gelu",
                   "paddle.nn.Hardtanh": "tanh",
@@ -175,11 +175,9 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()):
                 if layer.kernel.startswith("paddle.nn") and index == 0:
                     continue
                 if not output_name.startswith("x") or output_name in outputs \
-                        or layer.kernel == "prim.assert":
+                        or layer.kernel == "prim.assert" or \
+                        layer.kernel == "prim.if" or layer.kernel == "prim.loop":
                     continue
-                elif layer.kernel == "prim.if" or layer.kernel == "prim.loop":
-                    if index != 0:
-                        outputs.append(output_name)
                 elif output_name not in outputs:
                     outputs.append(output_name)
             continue
@@ -189,22 +187,15 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()):
                     if layer.kernel.startswith("paddle.nn") and index == 0 and "functional" not in layer.kernel:
                         continue
                     if not output_name.startswith("x") or output_name in outputs \
-                            or layer.kernel == "prim.assert":
+                            or layer.kernel == "prim.assert" or \
+                            layer.kernel == "prim.if" or layer.kernel == "prim.loop":
                         continue
-                    elif layer.kernel == "prim.if" or layer.kernel == "prim.loop":
-                        if index != 0:
-                            outputs.append(output_name)
                     else:
                         outputs.append(output_name)
     no_output_count = 0
     for i, (layer_id, layer) in enumerate(sub_layers.items()):
-        if ("paddle.nn" in layer.kernel and "functional" not in layer.kernel) or \
-                layer.kernel.startswith("custom_layer"):
-            line = "self.{}".format(layer.outputs[0])
-            if layer.kernel.startswith("custom_layer"):
-                line += "= x2paddle_nn.{}(".format(layer.kernel.split(":")[-1])
-            else:
-                line += " = {}(".format(layer.kernel)
+        if ("paddle.nn" in layer.kernel and "functional" not in layer.kernel):
+            line = "self.{} = {}(".format(layer.outputs[0], layer.kernel)
             for k, v in layer.attrs.items():
                 key_name = "{}_{}".format(layer.outputs[0], k)
                 if key_name in different_attrs:
@@ -298,10 +289,7 @@ def gen_layer_code(graph, sub_layers, sub_layers_name, different_attrs=list()):
                 else:
                     if v not in cur_outputs and v not in inputs:
                         inputs.append(v)
-                    if k == "args":
-                        line += v
-                    else:
-                        line += "{}={}, ".format(k, v)
+                    line += "{}={}, ".format(k, v)
             for k, v in layer.attrs.items():
                 key_name = "{}_{}".format(layer.outputs[0], k)
                 if key_name in different_attrs:

x2paddle/optimizer/code_optimizer/subgraphs_union.py

@@ -50,25 +50,21 @@ def get_inputs_outputs(pd_graph, layers):
     for layer_id, layer in layers.items():
         # 获取输出节点名字
         if layer_id not in pd_graph.edges_out:
-            for index, output_name in enumerate(layer.outputs):
+            for output_name in layer.outputs:
                 if not output_name.startswith("x") or output_name in outputs \
-                        or layer.kernel == "prim.assert":
+                        or layer.kernel == "prim.assert" or \
+                        layer.kernel == "prim.if" or layer.kernel == "prim.loop":
                     continue
-                elif layer.kernel == "prim.if" or layer.kernel == "prim.loop":
-                        if index != 0:
-                            outputs.append(output_name)
                 elif output_name not in outputs:
                     outputs.append(output_name)
         else:
             for out_layer_id in pd_graph.edges_out[layer_id]:
                 if out_layer_id not in layer_ids:
-                    for index, output_name in enumerate(layer.outputs):
+                    for output_name in layer.outputs:
                         if not output_name.startswith("x") or output_name in outputs \
-                                or layer.kernel == "prim.assert":
+                                or layer.kernel == "prim.assert" or \
+                                layer.kernel == "prim.if" or layer.kernel == "prim.loop":
                             continue
-                        elif layer.kernel == "prim.if" or layer.kernel == "prim.loop":
-                            if index != 0:
-                                outputs.append(output_name)
                         else:
                             outputs.append(output_name)
         # 获取输入节点名字

x2paddle/optimizer/elimination/static/__init__.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .transpose_elimination import StaticTransposeElimination
+from .transpose_eliminate_pass import StaticTransposeEliminatePass
\ No newline at end of file

x2paddle/optimizer/elimination/static/transpose_eliminate_pass.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.elimination.static import StaticTransposeElimination
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class StaticTransposeEliminatePass(Pass):
+    name = "static_transpose_eliminate_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = StaticTransposeElimination()
+        fuser.operate(graph)
+
+
+# 用于注册
+static_transpose_eliminate_pass = StaticTransposeEliminatePass()
\ No newline at end of file

x2paddle/optimizer/elimination/static/transpose_elimination.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import sys
+import numpy as np
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class StaticTransposeElimination(FuseBase):
+    def __init__(self):
+        super(StaticTransposeElimination, self).__init__(graph_type="static")
+        self.direct_layers = [
+            'paddle.nn.functional.relu', 'paddle.nn.functional.relu6', 'paddle.abs',
+            'paddle.nn.functional.sigmoid', 'paddle.exp', 'paddle.rsqrt',
+            'paddle.nn.functional.swish', 'paddle.tanh',
+            'paddle.nn.functional.softplus', 'paddle.nn.functional.leaky_relu',
+            'paddle.floor', 'paddle.erf', 'paddle.square'
+        ]
+        self.elementwise_layers = [
+            'paddle.add', 'fluid.layers.elementwise_sub',
+            'paddle.multiply', 'paddle.divide'
+        ]
+        self.reduce_layers = [
+            'paddle.mean', 'paddle.all',
+            'paddle.max', 'paddle.any',
+            'paddle.sum', 'paddle.prod'
+        ]
+
+    def get_transpose_num(self, graph):
+        count = 0
+        for layer_id, layer in graph.layers.items():
+            if layer.kernel == "paddle.transpose":
+                count += 1
+        return count
+    
+    def operate(self, graph):
+        total_layer_num = len(graph.layers)
+        scanned_layers = set()
+        optimized_transpose_layers = list()
+        optimized_reduce_layers = list()
+        optimized_concat_layers = list()
+        optimized_elementwise_layers = list()
+        
+        def get_index(layer):
+            if layer.kernel.startswith("paddle.nn") and "functional" not in layer.kernel:
+                return 1
+            else:
+                return 0 
+
+        def strip_transpose(_graph):
+            layers = copy.deepcopy(_graph.layers)
+            for layer_id, layer in layers.items():
+                if layer_id in scanned_layers:
+                    continue
+                scanned_layers.add(layer_id)
+                percent = round(len(scanned_layers) / total_layer_num * 100, 2)
+                sys.stderr.write("\rOptimize Transpose Layers...{}%".format(
+                    percent))
+
+                if layer.kernel != "paddle.transpose":
+                    continue
+                if layer.attrs["perm"] != [0, 2, 3, 1]:
+                    continue
+                transpose_layers = list()
+                propagate_layers = list()
+                reduce_layers = list()
+                concat_layers = list()
+                # 此elementwise_layers专用于存储shape(4) + shape(1)的形式layer
+                elementwise_layers = list()
+                can_be_optimized = True
+                for out in _graph.edges_out.get(layer_id, []):
+                    if _graph.layers[out].kernel == "paddle.transpose":
+                        if _graph.layers[out].attrs["perm"] != [0, 3, 1, 2]:
+                            can_be_optimized = False
+                            break
+                        transpose_layers.append(out)
+                    elif _graph.layers[out].kernel in self.elementwise_layers:
+                        propagate_layers.append(out)
+                    elif _graph.layers[out].kernel in self.direct_layers:
+                        ouput_index = get_index(_graph.layers[out])
+                        if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
+                            can_be_optimized = False
+                            break
+                        propagate_layers.append(out)
+                    elif _graph.layers[out].kernel in self.reduce_layers:
+                        ouput_index = get_index(_graph.layers[out])
+                        if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
+                            can_be_optimized = False
+                            break
+                        if not _graph.layers[out].attrs.get('keepdim', False):
+                            can_be_optimized = False
+                            break
+                        propagate_layers.append(out)
+                        reduce_layers.append(out)
+                    elif _graph.layers[out].kernel == "paddle.concat":
+                        ouput_index = get_index(_graph.layers[out])
+                        if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
+                            can_be_optimized = False
+                            break
+                        propagate_layers.append(out)
+                        concat_layers.append(out)
+                    else:
+                        can_be_optimized = False
+                        break
+
+                visited_layers = set()
+                while len(propagate_layers) > 0 and can_be_optimized:
+                    current_id = propagate_layers.pop(0)
+                    visited_layers.add(current_id)
+                    for out in _graph.edges_out.get(current_id, []):
+                        if _graph.layers[
+                                out].kernel == "paddle.transpose":
+                            if _graph.layers[out].attrs["perm"] != [0, 3, 1, 2]:
+                                can_be_optimized = False
+                                break
+                            transpose_layers.append(out)
+                        elif _graph.layers[
+                                out].kernel in self.elementwise_layers:
+                            output_index = get_index(_graph.layers[out])
+                            if _graph.layers[out].outputs[output_index] in _graph.outputs:
+                                can_be_optimized = False
+                                break
+                            if out not in visited_layers:
+                                propagate_layers.append(out)
+                        elif _graph.layers[out].kernel in self.direct_layers:
+                            output_index = get_index(_graph.layers[out])
+                            if _graph.layers[out].outputs[output_index] in _graph.outputs:
+                                can_be_optimized = False
+                                break
+                            if out not in visited_layers:
+                                propagate_layers.append(out)
+                        elif _graph.layers[out].kernel in self.reduce_layers:
+                            output_index = get_index(_graph.layers[out])
+                            if _graph.layers[out].outputs[output_index] in _graph.outputs:
+                                can_be_optimized = False
+                                break
+                            if not _graph.layers[out].attrs.get('keepdim',
+                                                                False):
+                                can_be_optimized = False
+                                break
+                            if out not in visited_layers:
+                                propagate_layers.append(out)
+                                reduce_layers.append(out)
+                        elif _graph.layers[out].kernel == "paddle.concat":
+                            output_index = get_index(_graph.layers[out])
+                            if _graph.layers[out].outputs[output_index] in _graph.outputs:
+                                can_be_optimized = False
+                                break
+                            if out not in visited_layers:
+                                propagate_layers.append(out)
+                                concat_layers.append(out)
+                        else:
+                            can_be_optimized = False
+                            break
+                    for ipt in _graph.edges_in.get(current_id, []):
+                        if _graph.layers[
+                                current_id].kernel in self.elementwise_layers:
+                            try:
+                                x_shape = _graph.layers[
+                                    current_id].input_shapes['x']
+                                y_shape = _graph.layers[
+                                    current_id].input_shapes['y']
+                                output_index = get_index(_graph.layers[ipt])
+                                if _graph.layers[ipt].outputs[
+                                        output_index] == _graph.layers[current_id].inputs[
+                                            'x']:
+                                    if len(x_shape) <= 1:
+                                        elementwise_layers.append(current_id)
+                                        continue
+                                elif _graph.layers[ipt].outputs[
+                                        output_index] == _graph.layers[current_id].inputs[
+                                            'y']:
+                                    if len(y_shape) <= 1:
+                                        elementwise_layers.append(current_id)
+                                        continue
+                                else:
+                                    raise Exception(
+                                        "Unexcepted situation happend while optimizing transpose"
+                                    )
+                            except Exception as e:
+                                can_be_optimized = False
+                                break
+                        output_index = get_index(_graph.layers[ipt])
+                        if _graph.layers[
+                                ipt].kernel == "paddle.transpose":
+                            if _graph.layers[ipt].attrs["perm"] != [0, 2, 3, 1]:
+                                can_be_optimized = False
+                                break
+                            if ipt not in visited_layers:
+                                transpose_layers.append(ipt)
+                        elif _graph.layers[
+                                ipt].kernel in self.elementwise_layers:
+                            if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
+                                can_be_optimized = False
+                                break
+                            if ipt not in visited_layers:
+                                propagate_layers.append(ipt)
+                        elif _graph.layers[ipt].kernel in self.direct_layers:
+                            if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
+                                can_be_optimized = False
+                                break
+                            if ipt not in visited_layers:
+                                propagate_layers.append(ipt)
+                        elif _graph.layers[ipt].kernel in self.reduce_layers:
+                            if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
+                                can_be_optimized = False
+                                break
+                            if not _graph.layers[ipt].attrs.get('keepdim',
+                                                                False):
+                                can_be_optimized = False
+                                break
+                            if ipt not in visited_layers:
+                                propagate_layers.append(ipt)
+                                reduce_layers.append(ipt)
+                        elif _graph.layers[ipt].kernel == "paddle.concat":
+                            if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
+                                can_be_optimized = False
+                                break
+                            if ipt not in visited_layers:
+                                propagate_layers.append(ipt)
+                                concat_layers.append(ipt)
+                        else:
+                            can_be_optimized = False
+                            break
+                    if not can_be_optimized:
+                        break
+                if not can_be_optimized:
+                    continue
+
+                transpose_layers.append(layer_id)
+                transpose_layers = list(set(transpose_layers))
+                for l in transpose_layers:
+                    output_index = get_index(graph.layers[l])
+                    if graph.layers[l].outputs[output_index] in graph.outputs:
+                        can_be_optimized = False
+                        break
+                if not can_be_optimized:
+                    continue
+
+                for l in transpose_layers:
+                    _graph.del_layer(l)
+
+                optimized_transpose_layers.extend(transpose_layers)
+                optimized_reduce_layers.extend(reduce_layers)
+                optimized_concat_layers.extend(concat_layers)
+                optimized_elementwise_layers.extend(elementwise_layers)
+                return True
+            return False
+
+        before_transpose_num = self.get_transpose_num(graph)
+        opt_graph = copy.deepcopy(graph)
+        total_layer_num = len(opt_graph.layers)
+
+        while strip_transpose(opt_graph):
+            pass
+
+        for layer_id in list(set(optimized_transpose_layers)):
+            graph.del_layer(layer_id)
+        for layer_id in list(set(optimized_reduce_layers)):
+            dim = graph.layers[layer_id].attrs.get('axis', None)
+            if dim is not None:
+                for i in range(len(dim)):
+                    dim[i] = [0, 2, 3, 1][dim[i]]
+                graph.layers[layer_id].attrs['axis'] = dim
+        for layer_id in list(set(optimized_concat_layers)):
+            axis = graph.layers[layer_id].attrs.get('axis', 0)
+            graph.layers[layer_id].attrs['axis'] = [0, 2, 3, 1][axis]
+        for layer_id in list(set(optimized_elementwise_layers)):
+            axis = graph.layers[layer_id].attrs.get('axis', -1)
+            graph.layers[layer_id].attrs['axis'] = [0, 2, 3, 1][axis]
+            if graph.layers[layer_id].kernel == "paddle.add":
+                graph.layers[layer_id].kernel = "fluid.layers.elementwise_add"
+
+        current_transpose_num = self.get_transpose_num(graph)
+        print(
+            "\nTranspose layers optimized, before: transpose_num={}, after: transpose_num={}".
+            format(before_transpose_num, current_transpose_num))

x2paddle/optimizer/fusion/dygraph/conv2d_add_fuser.py

@@ -105,10 +105,6 @@ class DygraphConv2DAddFuser(FuseBase):
             if layer.kernel == "paddle.nn.Conv2D":
                 conv_id = layer_id
         for layer_id, layer in matches.items():
-            if layer.kernel == "paddle.nn.functional.conv2d_transpose":
-                layer.bias = bias_name
-                if not is_transpose:
-                    layer.outputs[0] = output_name
             if layer.kernel == "paddle.nn.Conv2D":
                 layer.attrs["bias_attr"] = bias_name
                 if not is_transpose:

x2paddle/optimizer/fusion/static/__init__.py

@@ -13,4 +13,11 @@
 # limitations under the License.
 
 from .bn_scale_fuser import Static_BNScaleFuser
-from .bn_scale_fuse_pass import Static_BNScaleFusePass
\ No newline at end of file
+from .bn_scale_fuse_pass import Static_BNScaleFusePass
+from .conv2d_add_fuser import StaticConv2DAddFuser
+from .conv2d_add_fuse_pass import StaticConv2DAddFusePass
+from .prelu_fuser import StaticPReLUFuser
+from .prelu_fuse_pass import StaticPReLUFusePass
+from .tf_batchnorm_fuser import StaticTFBatchNormFuser
+from .tf_batchnorm_fuse_pass import StaticTFBatchNormFusePass
+

x2paddle/optimizer/fusion/static/bn_scale_fuser.py

@@ -79,7 +79,6 @@ class Static_BNScaleFuser(FuseBase):
         graph.layers[new_layer_id] = new_layer
         matches.pop(new_layer_id)
 
-
     def gen_new_layer(self, parameters, matches):
         layers_id = list(matches.keys())
         layer = matches[layers_id[0]]

x2paddle/optimizer/fusion/static/conv2d_add_fuse_pass.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion.static import StaticConv2DAddFuser
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class StaticConv2DAddFusePass(Pass):
+    name = "static_conv2d_add_fuse_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = StaticConv2DAddFuser()
+        fuser.operate(graph, match_kind="edge")
+
+
+# 用于注册
+static_conv2d_add_fuse_pass = StaticConv2DAddFusePass()
\ No newline at end of file

x2paddle/optimizer/fusion/static/conv2d_add_fuser.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import numpy as np
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class StaticConv2DAddFuser(FuseBase):
+    def __init__(self):
+        super(StaticConv2DAddFuser, self).__init__(graph_type="static")
+        self.patterns = list()
+
+    def build_pattern(self):
+        """ 描述需要替换的conv2d+add图结构。
+        conv2d+add层模式python实现代码示例:
+        模式一：
+        MobilenetV1_Logits_Conv2d_1c_1x1_biases = paddle.static.create_parameter(dtype='float32', shape=[1001], name='MobilenetV1_Logits_Conv2d_1c_1x1_biases', default_initializer=paddle.nn.initializer.Constant(value=0.0))
+        conv2d_transpose_14 = paddle.transpose(x=MobilenetV1_Logits_AvgPool_1a_AvgPool, perm=[0, 3, 1, 2])
+        MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.nn.functional.conv2d(x=conv2d_transpose_14, weight=MobilenetV1_Logits_Conv2d_1c_1x1_weights, bias=None, stride=[1, 1], dilation=[1, 1], padding='SAME')
+        MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.transpose(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, perm=[0, 2, 3, 1])
+        MobilenetV1_Logits_Conv2d_1c_1x1_BiasAdd = paddle.add(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, y=MobilenetV1_Logits_Conv2d_1c_1x1_biases)
+        
+        模式二：
+        MobilenetV1_Logits_Conv2d_1c_1x1_biases = paddle.static.create_parameter(dtype='float32', shape=[1001], name='MobilenetV1_Logits_Conv2d_1c_1x1_biases', default_initializer=paddle.nn.initializer.Constant(value=0.0))
+        MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.nn.functional.conv2d(x=conv2d_transpose_14, weight=MobilenetV1_Logits_Conv2d_1c_1x1_weights, bias=None, stride=[1, 1], dilation=[1, 1], padding='SAME')
+        MobilenetV1_Logits_Conv2d_1c_1x1_BiasAdd = paddle.add(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, y=MobilenetV1_Logits_Conv2d_1c_1x1_biases)
+        """
+
+        def gen_name(id):
+            return "x" + str(id)
+
+        pattern = PaddleGraph(graph_type="dygraph")
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        pattern.add_layer(
+            kernel="paddle.transpose",
+            inputs={"x": "conv-input-0"},
+            outputs=[gen_name(1)],
+            perm=[0, 3, 1, 2])
+        pattern.add_layer(
+            kernel="paddle.nn.functional.conv2d",
+            inputs={"input": gen_name(1),
+                    "weight": "conv-input-1"},
+            outputs=[gen_name(2)])
+        pattern.add_layer(
+            kernel="paddle.transpose",
+            inputs={"x": gen_name(2)},
+            outputs=[gen_name(2)],
+            perm=[0, 2, 3, 1])
+        pattern.add_layer(
+            kernel="paddle.add",
+            inputs={"x": gen_name(2),
+                    "y": gen_name(0)},
+            outputs=[gen_name(3)])
+        pattern.build(inputs={"input-0": "conv-input-0", 
+                              "input-1": "conv-input-1"})
+        self.patterns.append(pattern)
+        
+        pattern = PaddleGraph(graph_type="dygraph")
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        pattern.add_layer(
+            kernel="paddle.nn.functional.conv2d",
+            inputs={"input": "conv-input-0",
+                    "weight": "conv-input-1"},
+            outputs=[gen_name(1)])
+        pattern.add_layer(
+            kernel="paddle.add",
+            inputs={"x": gen_name(1),
+                    "y": gen_name(0)},
+            outputs=[gen_name(2)])
+        pattern.build(inputs={"input-0": "conv-input-0", 
+                              "input-1": "conv-input-1"})
+        self.patterns.append(pattern)
+
+    def insert_new_layer(self, graph, parameters, matches):
+        self.gen_new_layer(matches, graph)
+        matches_copy = copy.deepcopy(matches)
+        for layer_id, layer in matches_copy.items():
+            if layer.kernel not in ["paddle.add"]:
+                matches.pop(layer_id)
+
+    def gen_new_layer(self, matches, graph):
+        is_transpose = False
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.static.create_parameter":
+                bias_name = layer.attrs["name"][1: -1]
+            if layer.kernel == "paddle.transpose":
+                is_transpose = True
+            if layer.kernel == "paddle.add":
+                output_name = layer.outputs[0]
+            if layer.kernel == "paddle.nn.functional.conv2d":
+                conv_id = layer_id
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.nn.functional.conv2d":
+                layer.inputs["bias"] = bias_name
+                layer.attrs.pop("bias")
+                if not is_transpose:
+                    layer.outputs[0] = output_name
+            if layer.kernel == "paddle.transpose":
+                if conv_id in graph.edges_in[layer_id]:
+                    layer.outputs[0] = output_name
+            

x2paddle/optimizer/fusion/static/prelu_fuse_pass.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion.static import StaticPReLUFuser
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class StaticPReLUFusePass(Pass):
+    name = "static_prelu_fuse_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = StaticPReLUFuser()
+        fuser.operate(graph, match_kind="edge")
+
+
+# 用于注册
+static_prelu_fuse_pass = StaticPReLUFusePass()
\ No newline at end of file

x2paddle/optimizer/fusion/static/prelu_fuser.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import numpy as np
+from collections import OrderedDict
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class StaticPReLUFuser(FuseBase):
+    def __init__(self):
+        super(StaticPReLUFuser, self).__init__(graph_type="static")
+
+    def build_pattern(self):
+        """ 描述需要替换的prelu图结构。
+        prelu层模式python实现代码示例:
+            conv4_alphas = paddle.static.create_parameter(dtype='float32', shape=[128], name='conv4_alphas', default_initializer=paddle.nn.initializer.Constant(value=0.0))
+            conv4_mul_1_y = paddle.full(dtype='float32', shape=[1], fill_value=0.5)
+            conv4_Relu = paddle.nn.functional.relu(x=conv4_BiasAdd)
+            conv4_Abs = paddle.abs(x=conv4_BiasAdd)
+            conv4_sub = fluid.layers.elementwise_sub(x=conv4_BiasAdd, y=conv4_Abs)
+            conv4_mul = paddle.multiply(x=conv4_alphas, y=conv4_sub)
+            conv4_mul_1 = paddle.multiply(x=conv4_mul, y=conv4_mul_1_y)
+            conv4_add = paddle.add(x=conv4_Relu, y=conv4_mul_1)
+        """
+
+        def gen_name(id):
+            return "x" + str(id)
+
+        self.pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        self.pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(1)],
+            shape=[1],
+            fill_value=0.5)
+        self.pattern.add_layer(
+            "paddle.nn.functional.relu",
+            inputs={"x": "prelu-input-0"},
+            outputs=[gen_name(2)])
+        self.pattern.add_layer(
+            "paddle.abs",
+            inputs={"x": "prelu-input-0"},
+            outputs=[gen_name(3)])
+        self.pattern.add_layer(
+            "fluid.layers.elementwise_sub",
+            inputs={"x": "prelu-input-0",
+                    "y": gen_name(3)},
+            outputs=[gen_name(4)])
+        self.pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(0),
+                    "y": gen_name(4)},
+            outputs=[gen_name(5)])
+        self.pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(5),
+                    "y": gen_name(1)},
+            outputs=[gen_name(6)])
+        self.pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(2),
+                    "y": gen_name(6)},
+            outputs=[gen_name(7)])
+        self.pattern.build(inputs={"input-0": "prelu-input-0", })
+        
+
+    def insert_new_layer(self, graph, parameters, matches):
+        new_layers, last_layer_id = self.gen_new_layer(matches, parameters, graph)
+        matches_copy = copy.deepcopy(matches)
+        for layer_id, layer in matches_copy.items():
+            for i in range(4):
+                if layer_id == new_layers[i].id:
+                    matches.pop(new_layers[i].id)
+        prefix_layers = OrderedDict()
+        mid_layers = OrderedDict()
+        suffix_layers = OrderedDict()
+        is_need_id = False
+        for layer_id, layer in graph.layers.items():
+            if is_need_id:
+                suffix_layers[layer_id] = layer
+            else:
+                if layer_id == last_layer_id:
+                    for i in range(4):
+                        mid_layers[new_layers[i].id] = new_layers[i]
+                    is_need_id = True
+                prefix_layers[layer_id] = layer
+        prefix_layers.update(mid_layers)
+        prefix_layers.update(suffix_layers)
+        graph.layers = prefix_layers
+
+    def gen_new_layer(self, matches, parameters, graph):
+        layer_id_list = list(matches.keys())
+        layer_id_list.sort(key = int)
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.nn.functional.relu":
+                input_name = layer.inputs["x"]
+            if layer.kernel == "paddle.static.create_parameter":
+                param_layer = layer
+                param_name = layer.outputs[0]
+            if layer.kernel == "paddle.add":
+                output_name = layer.outputs[0]
+        transpose0 = PaddleLayer(
+            id=layer_id_list[-1] + "_1",
+            kernel="paddle.transpose",
+            inputs={"x": input_name},
+            outputs=["{}_transpose_for_prelu".format(input_name)],
+            perm=[0, 3, 1, 2])
+        param = parameters[param_name]
+        c = param.shape[0]
+        prelu = PaddleLayer(id=layer_id_list[-1] + "_2",
+                            kernel="paddle.nn.functional.prelu",
+                            inputs={"x": "{}_transpose_for_prelu".format(input_name),
+                                    "weight": param_name},
+                            outputs=["{}_prelu".format(input_name)])
+        transpose1 = PaddleLayer(
+            id=layer_id_list[-1] + "_3",
+            kernel="paddle.transpose",
+            inputs={"x": "{}_prelu".format(input_name)},
+            outputs=[output_name],
+            perm=[0, 2, 3, 1])
+        return [param_layer, transpose0, prelu, transpose1], layer_id_list[-1]
+

x2paddle/optimizer/fusion/static/tf_batchnorm_fuse_pass.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion.static import StaticTFBatchNormFuser
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class StaticTFBatchNormFusePass(Pass):
+    name = "static_tf_batchnorm_fuse_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = StaticTFBatchNormFuser()
+        fuser.operate(graph, match_kind="edge")
+
+
+# 用于注册
+static_tf_batchnorm_fuse_pass = StaticTFBatchNormFusePass()
\ No newline at end of file

x2paddle/optimizer/fusion/static/tf_batchnorm_fuser.py

+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import numpy as np
+from collections import OrderedDict
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class StaticTFBatchNormFuser(FuseBase):
+    def __init__(self):
+        super(StaticTFBatchNormFuser, self).__init__(graph_type="static")
+        self.patterns = list()
+
+    def build_pattern(self):
+        """ 描述需要替换的batchnorm图结构。
+        batchnorm层模式python实现代码示例:
+            
+        """
+
+        def gen_name(id):
+            return "x" + str(id)
+
+        pattern = PaddleGraph(graph_type="dygraph")
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(1)],
+            shape=[1])
+        pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(0), "y": gen_name(1)},
+            outputs=[gen_name(2)])
+        pattern.add_layer(
+            "paddle.rsqrt",
+            inputs={"x": gen_name(2)},
+            outputs=[gen_name(3)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(4)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(3), "y": gen_name(4)},
+            outputs=[gen_name(5)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(6)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(6), "y": gen_name(5)},
+            outputs=[gen_name(7)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(8)])
+        pattern.add_layer(
+            "fluid.layers.elementwise_sub",
+            inputs={"x": gen_name(8), "y": gen_name(7)},
+            outputs=[gen_name(9)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": "bn-input-0", "y": gen_name(5)},
+            outputs=[gen_name(10)])
+        pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(10), "y": gen_name(9)},
+            outputs=[gen_name(11)])
+        pattern.build(inputs={"input-0": "bn-input-0", })
+        self.patterns.append(pattern)
+        
+        pattern = PaddleGraph(graph_type="dygraph")
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(1)],
+            shape=[1])
+        pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(0), "y": gen_name(1)},
+            outputs=[gen_name(2)])
+        pattern.add_layer(
+            "paddle.rsqrt",
+            inputs={"x": gen_name(2)},
+            outputs=[gen_name(3)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(4)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(3), "y": gen_name(4)},
+            outputs=[gen_name(5)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": "bn-input-0", "y": gen_name(5)},
+            outputs=[gen_name(10)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(6)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(6), "y": gen_name(5)},
+            outputs=[gen_name(7)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(8)])
+        pattern.add_layer(
+            "fluid.layers.elementwise_sub",
+            inputs={"x": gen_name(8), "y": gen_name(7)},
+            outputs=[gen_name(9)])
+        pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(10), "y": gen_name(9)},
+            outputs=[gen_name(11)])
+        pattern.build(inputs={"input-0": "bn-input-0", })
+        self.patterns.append(pattern)
+
+    def insert_new_layer(self, graph, parameters, matches):
+        new_layers, last_layer_id = self.gen_new_layer(matches, parameters, graph)
+        matches_copy = copy.deepcopy(matches)
+        for layer_id, layer in matches_copy.items():
+            for i in range(7):
+                if layer_id == new_layers[i].id:
+                    matches.pop(new_layers[i].id)
+        prefix_layers = OrderedDict()
+        mid_layers = OrderedDict()
+        suffix_layers = OrderedDict()
+        is_need_id = False
+        for layer_id, layer in graph.layers.items():
+            if is_need_id:
+                suffix_layers[layer_id] = layer
+            else:
+                if layer_id == last_layer_id:
+                    for i in range(7):
+                        mid_layers[new_layers[i].id] = new_layers[i]
+                    is_need_id = True
+                prefix_layers[layer_id] = layer
+        prefix_layers.update(mid_layers)
+        prefix_layers.update(suffix_layers)
+        graph.layers = prefix_layers
+
+    def gen_new_layer(self, matches, parameters, graph):
+        layer_id_list = list(matches.keys())
+        layer_id_list.sort(key = int)
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.full":
+                full_layer = layer
+                out_layer_id = graph.edges_out[layer_id][0]
+                if matches[out_layer_id].kernel == "paddle.add":
+                    var_layer_id = graph.edges_in[out_layer_id][0]
+                    var_layer = matches[var_layer_id]
+            if layer.kernel == "paddle.rsqrt":
+                out_layer_id = graph.edges_out[layer_id][0]
+                if matches[out_layer_id].kernel == "paddle.multiply":
+                    gamma_layer_id = graph.edges_in[out_layer_id][1]
+                    gamma_layer = matches[gamma_layer_id]
+            if layer.kernel == "fluid.layers.elementwise_sub":
+                in_layer_id = graph.edges_in[layer_id][0]
+                beta_layer = matches[in_layer_id]
+                in_layer_id = graph.edges_in[layer_id][1]
+                in_layer_id = graph.edges_in[in_layer_id][0]
+                mean_layer = matches[in_layer_id]
+                out_layer_id = graph.edges_out[layer_id][0]
+                add_layer = matches[out_layer_id]
+            if layer.kernel == "paddle.multiply":
+                in_layer_id = graph.edges_in[layer_id][1]
+                mul_layer = matches[in_layer_id]
+                if mul_layer.kernel == "paddle.multiply":
+                    in_layer_id = graph.edges_in[layer_id][0]
+                    if in_layer_id not in matches:
+                        input_name = layer.inputs["x"]
+        transpose0 = PaddleLayer(
+            id=layer_id_list[-1] + "_1",
+            kernel="paddle.transpose",
+            inputs={"x": input_name},
+            outputs=["{}_transpose_for_bn".format(input_name)],
+            perm=[0, 3, 1, 2])
+        params = parameters[gamma_layer.outputs[0]]
+        c = params.shape[0]
+        bn = PaddleLayer(
+            id=layer_id_list[-1] + "_2",
+            kernel="paddle.nn.functional.batch_norm",
+            inputs={"x": "{}_transpose_for_bn".format(input_name),
+                    "running_mean": mean_layer.outputs[0],
+                    "running_var": var_layer.outputs[0],
+                    "weight": gamma_layer.outputs[0],
+                    "bias": beta_layer.outputs[0]},
+            outputs=["{}_bn".format(input_name)],
+            epsilon=full_layer.attrs["fill_value"])
+        transpose1 = PaddleLayer(
+            id=layer_id_list[-1] + "_3",
+            kernel="paddle.transpose",
+            inputs={"x": "{}_bn".format(input_name)},
+            outputs=add_layer.outputs,
+            perm=[0, 2, 3, 1])
+        mean_layer.id = layer_id_list[-1] + "_01"
+        var_layer.id = layer_id_list[-1] + "_02"
+        gamma_layer.id = layer_id_list[-1] + "_03"
+        beta_layer.id = layer_id_list[-1] + "_04"
+        return [mean_layer, var_layer, gamma_layer, beta_layer, 
+                transpose0, bn, transpose1], layer_id_list[-1]
+

x2paddle/optimizer/optimizer.py

@@ -16,13 +16,13 @@ from x2paddle.optimizer.pass_manager import PassManager
 from x2paddle.optimizer.fusion.dygraph import *
 from x2paddle.optimizer.fusion.static import *
 from x2paddle.optimizer.elimination.dygraph import *
+from x2paddle.optimizer.elimination.static import *
 
 class GraphOptimizer(object):
     def __init__(self, source_frame, paddle_type="dygraph", jit_type="trace"):
         if source_frame == "pytorch":
             if jit_type == "trace":
-                self.passes = ["dygraph_constant_fuse_pass", 
-                               "trace_fc_fuse_pass"]
+                self.passes = ["trace_fc_fuse_pass"]
             else:
                 self.passes = [
                     "dygraph_constant_fuse_pass", 
@@ -39,12 +39,20 @@ class GraphOptimizer(object):
             else:
                 self.passes = ["static_bn_scale_fuse_pass"]
         elif source_frame == "tf":
-            self.passes = [
-                "dygraph_conv2d_add_fuse_pass",
-                "dygraph_tf_batchnorm_fuse_pass",
-                "dygraph_prelu_fuse_pass",
-                "transpose_eliminate_pass"
-            ]
+            if paddle_type == "dygraph":
+                self.passes = [
+                    "dygraph_conv2d_add_fuse_pass",
+                    "dygraph_tf_batchnorm_fuse_pass",
+                    "dygraph_prelu_fuse_pass",
+                    "transpose_eliminate_pass"
+                ]
+            else:
+                self.passes = [
+                    "static_conv2d_add_fuse_pass",
+                    "static_tf_batchnorm_fuse_pass",
+                    "static_prelu_fuse_pass",
+                    "static_transpose_eliminate_pass"
+                ]
         else:
             self.passes = []