Merge pull request #380 from SunAhong1993/me

add pytorch convertor

x2paddle/__init__.py

 __version__ = "0.7.4"
 
-from .core.program import PaddleProgram
+from .core.program import PaddleGraph
 
-program = PaddleProgram()
+program = PaddleGraph()
 
 name_counter = dict()
 

x2paddle/core/program.py

@@ -18,15 +18,15 @@ import paddle.fluid as fluid
 from paddle.fluid.proto import framework_pb2
 from collections import OrderedDict
 import numpy
-import time
 import collections
 import sys
 import os
 import six
+import pickle
 
 
 class PaddleLayer(object):
-    def __init__(self, kernel, inputs, outputs, **kwargs):
+    def __init__(self, id, kernel, inputs, outputs, **kwargs):
         assert isinstance(
             inputs,
             dict), "parameter 'inputs' for PaddleLayer should be type of dict"
@@ -51,22 +51,29 @@ class PaddleLayer(object):
         self.inputs = inputs
         self.outputs = outputs
         self.attrs = kwargs
-        self.id = str(time.time())
+        self.id = id
+        self.blocks = list()
 
     def add_block(self, block):
-        block.father_layer = self
         self.blocks.append(block)
 
 
-class PaddleProgram(object):
-    def __init__(self):
+class PaddleGraph(object):
+    def __init__(self, parent_layer=None, graph_type="static"):
         self.layers = OrderedDict()
         self.edges_out = dict()
         self.edges_in = dict()
         self.inputs = list()
         self.outputs = list()
         self.parameters = dict()
-        self.father_layer = None
+        self.parent_layer = parent_layer
+        self.graph_type = graph_type
+
+    def set_name(self, name):
+        self.name = name
+
+    def set_parameters(self, parameters):
+        self.parameters = parameters
 
     def clear(self):
         self.layers = OrderedDict()
@@ -76,25 +83,44 @@ class PaddleProgram(object):
         self.outputs = list()
         self.parameters = dict()
 
+    def clear_edges(self):
+        self.edges_out = dict()
+        self.edges_in = dict()
+
     def add_layer(self, kernel, inputs, outputs, **kwargs):
-        layer = PaddleLayer(kernel, inputs, outputs, **kwargs)
         layer_id = str(len(self.layers))
-        if self.father_layer is not None:
-            layer_id = "{}.{}.{}".format(layer_id, len(self.father_layer.blocks()), self.father_layer.id)
+        if self.parent_layer is not None:
+            layer_id = "{}.{}.{}".format(self.parent_layer.id,
+                                         len(self.parent_layer.blocks),
+                                         layer_id)
+        layer = PaddleLayer(layer_id, kernel, inputs, outputs, **kwargs)
         self.layers[layer_id] = layer
         return layer_id
 
-    def build(self):
+    def build(self, inputs=None, outputs=None):
+        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):
                     vs = [vs]
                 for v in vs:
-                    assert v in outputs_from_nodes, "Couldn't find {} in previous layers, the layers should be make by topological sort".format(
+                    assert v in outputs_from_nodes or (
+                        inputs is not None and v in list(inputs.values())
+                    ) or (
+                        outputs is not None and v in outputs
+                    ), "Couldn't find {} in previous layers, the layers should be make by topological sort".format(
                         v)
+                    if v in outputs_from_nodes:
                         in_layer_id = outputs_from_nodes[v]
+                    else:
+                        in_layer_id = -1
                     if in_layer_id not in self.edges_out:
                         self.edges_out[in_layer_id] = list()
                     self.edges_out[in_layer_id].append(layer_id)
@@ -105,6 +131,27 @@ class PaddleProgram(object):
             for output in layer.outputs:
                 outputs_from_nodes[output] = layer_id
 
+            if len(layer.blocks) > 0:
+                for block in layer.blocks:
+                    block.build(layer.inputs, layer.outputs)
+
+        if self.graph_type == "dygraph":
+            self.get_dygraph_inputs()
+            self.get_dygraph_outputs()
+
+    def get_global_layers(self):
+        # 该全局layers的信息是按照拓扑排序组成的
+        def update(layers):
+            global_layers = dict()
+            for layer_id, layer in layers.items():
+                global_layers[layer_id] = layer
+                for block in layer.blocks:
+                    block_global_layers = update(block.layers)
+                    global_layers.update(block_global_layers)
+            return global_layers
+
+        return update(self.layers)
+
     def gen_code(self, code_dir):
         def write_code(f, code_list, indent=0):
             indent_blank = "    " * indent
@@ -162,6 +209,7 @@ class PaddleProgram(object):
         f.close()
 
     def gen_model(self, save_dir):
+        if self.graph_type == "static":
             code_dir = os.path.join(save_dir, 'model_with_code')
             infer_dir = os.path.join(save_dir, 'inference_model')
             self.gen_code(code_dir)
@@ -193,6 +241,9 @@ class PaddleProgram(object):
                         feeded_var_names=[i.name for i in inputs],
                         target_vars=outputs,
                         executor=exe)
+        else:
+            self.gen_dygraph_code(save_dir)
+            self.dump_dygraph_parameter(save_dir)
 
     def dump_parameter(self, param_name, param, save_dir):
         if not os.path.exists(save_dir):
@@ -227,3 +278,167 @@ class PaddleProgram(object):
         fp.write(tensor_desc.SerializeToString())
         param.tofile(fp)
         fp.close()
+
+    def get_dygraph_inputs(self):
+        def update(layers):
+            for layer_id, layer in layers.items():
+                if self.edges_in.get(layer_id, 0) == 0 and self.edges_out.get(
+                        layer_id, 0) == 0:
+                    continue
+                if layer.kernel == "fluid.dygraph.base.to_variable":
+                    value = layer.attrs["value"]
+                    if not value.startswith("params["):
+                        self.inputs.append(value)
+                if len(layer.blocks) > 0:
+                    for block in layer.blocks:
+                        block.get_dygraph_inputs()
+                        self.inputs.extend(block.inputs)
+
+        update(self.layers)
+        self.inputs = list(set(self.inputs))
+
+    def get_dygraph_outputs(self):
+        for layer_id, layer in self.layers.items():
+            if self.edges_in.get(layer_id, 0) == 0 and self.edges_out.get(
+                    layer_id, 0) == 0:
+                continue
+            if self.edges_out.get(layer_id, 0) == 0:
+                for output_name in layer.outputs:
+                    if not output_name.startswith("x"):
+                        continue
+                    self.outputs.append(output_name)
+        self.outputs = list(set(self.outputs))
+
+    def gen_dygraph_code(self, code_dir=None, indent=2):
+        def gen_codes(code_list, indent=0):
+            indent_blank = "    " * indent
+            codes = []
+            for code_line in code_list:
+                if code_line.strip() == "":
+                    codes.append('\n')
+                else:
+                    codes.append(indent_blank + code_line + '\n')
+            return codes
+
+        def gen_head():
+            self.head = gen_codes(
+                [
+                    "from paddle.fluid.initializer import Constant",
+                    "from paddle.fluid.param_attr import ParamAttr",
+                    "import paddle.fluid as fluid",
+                    "",
+                    "class {}(fluid.dygraph.Layer):".format(self.name),
+                ],
+                indent=0)
+            input_data_name = ', '.join(self.inputs)
+            self.init_func.extend(
+                gen_codes(
+                    ["def __init__(self, params):"], indent=1))
+            self.init_func.extend(
+                gen_codes(
+                    ["super({}, self).__init__()".format(self.name)], indent=2))
+            self.forward_func.extend(
+                gen_codes(
+                    ["def forward(self, {}):".format(input_data_name)],
+                    indent=1))
+
+        def write_code(code_dir):
+            f = open(os.path.join(code_dir, 'code.py'), 'w')
+            for code_line in self.head:
+                f.write(code_line)
+            init_writen_codes = []
+            for code_line in self.init_func:
+                if code_line in init_writen_codes:
+                    continue
+                f.write(code_line)
+                init_writen_codes.append(code_line)
+            f.write("\n")
+            return_code = "return {}".format(", ".join(self.outputs))
+            self.forward_func.extend(gen_codes([return_code], indent=2))
+            for code_line in self.forward_func:
+                f.write(code_line)
+            f.close()
+
+        self.init_func = []
+        self.forward_func = []
+        if indent == 2 and code_dir is not None:
+            gen_head()
+
+        for layer_id, layer in self.layers.items():
+            if len(self.layers) > 1:
+                if self.edges_in.get(layer_id, 0) == 0 and self.edges_out.get(
+                        layer_id, 0) == 0 and layer.kernel != "prim.assert" \
+                        and layer.kernel != "prim.exception" \
+                        and layer.kernel != "prim.warnings":
+                    continue
+            if "dygraph" in layer.kernel:
+                line = "{}".format(
+                    layer.outputs[0]
+                ) if layer.kernel == "fluid.dygraph.base.to_variable" and not layer.attrs[
+                    "value"].startswith("params[") else "self.{}".format(
+                        layer.outputs[0])
+                line += " = {}(".format(layer.kernel)
+                for k, v in layer.attrs.items():
+                    line += "{}={}, ".format(k, v)
+                line = line.strip(", ")
+                line += ")"
+
+                if layer.kernel == "fluid.dygraph.base.to_variable" and not layer.attrs[
+                        "value"].startswith("params["):
+                    self.forward_func.extend(gen_codes([line], indent=indent))
+                    continue
+                else:
+                    self.init_func.extend(gen_codes([line], indent=2))
+
+                if len(layer.outputs) == 1:
+                    line = layer.outputs[0]
+                elif len(layer.outputs) == 2:
+                    line = layer.outputs[1]
+                else:
+                    line = ','.join(layer.outputs[1:])
+                if layer.kernel == "fluid.dygraph.base.to_variable" and layer.attrs[
+                        "value"].startswith("params["):
+                    line += " = self.{}".format(layer.outputs[0])
+                else:
+                    line += " = self.{}(".format(layer.outputs[0])
+                    for k, v in layer.inputs.items():
+                        line += "{}, ".format(v)
+                    line = line.strip(", ")
+                    line += ")"
+                self.forward_func.extend(gen_codes([line], indent=indent))
+            elif "prim" in layer.kernel:
+                func_name = layer.kernel.replace(".", "_")
+                from x2paddle.op_mapper.pytorch2paddle import prim2code
+                if hasattr(prim2code, func_name):
+                    func = getattr(prim2code, func_name)
+                    func(
+                        layer,
+                        indent=indent,
+                        init_func=self.init_func,
+                        forward_func=self.forward_func)
+                else:
+                    raise Exception(
+                        "The kind {} in paddle model is not supported yet.".
+                        format(layer.kernel))
+            else:
+                if len(layer.outputs) == 1:
+                    line = layer.outputs[0]
+                else:
+                    line = ','.join(layer.outputs)
+                line += " = {}(".format(layer.kernel)
+                for k, v in layer.inputs.items():
+                    line += "{}={}, ".format(k, v)
+                for k, v in layer.attrs.items():
+                    line += "{}={}, ".format(k, v)
+                line = line.strip(", ")
+                line += ")"
+                self.forward_func.extend(gen_codes([line], indent=indent))
+        if indent == 2:
+            write_code(code_dir)
+        else:
+            return self.init_func, self.forward_func
+
+    def dump_dygraph_parameter(self, code_dir):
+        params_output = open(os.path.join(code_dir, 'model.pdparams'), 'wb')
+        pickle.dump(self.parameters, params_output)
+        params_output.close()

x2paddle/decoder/pytorch_decoder.py

+#   Copyright (c) 2019  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 torch
+
+
+class PyTorchDecoder(object):
+    def __init__(self, script_path):
+        self.script = torch.jit.load(script_path)
+        self.graph = self._optimize_graph(self.script.inlined_graph)
+
+    def _optimize_graph(self, graph):
+        torch._C._jit_pass_constant_propagation(graph)
+        torch._C._jit_pass_dce(graph)
+        torch._C._jit_pass_lint(graph)
+        torch._C._jit_pass_peephole(graph)
+        torch._C._jit_pass_lint(graph)
+        torch._C._jit_pass_dce(graph)
+        torch._C._jit_pass_lint(graph)
+        graph = torch._C._jit_pass_canonicalize(graph)
+        torch._C._jit_pass_lint(graph)
+        return graph

x2paddle/op_mapper/pytorch2paddle/prim.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 torch
+from x2paddle.core.util import *
+
+
+def prim_Constant(mapper, graph, node):
+    """ 构造constant的PaddleLayer，该节点实现常量赋值。
+
+    TorchScript示例:
+        %2 : int = prim::Constant[value=-1]()
+        参数含义:
+        %2 (常量类型由赋值类型定义，该示例中为int型): 常量赋值结果输出。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    output = list(node.outputs())[0]
+    value = output.toIValue()
+    mapper.attrs[output_name] = value
+    if isinstance(value, str):
+        value = string(value)
+    graph.add_layer(
+        "prim.constant", inputs={}, outputs=[output_name], value=value)
+    return [], [output_name]
+
+
+def prim_GetAttr(mapper, graph, node):
+    """ 获取attribute信息。
+
+    TorchScript示例:
+        %27 : Tensor? = prim::GetAttr[name="bias"](%7)
+        参数含义:
+        %7 (Tensor): 输入Tensor。
+        %27 (Tensor): 输入Tensor。
+    """
+    current_node = node
+    field_name_list = [node.s('name')]
+    while True:
+        input_node = list(node.inputs())[0].node()
+        try:
+            field_name_list.insert(0, input_node.s('name'))
+            node = input_node
+        except Exception:
+            break
+    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]
+
+
+def prim_ListConstruct(mapper, graph, node):
+    """ 构造list的PaddleLayer。
+
+    TorchScript示例:
+        %86 : int[] = prim::ListConstruct(%84, %85)
+        参数含义:
+        %86 (list): list节点输出。
+        %84 (int/其他): list第一个元素信息。
+        %85 (int/其他): list第二个元素信息。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理每个输入
+    for i, input_name in enumerate(inputs_name):
+        layer_inputs["input{}".format(i)] = input_name
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer("prim.list", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
+def prim_RaiseException(mapper, graph, node):
+    """ 构造抛出异常的PaddleLayer。
+
+    TorchScript示例:
+        = prim::RaiseException(%76)
+        参数含义:
+        %76 (str): 异常信息。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%76
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer(
+        "prim.exception", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
+def prim_Loop(mapper, graph, node):
+    """ 构造loop循环的PaddleLayer。
+
+    TorchScript示例:
+        %x : Tensor = prim::Loop(%4, %3, %x.3)
+        block0(%i : int, %x.12 : Tensor):
+          %72 : int[] = prim::Constant[value=[6, 6]]()
+          ...
+          %x.5 : Tensor = aten::adaptive_avg_pool2d(%x.12, %_output_size.1)
+          -> (%3, %x.5)
+       参数含义:
+       %4 (int): 循环次数。
+       %3 (bool): 是否进入退出。
+       %x.3 (Tensor): 循环中修改的Tensor。
+       %x (Tensor): loop循环的输出，与%x.5对应。
+    """
+    node_outputs = mapper._get_outputs_name(node)
+    loop_inputs = {}
+    block = list(node.blocks())[0]
+    loop_outputs = node_outputs.copy()
+    for i, block_input_ivalue in enumerate(block.inputs()):
+        if i == 0:
+            block_input_node_name = '_x' + str(mapper.output_index)
+        else:
+            block_input_node_name = 'x' + str(mapper.output_index)
+        unique_id = block_input_ivalue.unique()
+        if unique_id not in mapper.outputs_info:
+            mapper.outputs_info[unique_id] = block_input_node_name
+            mapper.output_index += 1
+        if i == 0:
+            loop_input_node = list(node.inputs())[0].node()
+            script_loop_input_unique_id = list(node.inputs())[0].unique()
+            loop_input_node_name = mapper.outputs_info[
+                script_loop_input_unique_id]
+            mapper._check_input(graph, loop_input_node, loop_input_node_name,
+                                node_outputs)
+            loop_inputs['input'] = loop_input_node_name
+            loop_outputs.append(block_input_node_name)
+            node_outputs.append(block_input_node_name)
+        else:
+            loop_input_node = list(node.inputs())[i + 1].node()
+            script_loop_input_unique_id = list(node.inputs())[i + 1].unique()
+            loop_input_node_name = mapper.outputs_info[
+                script_loop_input_unique_id]
+            mapper._check_input(graph, loop_input_node, loop_input_node_name,
+                                node_outputs)
+            graph.add_layer(
+                "prim.equal",
+                inputs={'input': loop_input_node_name},
+                outputs=[block_input_node_name])
+            node_outputs.append(block_input_node_name)
+
+    graph.add_layer("prim.loop", inputs=loop_inputs, outputs=loop_outputs)
+    current_layer = list(graph.layers.values())[-1]
+    block_graph, graph_inputs = mapper.traverse(block, current_layer)
+    for i, input_name in enumerate(graph_inputs):
+        if input_name == loop_outputs[1]:
+            continue
+        current_layer.inputs['input-{}'.format(i)] = input_name
+    current_layer.add_block(block_graph)
+    return list(current_layer.inputs.values()), node_outputs
+
+
+def prim_If(mapper, graph, node):
+    """ 构造if控制流的PaddleLayer。
+
+    TorchScript示例:
+        %input.5 : Tensor = prim::If(%107)
+          block0():
+            %109 : Tensor = aten::t(%102)
+            %ret.2 : Tensor = aten::addmm(%103, %101, %109, %104, %104)
+            -> (%ret.2)
+          block1():
+            %111 : Tensor = aten::t(%102)
+            ...
+            -> (%output.4)
+        参数含义:
+        %107 (bool): if判断条件。
+        %input.5 (Tensor): if控制流的输出，与%output.4对应。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    node_outputs = [output_name]
+    input_node = list(node.inputs())[0].node()
+    script_input_unique_id = list(node.inputs())[0].unique()
+    input_node_name = mapper.outputs_info[script_input_unique_id]
+    mapper._check_input(graph, input_node, input_node_name, node_outputs)
+    graph.add_layer("prim.if", {'input': input_node_name}, [output_name])
+    current_layer = list(graph.layers.values())[-1]
+    block0 = list(node.blocks())[0]
+    block0_graph, graph_inputs0 = mapper.traverse(block0, current_layer)
+    len0 = 0
+    for i, input_name in enumerate(graph_inputs0):
+        current_layer.inputs['input-{}'.format(i)] = input_name
+        len0 = i
+    current_layer.add_block(block0_graph)
+    block1 = list(node.blocks())[1]
+    block1_graph, graph_inputs1 = mapper.traverse(block1, current_layer)
+    for i, input_name in enumerate(graph_inputs1):
+        current_layer.inputs['input-{}'.format(len0 + 1 + i)] = input_name
+    current_layer.add_block(block1_graph)
+    return list(current_layer.inputs.values()), node_outputs
+
+
+def prim_min(mapper, graph, node):
+    """ 构造min的PaddleLayer。
+
+    TorchScript示例:
+        %87 : int = prim::min(%86)
+        参数含义:
+        %86 (list): 输入。
+        %87 (int): 输出。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%86
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer("prim.min", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
+def prim_requires_grad(mapper, graph, node):
+    """ 构造是否计算梯度的PaddleLayer。
+
+    TorchScript示例:
+        %356 : bool = prim::requires_grad(%tensor.31)
+        参数含义:
+        %356 (bool): 输出，当前Tensor是否计算梯度。
+        %tensor.31 (Tensor): 输入的Tensor。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%86
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer(
+        "prim.requires_grad", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
+def prim_SetAttr(mapper, graph, node):
+    """ 设置attribute信息。
+
+    TorchScript示例:
+        = prim::SetAttr[name="num_batches_tracked"](%260, %277)
+        参数含义:
+        %260 (-): 属性名前缀。
+        %277 (-): 需要设置的值。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    field_name_list = []
+    tmp_node = node
+    while True:
+        input_node = list(tmp_node.inputs())[0].node()
+        try:
+            field_name_list.insert(0, input_node.s('name'))
+            tmp_node = input_node
+        except Exception:
+            break
+    field_name_list.append(node.s('name'))
+
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    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]
+
+
+def prim_shape(mapper, graph, node):
+    """ 构造获取shape的PaddleLayer。
+
+    TorchScript示例:
+        %4701 : int[] = prim::shape(%result.1)
+        参数含义:
+        %4701 (list): 输出，shape信息。
+        %result.1 (Tensor): 需要获取shape的值。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理输入0，即%input.8
+    mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs)
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer("prim.shape", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
+def prim_TupleConstruct(mapper, graph, node):
+    """ 构造tuple的PaddleLayer。
+
+    TorchScript示例:
+        %4492 : (Tensor, Tensor?) = prim::TupleConstruct(%x.46, %aux)
+        参数含义:
+        %4492 (tuple): 输出，tuple。
+        %x.46 (Tensor/其他): tuple第一个元素信息。
+        %aux (Tensor/其他): tuple第二个元素信息。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    layer_outputs = [output_name]
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = [output_name]
+    # 处理每个输入
+    for i, input_name in enumerate(inputs_name):
+        layer_inputs["input{}".format(i)] = input_name
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer("prim.tuple", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
+def prim_TupleUnpack(mapper, graph, node):
+    """ 构造获取tuple元素的PaddleLayer。
+
+    TorchScript示例:
+        %x.223 : Tensor, %aux.3 : Tensor? = prim::TupleUnpack(%4492)
+        参数含义:
+        %x.223 (Tensor/其他): 输出，tuple第一个元素信息。
+        %aux.3 (Tensor/其他): 输出，tuple第二个元素信息。
+        %4492 (tuple): 需要获取元素的tuple。
+    """
+    outputs_name = mapper._get_outputs_name(node)
+    layer_outputs = outputs_name
+    layer_inputs = {}
+    inputs_name, inputs_node = mapper._get_inputs_name(node)
+    # 获取当前节点输出的list
+    current_outputs = outputs_name
+    layer_inputs["input"] = inputs_name[0]
+    # 获取当前节点输入的list
+    current_inputs = list(layer_inputs.values())
+
+    graph.add_layer(
+        "prim.tuple_unpack", inputs=layer_inputs, outputs=layer_outputs)
+    return current_inputs, current_outputs
+
+
+def prim_Uninitialized(mapper, graph, node):
+    """ 构造表示编译器永远不会使用的值的PaddleLayer，该节点转换为None。
+
+    TorchScript示例:
+        %345 : bool = prim::Uninitialized()
+        参数含义:
+        %345 (bool): 输出，为赋值的bool。
+    """
+    output_name = mapper._get_outputs_name(node)[0]
+    output = list(node.outputs())[0]
+    mapper.attrs[output_name] = None
+    graph.add_layer(
+        "prim.constant", inputs={}, outputs=[output_name], value=None)
+    return [], [output_name]

x2paddle/op_mapper/pytorch2paddle/prim2code.py

+#   Copyright (c) 2019  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.
+
+
+def gen_codes(code_list, indent=0):
+    indent_blank = "    " * indent
+    codes = []
+    for code_line in code_list:
+        if code_line.strip() == "":
+            codes.append('\n')
+        else:
+            codes.append(indent_blank + code_line + '\n')
+    return codes
+
+
+def get_value(layer, key):
+    """ 进行optimizer后可能把inputs的value直接用数值代替（ConstantFuser），
+        会把input换成attr，所以需要此处的操作。
+    """
+    if key in layer.inputs:
+        return layer.inputs[key]
+    else:
+        return str(layer.attrs[key])
+
+
+def prim_add(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} + {}".format(layer.outputs[0],
+                                 get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_add_(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} + {} * {}".format(layer.outputs[0],
+                                      get_value(layer, "x"),
+                                      layer.attrs["alpha"],
+                                      get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_and(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} and {}".format(layer.outputs[0],
+                                   get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_append(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{}.append({})".format(
+        get_value(layer, "list"), get_value(layer, "element"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_assert(layer, indent=1, init_func=[], forward_func=[]):
+    if layer.attrs["type"] == "eq":
+        if isinstance(layer.attrs["value"], list):
+            s = ""
+            for v in layer.attrs["value"]:
+                s += "{} == {} or ".format(layer.attrs["key"], v)
+            if len(s) > 0:
+                s = s[:-4]
+            line = "assert {}, \'The {} must be {}!\'".format(
+                s, layer.attrs["key"], layer.attrs["value"])
+        else:
+            line = "assert {} == {}, \'The {} must be {}!\'".format(
+                layer.attrs["key"], layer.attrs["value"], layer.attrs["key"],
+                layer.attrs["value"])
+    else:
+        raise Exception("Not implement yet!")
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_constant(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {}".format(layer.outputs[0], layer.attrs["value"])
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_eq(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} == {}".format(layer.outputs[0],
+                                  get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_equal(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_exception(layer, indent=1, init_func=[], forward_func=[]):
+    line = "raise RaiseException({})".format(get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_if(layer, indent=1, init_func=[], forward_func=[]):
+    line = "if {} :".format(get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+    block = layer.blocks[0]
+    b_init_lines, b_forward_lines = block.gen_dygraph_code(indent=indent + 1)
+    init_func.extend(b_init_lines)
+    forward_func.extend(b_forward_lines)
+    block = layer.blocks[1]
+    if len(block.layers) > 0:
+        line = "else:"
+        forward_func.extend(gen_codes([line], indent=indent))
+        b_init_lines, b_forward_lines = block.gen_dygraph_code(
+            indent=indent + 1)
+        init_func.extend(b_init_lines)
+        forward_func.extend(b_forward_lines)
+
+
+def prim_getitem(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {}[{}]".format(layer.outputs[0],
+                                get_value(layer, "list"),
+                                get_value(layer, "index"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_gt(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} > {}".format(layer.outputs[0],
+                                 get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_le(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} <= {}".format(layer.outputs[0],
+                                  get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_len(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = len({})".format(layer.outputs[0], get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_lt(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} < {}".format(layer.outputs[0],
+                                 get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_list(layer, indent=1, init_func=[], forward_func=[]):
+    input_len = len(layer.inputs) + len(layer.attrs)
+    inputs_list = list()
+    for i in range(input_len):
+        inputs_list.append(get_value(layer, "input{}".format(i)))
+    inputs_str = ', '.join(inputs_list)
+    line = "{} = [{}]".format(layer.outputs[0], inputs_str)
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_loop(layer, indent=1, init_func=[], forward_func=[]):
+    loop_range = get_value(layer, "input")
+    line = "for {} in range({}):".format(layer.outputs[1], loop_range)
+    forward_func.extend(gen_codes([line], indent=indent))
+    block = layer.blocks[0]
+    b_init_lines, b_forward_lines = block.gen_dygraph_code(indent=indent + 1)
+    init_func.extend(b_init_lines)
+    forward_func.extend(b_forward_lines)
+
+
+def prim_min(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = min({})".format(layer.outputs[0], get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_mul(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} * {}".format(layer.outputs[0],
+                                 get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_ne(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} != {}".format(layer.outputs[0],
+                                  get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_neg(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_not(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = not {}".format(layer.outputs[0], get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_requires_grad(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = not {}.stop_gradient".format(layer.outputs[0],
+                                              get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_select(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {}[".format(layer.outputs[0], get_value(layer, "input"))
+    for dim in range(layer.attrs["dim"]):
+        line += ":, "
+    line += (get_value(layer, "index") + "]")
+    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))
+
+
+def prim_slice(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {}[{}: {}: {}]".format(layer.outputs[0],
+                                        get_value(layer, "input"),
+                                        get_value(layer, "start"),
+                                        get_value(layer, "end"),
+                                        get_value(layer, "step"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_sub(layer, indent=1, init_func=[], forward_func=[]):
+    line = "{} = {} - {}".format(layer.outputs[0],
+                                 get_value(layer, "x"), get_value(layer, "y"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_tuple(layer, indent=1, init_func=[], forward_func=[]):
+    input_len = len(layer.inputs) + len(layer.attrs)
+    inputs_list = list()
+    for i in range(input_len):
+        inputs_list.append(get_value(layer, "input{}".format(i)))
+    inputs_str = ', '.join(inputs_list)
+    line = "{} = ({})".format(layer.outputs[0], inputs_str)
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_tuple_unpack(layer, indent=1, init_func=[], forward_func=[]):
+    outputs_str = ', '.join(layer.outputs)
+    line = "{} = {}".format(outputs_str, get_value(layer, "input"))
+    forward_func.extend(gen_codes([line], indent=indent))
+
+
+def prim_warnings(layer, indent=1, init_func=[], forward_func=[]):
+    lines = ["import warnings"]
+    line = "warnings.warn({}, stacklevel={})".format(
+        get_value(layer, "input"), layer.attrs["stacklevel"])
+    lines.append(line)
+    forward_func.extend(gen_codes(lines, indent=indent))

x2paddle/op_mapper/pytorch2paddle/pytorch_op_mapper.py

+#   Copyright (c) 2019  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 torch
+import numpy as np
+from x2paddle.core.op_mapper import OpMapper
+from x2paddle.core.util import *
+from x2paddle.core.program import PaddleGraph
+from x2paddle.op_mapper.pytorch2paddle import prim
+from x2paddle.op_mapper.pytorch2paddle import aten
+
+
+class PyTorchOpMapper(OpMapper):
+    def __init__(self, decoder):
+        super(PyTorchOpMapper, self).__init__()
+        self.script = decoder.script
+        self.paddle_params = dict()
+        self.outputs_info = {}  # key为output unique id，value为当前节点的输出名字
+        self.pytorch_params = {}  # key为节点名，value为参数
+        self.attrs = {}  # key为节点名，value为属性值
+        self.output_index = 0
+        self.dygraph_name_id = {}  # 动态图__init__输出名字中的id，key为kernel类型，value为id
+        # 转换
+        self.check_op(decoder.graph)
+        self.graph, _ = self.traverse(decoder.graph)
+
+    def check_op(self, script_graph):
+        def _update_op_list(graph):
+            for node in graph.nodes():
+                op_list.append(node.kind())
+                for block in node.blocks():
+                    _update_op_list(block)
+
+        op_list = list()
+        _update_op_list(script_graph)
+        op_list = list(set(op_list))
+        unsupported_op_list = []
+        for op in op_list:
+            func_name = op.replace('::', '_')
+            if not (hasattr(prim, func_name) or hasattr(aten, func_name)):
+                unsupported_op_list.append(op)
+        if len(unsupported_op_list) > 0:
+            raise Exception("The kind {} in model is not supported yet.".format(
+                unsupported_op_list))
+
+    def traverse(self, script_graph, parent_layer=None):
+        # 用于获取graph的输入
+        def _update_graph_inputs(kind, inputs, outputs):
+            # extend只能放更新graph_inputs之前的情况：
+            # 1. loop的输出i也是输入；i是输入的原因是：子图中为父图得到的。
+            # 2. 在_check_input中需要使用to_variable。
+            # extend只能放更新graph_inputs之后的情况：
+            # 使用了append。
+            if kind != "aten::append":
+                current_node_outputs.extend(outputs)
+            for name in inputs:
+                if name not in current_node_outputs:
+                    graph_inputs.append(name)
+            if kind == "aten::append":
+                current_node_outputs.extend(outputs)
+
+        # 初始化
+        graph = PaddleGraph(parent_layer, graph_type="dygraph")
+        current_node_outputs = []
+        graph_inputs = []
+        # 转换输入节点
+        if isinstance(script_graph, torch._C.Graph):
+            for i, ivalue in enumerate(script_graph.inputs()):
+                node = ivalue.node()
+                if str(ivalue.type()) != "Tensor":
+                    graph.set_name(str(ivalue.type()).split(".")[-1])
+                inputs, outputs = self.data(graph, node, ivalue.unique())
+        # 转换中间节点
+        for node in script_graph.nodes():
+            kind = node.kind()
+            func_name = kind.replace('::', '_')
+            if hasattr(prim, func_name):
+                func = getattr(prim, func_name)
+                inputs, outputs = func(self, graph, node)
+                _update_graph_inputs(kind, inputs, outputs)
+            elif hasattr(aten, func_name):
+                func = getattr(aten, func_name)
+                inputs, outputs = func(self, graph, node)
+                _update_graph_inputs(kind, inputs, outputs)
+
+        # 转换输出节点
+        if hasattr(script_graph, 'returnNode'):
+            for i, ivalue in enumerate(script_graph.returnNode().inputs()):
+                if parent_layer.kernel == "prim.loop" and i == 0:
+                    continue
+                node = ivalue.node()
+                script_unique_id = ivalue.unique()
+                inputs, outputs = self.equal(
+                    graph,
+                    node,
+                    uid=script_unique_id,
+                    parent_layer=parent_layer,
+                    index=i)
+                _update_graph_inputs("equal", inputs, outputs)
+        # 设置graph的参数
+        if isinstance(script_graph, torch._C.Graph):
+            graph.set_parameters(self.paddle_params)
+        return graph, graph_inputs
+
+    def _get_outputs_name(self, node, attr_name=None):
+        outputs_name = []
+        for output_ivalue in node.outputs():
+            script_unique_id = output_ivalue.unique()
+            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:
+            output_name = '_x' + str(self.output_index)
+            self.output_index += 1
+            outputs_name.append(output_name)
+        return outputs_name
+
+    def _check_input(self,
+                     graph,
+                     node,
+                     output_name,
+                     node_outputs,
+                     add_dim=False):
+        if node.kind() == "prim::GetAttr":
+            param = self.pytorch_params[output_name]
+            if isinstance(param, np.ndarray):
+                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)))
+            else:
+                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={},
+                        outputs=[output_name],
+                        value=param["Tensor"])
+                else:
+                    graph.add_layer(
+                        "prim.constant",
+                        inputs={},
+                        outputs=[output_name],
+                        value=string(param)
+                        if isinstance(param, str) else param)
+            node_outputs.append(output_name)
+
+    def _get_inputs_name(self, node):
+        inputs_name = []
+        inputs_node = []
+        for script_input_ivalue in node.inputs():
+            script_input_node = script_input_ivalue.node()
+            script_input_unique_id = script_input_ivalue.unique()
+            input_name = self.outputs_info[script_input_unique_id]
+            inputs_node.append(script_input_node)
+            inputs_name.append(input_name)
+        return inputs_name, inputs_node
+
+    def data(self, graph, node, uid):
+        for output_ivalue in node.outputs():
+            script_unique_id = output_ivalue.unique()
+            if script_unique_id in self.outputs_info or script_unique_id != uid:
+                continue
+            node_name = 'x' + str(self.output_index)
+            self.outputs_info[script_unique_id] = node_name
+            self.output_index += 1
+        output_name = self.outputs_info[uid]
+        graph.add_layer(
+            "fluid.dygraph.base.to_variable",
+            inputs={},
+            outputs=[node_name],
+            value=output_name)
+        return [], [output_name]
+
+    def equal(self, graph, node, uid=None, parent_layer=None, index=None):
+        if parent_layer is not None and index is not None:
+            # block的输出
+            input_node_name = self.outputs_info[uid]
+            control_output_id = index
+            if parent_layer.kernel == "prim.loop":
+                control_output_id = index - 1
+            output_node_name = parent_layer.outputs[control_output_id]
+            current_outputs = [output_node_name]
+            self._check_input(graph, node, input_node_name, current_outputs)
+            graph.add_layer(
+                "prim.equal",
+                inputs={'input': input_node_name},
+                outputs=[output_node_name])
+            return [input_node_name], current_outputs

x2paddle/optimizer/fusion/__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 .fc_fuser import FcFuser
+from .fc_fuse_pass import FcFusePass
+from .nn_adaptive_pool2d_fuser import NnAdaptivePool2dFuser
+from .nn_adaptive_pool2d_fuse_pass import NnAdaptivePool2dFusePass
+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/fusion/fc_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 import FcFuser
+from x2paddle.optimizer.pass_manager import pass_register
+
+
+@pass_register
+class FcFusePass(Pass):
+    name = "fc_fuse_pass"
+
+    def __init__(self):
+        Pass.__init__(self)
+
+    def apply(self, graph):
+        fuser = FcFuser()
+        fuser.operate(graph, match_kind="topo")
+
+
+# 用于注册
+fc_fuse_pass = FcFusePass()

x2paddle/optimizer/fusion/fc_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 numpy as np
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+
+
+class FcFuser(FuseBase):
+    def __init__(self):
+        self.linear_index = 0
+        super(FcFuser, self).__init__(graph_type="dygraph")
+
+    def build_pattern(self):
+        """ 描述需要替换的fc图结构。
+        fc层模式python实现代码示例:
+            x149 = 2
+            x151 = x146.shape
+            x151 = len(x151)
+            x152 = x151 == x149
+            if x152 :
+                x147 = self.x147
+                x154 = fluid.layers.transpose(x=x147, perm=[1, 0])
+                x148 = self.x148
+                x155 = fluid.layers.addmm(input=x148, x=x146, y=x154, beta=1, alpha=1)
+                x153 = x155
+            else:
+                x147 = self.x147
+                x157 = fluid.layers.transpose(x=x147, perm=[1, 0])
+                x158 = fluid.layers.matmul(x=x146, y=x157)
+                x159 = True
+                if x159 :
+                    x148 = self.x148
+                    x161 = x158 + 1 * x148
+                    x160 = x161
+                else:
+                    x160 = x158
+                x153 = x160
+        """
+
+        def gen_name(id):
+            return "x" + str(id)
+
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(0)], value=2)
+        self.pattern.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(1)], value=1)
+        self.pattern.add_layer(
+            "prim.shape", inputs={'input': "fc-input-0"},
+            outputs=[gen_name(2)])
+        self.pattern.add_layer(
+            "prim.len", inputs={'input': gen_name(2)}, outputs=[gen_name(2)])
+        self.pattern.add_layer(
+            "prim.eq",
+            inputs={"eq0": gen_name(2),
+                    "eq1": gen_name(0)},
+            outputs=[gen_name(3)])
+        self.pattern.add_layer("prim.if", {'input': gen_name(3)}, [gen_name(4)])
+        self.pattern.outputs.append(gen_name(4))
+        if_layer1 = self.pattern.layers[list(self.pattern.layers.keys())[-1]]
+        pattern_block0 = PaddleGraph(if_layer1, graph_type="dygraph")
+        pattern_block0.add_layer(
+            "fluid.dygraph.base.to_variable",
+            inputs={},
+            outputs=[gen_name(5)],
+            value="params[{}]".format(string(gen_name(5))))
+        pattern_block0.add_layer(
+            "fluid.layers.transpose",
+            inputs={"x": gen_name(5)},
+            outputs=[gen_name(6)],
+            perm=[1, 0])
+        pattern_block0.add_layer(
+            "fluid.dygraph.base.to_variable",
+            inputs={},
+            outputs=[gen_name(7)],
+            value="params[{}]".format(string(gen_name(7))))
+        pattern_block0.add_layer(
+            "fluid.layers.addmm",
+            inputs={"input": gen_name(7),
+                    "x": "fc-input-0",
+                    "y": gen_name(6)},
+            outputs=[gen_name(8)],
+            beta=1,
+            alpha=1)
+        if_layer1.inputs["input-0"] = "fc-input-0"
+        self.pattern.inputs.append("fc-input-0")
+        pattern_block0.add_layer(
+            "prim.equal", inputs={'input': gen_name(8)}, outputs=[gen_name(4)])
+        if_layer1.add_block(pattern_block0)
+        pattern_block1 = PaddleGraph(if_layer1, graph_type="dygraph")
+        pattern_block1.add_layer(
+            "fluid.dygraph.base.to_variable",
+            inputs={},
+            outputs=[gen_name(5)],
+            value="params[{}]".format(string(gen_name(5))))
+        pattern_block1.add_layer(
+            "fluid.layers.transpose",
+            inputs={"x": gen_name(5)},
+            outputs=[gen_name(6)],
+            perm=[1, 0])
+        pattern_block1.add_layer(
+            "fluid.layers.matmul",
+            inputs={"x": "fc-input-0",
+                    "y": gen_name(6)},
+            outputs=[gen_name(9)])
+        if_layer1.inputs["input-1"] = "fc-input-0"
+        pattern_block1.add_layer(
+            "prim.constant", inputs={}, outputs=[gen_name(10)], value=True)
+        pattern_block1.add_layer("prim.if", {'input': gen_name(10)},
+                                 [gen_name(11)])
+        if_layer2 = pattern_block1.layers[list(pattern_block1.layers.keys())[
+            -1]]
+        pattern_block1_block0 = PaddleGraph(if_layer2, graph_type="dygraph")
+        pattern_block1_block0.add_layer(
+            "fluid.dygraph.base.to_variable",
+            inputs={},
+            outputs=[gen_name(12)],
+            value="params[{}]".format(string(gen_name(12))))
+        pattern_block1_block0.add_layer(
+            "prim.add_",
+            inputs={"x": gen_name(9),
+                    "y": gen_name(12)},
+            outputs=[gen_name(13)],
+            alpha=1)
+        if_layer2.inputs["input-0"] = gen_name(9)
+        pattern_block1_block0.add_layer(
+            "prim.equal",
+            inputs={'input': gen_name(13)},
+            outputs=[gen_name(11)])
+        if_layer2.add_block(pattern_block1_block0)
+        pattern_block1_block1 = PaddleGraph(if_layer2, graph_type="dygraph")
+        pattern_block1_block1.add_layer(
+            "prim.equal", inputs={'input': gen_name(9)},
+            outputs=[gen_name(11)])
+        if_layer2.inputs["input-1"] = gen_name(9)
+        pattern_block1.add_layer(
+            "prim.equal", inputs={'input': gen_name(11)},
+            outputs=[gen_name(4)])
+        if_layer2.add_block(pattern_block1_block1)
+        if_layer1.add_block(pattern_block1)
+        self.pattern.build(inputs={"input-0": "fc-input-0"})
+
+    def insert_new_layer(self, graph, parameters, matches):
+        new_layer = self.gen_new_layer(parameters, matches)
+        new_layer_id = list(matches.keys())[0]
+        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[2]]
+        input_name = layer.inputs["input"]
+        layer = matches[layers_id[5]]
+        output_name = layer.outputs[0]
+        layer = matches[layers_id[6]]
+        weight_name = layer.attrs["value"][8:-2]
+        layer = matches[layers_id[8]]
+        bias_name = layer.attrs["value"][8:-2]
+        attrs = dict()
+        attrs["input_dim"] = parameters[weight_name].shape[1]
+        attrs["output_dim"] = parameters[weight_name].shape[0]
+        linear_name = "linear{}".format(self.linear_index)
+        self.linear_index += 1
+        parameters["{}.weight".format(linear_name)] = parameters[
+            weight_name].transpose((1, 0))
+        parameters["{}.bias".format(linear_name)] = np.squeeze(parameters[
+            bias_name])
+        new_layer = PaddleLayer(
+            layers_id[0],
+            "fluid.dygraph.Linear",
+            inputs={"input": input_name},
+            outputs=[linear_name, output_name],
+            **attrs)
+        return new_layer

x2paddle/optimizer/optimizer.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.fusion import *
+from x2paddle.optimizer.pass_manager import PassManager
+
+
+class GraphOptimizer(object):
+    def __init__(self):
+        self.passes = [
+            "fc_fuse_pass",
+            #             "nn_adaptive_pool2d_fuse_pass",
+            #             "functional_adaptive_pool2d_fuse_pass",
+            #             "batchnorm2d_fuse_pass",
+            "constant_fuse_pass"
+        ]
+
+    def optimize(self, graph):
+        for pass_name in self.passes:
+            pass_ = PassManager.lookup(pass_name)()
+            pass_.apply(graph)
+            print("{} done!".format(pass_name))
+        return graph

x2paddle/optimizer/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.
+
+
+class Pass(object):
+    name = "pass"
+
+    def __init__(self):
+        pass
+
+    def apply(self, graph):
+        raise NotImplementedError("The apply function must be implemented!")
+
+    @classmethod
+    def get_name(cls):
+        return cls.name

x2paddle/optimizer/pass_manager.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.
+
+
+class PassManager(object):
+    """ pass管理器。
+    """
+    # pass_map存储name与其对应的pass
+    pass_map = dict()
+
+    def __init__(self):
+        pass
+
+    @staticmethod
+    def add_new_pass(name, pass_):
+        if name not in PassManager.pass_map:
+            PassManager.pass_map[name] = pass_
+
+    @staticmethod
+    def clear():
+        PassManager.passes = list()
+
+    @staticmethod
+    def lookup(name):
+        return PassManager.pass_map[name]
+
+
+def pass_register(cls):
+    name = cls.get_name()
+    PassManager.add_new_pass(name, cls)
+    return cls

x2paddle/optimizer/pattern_matcher.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.core.program import PaddleGraph
+
+
+class PatternMatcher(object):
+    def __init__(self, pattern):
+        self.pattern = pattern
+        # matches的每个match是按照拓扑排序组成layer的dict
+        self.matches = list()
+
+    def operate(self, graph, match_kind="topo"):
+        if match_kind == "topo":
+            self.detect_patterns_by_topo(graph)
+        elif match_kind == "edge":
+            self.detect_patterns_by_edge(graph)
+        self.remove_overlapped_match()
+        return self.matches
+
+    def detect_patterns_by_topo(self, graph):
+        """ 找到与模式匹配的子图，
+            并将子图的id以拓扑排序存放到subgraph_id2layers。
+        """
+
+        def get_subgraph(pattern, graph, start_index):
+            pattern_index = 0
+            pattern_id2layers = pattern.get_global_layers()
+            pattern_ids = list(pattern_id2layers.keys())
+            subgraph_id2layers = dict()
+            graph_layers = dict(list(graph.layers.items())[start_index:])
+            for layer_id, layer in graph_layers.items():
+                pattern_layer = pattern.layers[list(pattern.layers.keys())[
+                    pattern_index]]
+                if layer.kernel == pattern_layer.kernel:
+                    subgraph_id2layers[layer_id] = layer
+                    pattern_layer_id = pattern_layer.id
+                    # 判断输入连接是否一致
+                    if layer_id in graph.edges_in:
+                        if pattern_layer_id not in pattern.edges_in:
+                            return False
+                        else:
+                            if len(graph.edges_in[layer_id]) != len(
+                                    pattern.edges_in[pattern_layer_id]):
+                                return False
+                        layer_in = graph.edges_in[layer_id]
+                        pattern_layer_in = pattern.edges_in[pattern_layer_id]
+                        for i in range(len(layer_in)):
+                            layer_id_in = layer_in[i]
+                            pattern_layer_id_in = pattern_layer_in[i]
+                            if pattern_layer_id_in != -1:
+                                subgraph_ids = list(subgraph_id2layers.keys())
+                                if pattern_ids.index(pattern_layer_id_in) == \
+                                subgraph_ids.index(layer_id_in):
+                                    # 判断pattern输入在pattern_ids的索引
+                                    # 和graph输入在subgraph_ids的索引一致
+                                    continue
+                                return False
+                    # 判断subgraph中的节点是否被外部图使用到(如若被使用到则无效)
+                    if layer_id in graph.edges_out:
+                        if pattern_layer_id not in pattern.edges_out:
+                            if not set(pattern_layer.outputs).issubset(
+                                    pattern.outputs):
+                                # 若pattern当前layer的输出是pattern的输出，则是正确的
+                                return False
+                        else:
+                            if len(graph.edges_out[layer_id]) != len(
+                                    pattern.edges_out[pattern_layer_id]):
+                                # 如果在每个节点edges_in相同的情况下，edges_out数目相同则说明无节点在subgraph外被用到
+                                if not set(pattern_layer.outputs).issubset(
+                                        pattern.outputs):
+                                    # 若pattern当前layer的输出是pattern的输出，则是正确的
+                                    return False
+                    # 当为控制流时的处理
+                    if layer.kernel == "prim.if" or layer.kernel == "prim.loop":
+                        if len(pattern_layer.blocks) != len(layer.blocks):
+                            return False
+                        for i, b in enumerate(pattern_layer.blocks):
+                            match_info = get_subgraph(pattern_layer.blocks[i],
+                                                      layer.blocks[i], 0)
+                            if match_info is not False:
+                                subgraph_id2layers.update(match_info)
+                            else:
+                                return False
+                    pattern_index += 1
+                    if pattern_index == len(pattern.layers):
+                        return subgraph_id2layers
+                else:
+                    return False
+            return subgraph_id2layers
+
+        for i, (layer_id, layer) in enumerate(graph.layers.items()):
+            match_info = get_subgraph(self.pattern, graph, i)
+            if match_info:
+                self.matches.append(match_info)
+            for j, block in enumerate(layer.blocks):
+                if len(block.layers) > 0:
+                    self.detect_patterns_by_topo(layer.blocks[j])
+
+    def detect_patterns_by_edge(self, graph):
+        """当遇见顺序没有强制规定的pattern时使用该方式
+        """
+        pass
+
+    def remove_overlapped_match(self):
+        """ 如果2个子图有重叠，只取前一个子图。
+        """
+        match_ids = []
+        for i, match in enumerate(self.matches):
+            is_overlapped = False
+            for id in match.keys():
+                if id in match_ids:
+                    self.matches.pop(i)
+                    is_overlapped = True
+                    break
+            if not is_overlapped:
+                match_ids.extend(list(match.keys()))
+
+
+def get_subgraph(prefix_layer_id, suffix_layer_id, graph):
+    """ 根据prefix_layer_id和suffix_layer_id获取需要子图。
+        Args:
+            prefix_layer_id (str): 起初为一个空字符串，之后为suffix_layer_id分割出来的前缀。
+            suffix_layer_id (str): 起初为以一个layer的id，之后将分割部分给prefix_layer_id；例如”57.0.1“；
+            graph (x2paddle.core.program.PaddleGraph): 需要进行pass的子图。
+    """
+    id_part = suffix_layer_id.split(".")
+    if len(id_part) == 1:
+        return graph
+    if prefix_layer_id == "":
+        layer_id = id_part[0]
+        prefix_layer_id += ".".join(id_part[:2])
+    else:
+        layer_id = prefix_layer_id + "." + id_part[0]
+        prefix_layer_id += ("." + ".".join(id_part[:2]))
+    subgraph = graph.layers[layer_id].blocks[int(id_part[1])]
+    suffix_layer_id = ".".join(id_part[2:])
+    return get_subgraph(prefix_layer_id, suffix_layer_id, subgraph)
+
+
+class FuseBase(object):
+    def __init__(self, graph_type):
+        self.pattern = PaddleGraph(graph_type=graph_type)
+
+    def operate(self, graph, match_kind="topo"):
+        parameters = graph.parameters
+        self.build_pattern()
+        self.perform_pattern_matcher(graph, match_kind)
+        for match in self.matches:
+            first_layer_id = list(match.keys())[0]
+            subgraph = get_subgraph("", first_layer_id, graph)
+            self.insert_new_layer(subgraph, parameters, match)
+        self.delete_inter_layer(graph)
+        graph.build()
+
+    def perform_pattern_matcher(self, graph, match_kind="topo"):
+        """ 执行模式匹配，找到匹配的子图。
+        """
+        pattern_matcher = PatternMatcher(self.pattern)
+        self.matches = pattern_matcher.operate(graph, match_kind)
+
+    def delete_inter_layer(self, graph):
+        """ 删除不需要的中间layer及其对应参数。
+        """
+        for match in self.matches:
+            first_layer_id = list(match.keys())[0]
+            subgraph = get_subgraph("", first_layer_id, graph)
+            for layer_id, layer in match.items():
+                if layer.kernel == "fluid.dygraph.base.to_variable" and \
+                layer.attrs["value"].startswith("params["):
+                    param_name = layer.attrs["value"][8:-2]
+                    if param_name in graph.parameters:
+                        graph.parameters.pop(param_name)
+                if layer_id in subgraph.layers:
+                    # layer_id可能是属于子图的，此时删除父layer，即删除整个子图
+                    subgraph.layers.pop(layer_id)