# -*- coding:UTF-8 -*- # 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 os.path as osp from treelib import Tree from queue import Queue from x2paddle.optimizer.code_optimizer.layer_code_generator import gen_layer_code, rename_layers, NN_KERNEL_WITH_PARAMS, NN_KERNEL_NAME from x2paddle.optimizer.code_optimizer.subgraphs_union import distinguish_sequential, get_inputs_outputs from x2paddle.core.program import PaddleLayer from x2paddle.optimizer.code_optimizer.parameter_tree import PamareterNode, PamareterTree SEPARATOR_IN_SCOPE = "/" class HierarchicalTree(Tree): """ 定义层次树。 """ def __init__(self, pd_graph): super(HierarchicalTree, self).__init__() self.pd_graph = pd_graph self.script = pd_graph.script self.create_node("Module", self.pd_graph.name) # create root self._hierarchical_order = dict() self.codes = list() self.identifier_idx = dict() self.param_tree = PamareterTree() self.module_name2count = dict() def insert(self, layer): """ 往层次树中插入节点。 Args: layer (PaddleLayer): 需要插入的节点。 """ scope_name = layer.scope_name if scope_name == "": if layer.kernel == "prim.tuple" or layer.kernel == "prim.tuple_unpack": layer_id = layer.id layer_id_list = list() for input_layer_id in self.pd_graph.edges_in[layer_id]: layer_id_list.append(int(input_layer_id)) layer_id_list = list(set(layer_id_list)) layer_id_list.sort(reverse=True) for input_layer_id in layer_id_list: input_layer_id_str = str(input_layer_id) if self.pd_graph.layers[input_layer_id_str].scope_name != "": scope_name = self.pd_graph.layers[input_layer_id_str].scope_name break layer.scope_name = scope_name else: self.create_node(tag=layer.id, identifier="no_scope_" + layer.id, parent=self.pd_graph.name, data=layer) return scopes = scope_name.split(SEPARATOR_IN_SCOPE) for idx, scope in enumerate(scopes): parent = SEPARATOR_IN_SCOPE.join(scopes[:idx])#.lower() identifier = SEPARATOR_IN_SCOPE.join(scopes[:idx + 1])#.lower() if self.contains(identifier): if idx != len(scopes) - 1: parent_node = self.parent(identifier) self.move_node(identifier, parent_node.identifier) continue else: if self.get_node(identifier).data is None: data = layer if identifier not in self.identifier_idx: self.identifier_idx[identifier] = 0 else: self.identifier_idx[identifier] += 1 identifier_name = identifier + SEPARATOR_IN_SCOPE + str(self.identifier_idx[identifier]) self.create_node(tag=scopes[idx], identifier=identifier_name, parent=identifier, data=data) data.scope_name = identifier_name continue else: data = self[identifier].data self[identifier].data = None parent_node = self.parent(identifier) self.move_node(identifier, parent_node.identifier) if identifier not in self.identifier_idx: self.identifier_idx[identifier] = 0 else: self.identifier_idx[identifier] += 1 self.create_node(tag=scopes[idx], identifier=identifier + SEPARATOR_IN_SCOPE + str(self.identifier_idx[identifier]), parent=identifier, data=data) self.identifier_idx[identifier] += 1 data = layer self.create_node(tag=scopes[idx], identifier=identifier + SEPARATOR_IN_SCOPE + str(self.identifier_idx[identifier]), parent=identifier, data=data) continue if idx == 0 and not self.contains(identifier): data = layer if idx == len(scopes) - 1 else None self.create_node(tag=scopes[idx], identifier=identifier, parent=self.pd_graph.name, data=data) else: if idx == len(scopes) - 1: if parent == "": childs = self.children(self.pd_graph.name) parent = self.pd_graph.name else: childs = self.children(parent) prefix = identifier identifiers = list() for child in childs: child_identifier = child.identifier if child_identifier.startswith(prefix) and child_identifier != prefix: identifiers.append(child_identifier) if len(identifiers) == 0: identifier = prefix + "_0" else: identifier_ids = list() for id_obj in identifiers: identifier_ids.append(int(id_obj.split("_")[-1])) identifier_ids.sort() identifier = prefix + "_{}".format(identifier_ids[-1] + 1) data = layer if idx == len(scopes) - 1 else None self.create_node(tag=scopes[idx], identifier=identifier, parent=parent, data=data) def update_hierarchical_order(self): """ 更新层次排序，使用一个字典存储该信息， 关键字为当前层次，值为节点名字。 """ hierarchical_order = dict() queue = Queue() queue.put(item=(self.pd_graph.name, 0), block=False) while not queue.empty(): node_name, cur_level = queue.get(block=False) node_inst = self[node_name] if cur_level not in hierarchical_order: hierarchical_order[cur_level] = [] hierarchical_order[cur_level].append(node_name) for successor_name in node_inst.successors(self.identifier): queue.put(item=(successor_name, cur_level + 1), block=False) self._hierarchical_order = hierarchical_order def analyze_attrs_table(self, attrs_table): """ 分析属性表格，哪些属性取值不一致。 """ diff_attrs_column = list() for column in list(attrs_table.columns): elements = list(attrs_table.get(column)) base = elements[0] for element in elements[1:]: if isinstance(base, str) and "'" not in base: break if element != base: diff_attrs_column.append(column) break return diff_attrs_column def merge_node(self, sub_layers_list, attrs_table, node_name2sub_layers, module_name): """ 将一个scope的节点合成一个Module（Class），并将对应的Class代码 放到code字符串中。 """ def get_node_name(sub_layers): for k, v in node_name2sub_layers.items(): if v == sub_layers: node_name = k break return node_name sub_layers = sub_layers_list[0] node_name = get_node_name(sub_layers) sub_layers, _, _ = rename_layers(sub_layers) diff_attrs_column = self.analyze_attrs_table(attrs_table) if module_name is None: module_name = node_name.replace("/", "_") #node_name.split("/")[-1] module_name = module_name[0].upper() + module_name[1:] if module_name in self.module_name2count: module_name = module_name + "_0" code_str = gen_layer_code(self.pd_graph, sub_layers, module_name, different_attrs=diff_attrs_column) self.codes.append(code_str) for sub_layers in sub_layers_list: inputs, outputs = get_inputs_outputs(self.pd_graph, sub_layers) inputs_dict = dict() for i, input in enumerate(inputs): inputs_dict["input_{}".format(i)] = input if module_name in self.module_name2count: self.module_name2count[module_name] += 1 else: self.module_name2count[module_name] = 0 if module_name.lower() in NN_KERNEL_NAME.values(): mn = module_name.lower() + "__" else: mn = module_name.lower() outputs = ["{}/{}".format(mn, self.module_name2count[module_name])] + outputs node_name = get_node_name(sub_layers) diff_attrs = dict() for column in diff_attrs_column: diff_attrs[column] = attrs_table.get(column).loc[node_name] node_name_seg = node_name.split(SEPARATOR_IN_SCOPE) node_name_seg[-1] = module_name.lower() new_node_name = SEPARATOR_IN_SCOPE.join(node_name_seg) new_layer = PaddleLayer(id=list(sub_layers.keys())[-1], kernel="module", inputs=inputs_dict, outputs=outputs, scope_name=new_node_name, module=module_name, **diff_attrs) _, nn_param_nodes, _ = rename_layers(sub_layers, self.param_tree) param_node = PamareterNode(old_name=outputs[0]) for node in nn_param_nodes: param_node.add_child(node) self.param_tree.add_node(param_node) for i, (layer_id, layer) in enumerate(sub_layers.items()): if i == len(sub_layers) - 1: self.pd_graph.layers[layer_id] = new_layer else: self.pd_graph.layers.pop(layer_id) self.pd_graph.build() self[node_name].data = new_layer def find_subgraph_diff(self, module_name2sub_layers, module_name2sub_identifiers, node_name2sub_layers, name): """ 查找子图的diff，主要是输入参数的diff。 """ sub_layers = module_name2sub_layers[name] sub_identifiers = module_name2sub_identifiers[name] new_sub_layers, new_sub_sequentials, sequentials2attrs_table = distinguish_sequential(self.pd_graph, name, sub_layers, sub_identifiers, node_name2sub_layers) module_name2sub_layers.pop(name) module_name2sub_identifiers.pop(name) for k, v in new_sub_layers.items(): module_name2sub_layers[k] = v module_name2sub_identifiers[k] = new_sub_sequentials[k] return sequentials2attrs_table def convert_subgraph_to_layer(self): """ 1. 根据_hierarchical_order，从最深的层次开始将 子图合并成layer（即合成节点）。 2. 根据参数名新旧对应关系，更新参数名。 """ depths = sorted(list(self._hierarchical_order.keys()), reverse=True) all_name_old2new = dict() for depth in depths[1:]: # Module的名字与子图的对应关系 module_name2sub_layers = dict() # Module的名字与子图中layer命名的对应关系 module_name2sub_identifiers = dict() # 层次树中包含子树的节点，其节点名与子图对用关系 node_name2sub_layers = dict() for node_name in self._hierarchical_order[depth]: node_inst = self[node_name] if node_inst.data is None: sub_layers = dict() sub_identifiers = dict() for successor_name in node_inst.successors(self.identifier): sub_layers[self[successor_name].data.id] = self[successor_name].data sub_identifiers[self[successor_name].data.id] = self[successor_name].data.scope_name.split("/")[-1] node_name2sub_layers[node_name] = sub_layers node_name_segs = node_name.split("/") # 获取Module的名字 module = self.script is_largest_module = False # 当前module是否是最外层的Module for name_id, name in enumerate(node_name_segs): name = name.split("__")[0] if not hasattr(module, name): is_largest_module = True break module = getattr(module, name) if is_largest_module: if name_id == 0: module_name = name else: module_name = "_".join(node_name_segs) else: module_name = module._get_name() if module_name in module_name2sub_layers: module_name2sub_layers[module_name].append(sub_layers) module_name2sub_identifiers[module_name].append(sub_identifiers) else: module_name2sub_layers[module_name] = [sub_layers] module_name2sub_identifiers[module_name] = [sub_identifiers] module_names = list(module_name2sub_layers.keys()) for module_name in module_names: sequentials2attrs_table = self.find_subgraph_diff(module_name2sub_layers, module_name2sub_identifiers, node_name2sub_layers, module_name) for name in sequentials2attrs_table.keys(): if name.startswith("Sequential"): # 若Module的名字为Sequential，则以scope_name的名字来命名，在merge_node中实现 module_name = None else: module_name = name self.merge_node(module_name2sub_layers[name], sequentials2attrs_table[name], node_name2sub_layers, module_name) def update_parameters(self): """ 更新参数。 """ self.param_tree.traverse() full_old_name_list = copy.deepcopy(list(self.pd_graph.parameters.keys())) for old_name, new_name in self.param_tree.old2new.items(): for full_old_name in full_old_name_list: if full_old_name.startswith("{}.".format(old_name)): full_new_name = full_old_name.replace("{}.".format(old_name), "{}.".format(new_name)) params = self.pd_graph.parameters.pop(full_old_name) self.pd_graph.parameters[full_new_name] = params if full_old_name == old_name: full_new_name = full_old_name.replace(old_name, new_name) params = self.pd_graph.parameters.pop(full_old_name) self.pd_graph.parameters[full_new_name] = params def save_source_files(self, save_dir): def gen_main_code(): input_data_name = ', '.join(self.pd_graph.inputs) run_func_list = list() run_func_list.append("def main({}):".format(input_data_name)) run_func_list.append(" # 共{}个输入".format(len(self.pd_graph.inputs_info))) for k, v in self.pd_graph.inputs_info.items(): 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)), " model = {}()".format(self.pd_graph.name), " model.set_dict(params)", " model.eval()", " out = model({})".format(input_data_name), " return out"]) return "\n".join(run_func_list) self.update_hierarchical_order() self.convert_subgraph_to_layer() 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) f = open(osp.join(save_dir, 'x2paddle_code.py'), 'w') f.write(import_str) for code in self.codes: f.write(code) f.write("\n") run_func = gen_main_code() f.write(run_func) f.close()