# -*- 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 import os.path as osp from x2paddle.core.program import PaddleLayer from x2paddle.optimizer.pytorch_code_optimizer.subgraphs_union import construct_attrs_table, get_inputs_outputs from x2paddle.optimizer.pytorch_code_optimizer.layer_code_generator import gen_layer_code, rename_layers from x2paddle.optimizer.pytorch_code_optimizer.parameter_tree import PamareterNode, PamareterTree NoModuleStart = ["paddle.nn.ReLU"] class Apriori(object): """ 使用Apriori算法挖掘频繁子图 1. 构建频繁1项集 2. 挖掘频繁k项集 3. 最终k项集和节点数满足最少节点数的子图组成集合GS Args: min_support (int): 子图出现次数的最小值。 """ def __init__(self, min_support): self.min_support = min_support def is_match(self, item, sublayers): for i in range(len(item)): if len(sublayers) <= i or item[i] != sublayers[i].kernel: return False return True def create_C1(self): # 构建候选1-项集 C1 = list() for layer_id, layer in self.layers.items(): if layer.kernel == "paddle.to_tensor" or \ layer.kernel == "prim.if" or \ layer.kernel == "prim.loop": #or \ # layer.kernel == "prim.list" or \ # layer.kernel == "prim.tuple" or \ # layer.kernel == "prim.dict_construct": continue if self.pd_graph.edges_in.get(layer_id, 0) == 0 and \ self.pd_graph.edges_out.get(layer_id, 0) == 0: continue if [layer.kernel] not in C1: C1.append([layer.kernel]) return C1 def create_Ck(self, Lk_last, C1): # 构建候选k-项集 Ck = list() for item in Lk_last: for item_one in C1: new_item = copy.deepcopy(item) new_item.append(item_one[0]) if new_item[0] in NoModuleStart: continue Ck.append(new_item) return Ck def generate_Lk_by_Ck(self, Ck): # 生成频繁k-项集 Lk = list() for item in Ck: count = 0 for i in range(len(self.layers)): sublayers = list(self.layers.values())[i:] if self.is_match(item, sublayers): count += 1 if count >= self.min_support: Lk.append(item) return Lk def run(self, graph): self.pd_graph = graph self.layers = graph.layers itemset = list() C1 = self.create_C1() L1 = self.generate_Lk_by_Ck(C1) Lk = L1 while len(Lk) > 0: Ck = self.create_Ck(Lk, C1) Lk = self.generate_Lk_by_Ck(Ck) itemset.extend(Lk) return itemset class DP(object): """ 使用动动态规划找到使代码最短的组合方式。 """ def __init__(self, combination_itemset): self.combination_itemset = combination_itemset def get_combination_id(self, combination, layers): combination_id = list() for layer_obj in combination: if len(layer_obj) > 1: kernel_itemset = list() for layer_id in layer_obj: kernel_itemset.append(layers[layer_id].kernel) id = self.combination_itemset.index(kernel_itemset) combination_id.append(id) else: combination_id.append(-1) return combination_id def run(self, graph): layers = graph.layers layer_combination_list = list() for i, (layer_id, layer) in enumerate(layers.items()): if i == 0: layer_combination_list.append([[layer_id]]) continue current_itemset = [layer_id] kernel_itemset = [layer.kernel] candidate_itemset = list() min_count = len(layers) prefix_ids = list(range(i)) prefix_ids.reverse() for j in prefix_ids: current_layer_id = list(layers.keys())[j] current_layer = list(layers.values())[j] current_itemset.insert(0, current_layer_id) kernel_itemset.insert(0, current_layer.kernel) if kernel_itemset in self.combination_itemset: current_count = len(layer_combination_list[j - 1]) all_count = current_count + 1 if all_count < min_count: min_count = all_count candidate_itemset = copy.deepcopy(current_itemset) if j - 1 < 0: last_itemset = list() else: last_itemset = copy.deepcopy(layer_combination_list[j - 1]) else: if j == prefix_ids[0]: min_count = len(layer_combination_list[j]) + 1 current_itemset.pop(0) candidate_itemset = copy.deepcopy(current_itemset) last_itemset = copy.deepcopy(layer_combination_list[j]) break last_itemset.append(candidate_itemset) layer_combination_list.append(last_itemset) final_combination = layer_combination_list[-1] combination_id = self.get_combination_id(final_combination, layers) return final_combination, combination_id class ModuleGraph(object): """ 更新PaddleGraph，生成代码。 """ def __init__(self, graph): self.pd_graph = graph self.global_layers = graph.get_global_layers() self.codes = list() self.param_tree = PamareterTree() def get_updation_information(self): aprior = Apriori(3) combination_itemset = aprior.run(self.pd_graph) dp = DP(combination_itemset) combination, combination_id = dp.run(self.pd_graph) return combination, combination_id 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 = dict() for column in list(attrs_table.columns): elements = list(attrs_table.get(column)) elements_list = list() count_list = list() for element in elements: if isinstance(element, str) and "'" not in element: break if element not in elements_list: count_list.append(1) elements_list.append(element) else: index = elements_list.index(element) count_list[index] += 1 if len(elements_list) > 1: max_ct = 0 for k, v in zip(elements_list, count_list): if v > max_ct and str(k) != "nan" : max_ele = k max_ct = v diff_attrs_column[column] = max_ele return diff_attrs_column def analyze_graph(self, sub_layers_list): def is_same(sub_layers1, sub_layers2, id1, id2): inputs1, outputs1 = ipt_opt_list[id1] inputs2, outputs2 = ipt_opt_list[id2] if len(inputs1) != len(inputs2) or len(outputs1) != len(outputs2): return False layer_id_list1 = list(sub_layers1.keys()) layer_id_list2 = list(sub_layers2.keys()) for i, layer_id1 in enumerate(layer_id_list1): layer_id2 = layer_id_list2[i] if len(self.pd_graph.edges_in[layer_id1]) != len(self.pd_graph.edges_in[layer_id2]): return False for j, ipt_layer_id1 in enumerate(self.pd_graph.edges_in[layer_id1]): ipt_layer_id2 = self.pd_graph.edges_in[layer_id2][j] if (ipt_layer_id1 in layer_id_list1) ^ (ipt_layer_id2 in layer_id_list2): return False if (layer_id1 in self.pd_graph.edges_out) ^ (layer_id2 in self.pd_graph.edges_out): return False if (layer_id1 in self.pd_graph.edges_out) and (layer_id2 in self.pd_graph.edges_out): if (len(self.pd_graph.edges_out[layer_id1]) > 1 and len(self.pd_graph.edges_out[layer_id2]) == 1) or \ (len(self.pd_graph.edges_out[layer_id1]) == 1 and len(self.pd_graph.edges_out[layer_id2]) > 1): return False for j, opt_layer_id1 in enumerate(self.pd_graph.edges_out[layer_id1]): if len(self.pd_graph.edges_out[layer_id1]) == 1 and len(self.pd_graph.edges_out[layer_id2]) == 1: opt_layer_id2 = self.pd_graph.edges_out[layer_id2][j] if (opt_layer_id1 in layer_id_list1) ^ (opt_layer_id2 in layer_id_list2): return False return True sub_layers_list_list = list() id_list = list() ipt_opt_list = list() sub_layers_list_list.append([sub_layers_list[0]]) id_list.append(0) for i, sub_layer in enumerate(sub_layers_list): ipt_opt_list.append(get_inputs_outputs(self.pd_graph, sub_layer)) if i == 0: continue id_list_cp = copy.deepcopy(id_list) for j, index in enumerate(id_list_cp): if is_same(sub_layers_list[index], sub_layer, index, i): sub_layers_list_list[j].append(sub_layer) break if j == len(id_list_cp) - 1: sub_layers_list_list.append(list()) sub_layers_list_list[j + 1].append(sub_layer) id_list.append(i) return sub_layers_list_list def merge_node(self, sub_layers_list, attrs_table, module_name): sub_layers = sub_layers_list[0] diff_attrs_column = self.analyze_attrs_table(attrs_table) sub_layers, _, _ = rename_layers(sub_layers) code_str = gen_layer_code(self.pd_graph, sub_layers, module_name, different_attrs=diff_attrs_column) self.codes.append(code_str) for index, sub_layers in enumerate(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 mn = module_name.lower() outputs = ["{}_{}".format(mn, index)] + outputs node_name = "{}_{}".format(module_name, index) diff_attrs = dict() for column, element in diff_attrs_column.items(): current_element = attrs_table.get(column).loc[node_name] if current_element != element: diff_attrs[column] = current_element new_layer = PaddleLayer(id=list(sub_layers.keys())[-1], kernel="module", inputs=inputs_dict, outputs=outputs, 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: if len(layer_id.split(".")) > 1: continue self.pd_graph.layers.pop(layer_id) self.pd_graph.build() def convert_subgraph_to_layer(self, combination, combination_id): combination_id_set = set(combination_id) for s in list(combination_id_set): if s == -1: continue module_name = "Block{}".format(s) sub_layers_list = list() for i, c in enumerate(combination): if len(c) > 1 and combination_id[i] == s: sub_layers = dict() for layer_id in c: sub_layers[layer_id] = self.global_layers[layer_id] sub_layers_list.append(sub_layers) sub_layers_list_list = self.analyze_graph(sub_layers_list) for i, sub_layers_list in enumerate(sub_layers_list_list): if i == 0: real_module_name = module_name else: real_module_name = module_name + "__{}".format(i) if len(sub_layers_list) > 1: attrs_table = construct_attrs_table(sub_layers_list, module_name=real_module_name) self.merge_node(sub_layers_list, attrs_table, real_module_name) layers, nn_param_nodes, _ = rename_layers(self.pd_graph.layers, self.param_tree, is_rename_module=True) code_str = gen_layer_code(self.pd_graph, layers, self.pd_graph.name) self.codes.append(code_str) param_node = PamareterNode(old_name="Module") for node in nn_param_nodes: param_node.add_child(node) self.param_tree.add_node(param_node) 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(" # There are {} inputs.".format(len(self.pd_graph.inputs_info))) for k, v in self.pd_graph.inputs_info.items(): run_func_list.append(" # {}: shape-{}, type-{}.".format(k, v[0], v[1])) run_func_list.extend( [" paddle.disable_static()", " params = paddle.load('{}')".format(osp.join(osp.abspath(save_dir), "model.pdparams")), " 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) combination, combination_id = self.get_updation_information() self.convert_subgraph_to_layer(combination, combination_id) self.update_parameters() import_list = ["import paddle", "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()