TVM中遗传算法的研究与改进_Improvement-of-GA-in-TVM.zip

[TOC] # tvm.autotvm.MYGA 包含改进遗传算法的各类参数的枚举类及对应工具类，包括selection,crossover,mutation,aga,sa 包含配置类，用以设置遗传算法的各类参数 包含ga，用以进行遗传算法的迭代 ## tvm.autotvm.MYGA.selection 包含枚举类EnumSelection和工具类GASelection ### EnumSelection 继承Enum，包含已配置选择算子种类，并可返回一随机选择算子 ```python class EnumSelection(Enum): dup = 0 ecs = 1 etour = 2 otos = 3 rcs = 4 rps = 5 rws = 6 sus = 7 tour = 8 urs = 9 crws = 10 @staticmethod def randomSelection(): l = len(EnumSelection) idx = np.random.randint(l) selection = EnumSelection(idx) return selection ``` ### GASeletion 提供具体选择算子函数，包括dup，ecs等11种选择算子函数 如 ```python @staticmethod def dup(genes, scores, num=2): tmp_scores = np.copy(scores)[:, np.newaxis] p1 ,p2 = geatpy.dup(tmp_scores, num) p1 %= len(genes) p2 %= len(genes) return p1, p2 ``` ## tvm.autotvm.MYGA.crossover 包含枚举类EnumCrossover和工具类GACrossover ### EnumCrossover 继承Enum，包含已配置交叉算子种类，并可返回一随机交叉算子 ```python class EnumCrossover(Enum): spc = 0 # single point cross tpc = 1 # two point cross mpc = 2 # multipoint cross ec = 3 # even cross etpc = 4 # Even two-point cross dc = 5 # Discrete cross ac = 6 # Arithmetic cross hc = 7 # Heuristic cross @staticmethod def randomCrossover(): l = len(EnumCrossover) idx = np.random.randint(l) crossover = EnumCrossover(idx) return crossover ``` ### GACrossover 提供具体交叉算子函数，包括spc，tpc等8种交叉算子函数 如 ```python @staticmethod def spc(g1, g2, size): if size >= 1: point = np.random.randint(size) tmp_gene = g1[:point] + g2[point:] return tmp_gene return g1 ``` ## tvm.autotvm.MYGA.mutation 包含枚举类EnumMutation和工具类GAMutation ### EnumMutation 继承Enum，包含已配置变异算子种类，并可返回一随机变异算子 ```python class EnumMutation(Enum): rm = 0 # random mutation cm = 1 # crossover mutation sm = 2 # swap mutation im = 3 # inversion mutation Rm = 4 @staticmethod def randomMutation(): l = len(EnumMutation) idx = np.random.randint(l) mutation = EnumMutation(idx) return mutation ``` ### GAMutation 提供具体变异算子函数，包括rm，cm等5种变异算子函数 如 ```python @staticmethod def rm(gene, dims, rate): for j, dim in enumerate(dims): if np.random.random() < rate: gene[j] = np.random.randint(dim) return gene ``` ## tvm.autotvm.MYGA.aga 增加自适应遗传算法，根据适应度动态调整个体交叉率和变异率，包含枚举类EnumAGA和工具类AGA ### EnumAGA 继承Enum，包含已配置AGA算子种类，并可返回一随机AGA算子 ```python class EnumAGA(Enum): GA = 0 # normal Genetic Algorithm AGA = 1 # Adaptive Genetic Algorithm LAGA = 2 # Linear Adaptive Genetic Algorithm CAGA = 3 # Cosine Adaptive Genetic Algorithm IAGA = 4 # Improved Adaptive Genetic Algorithm @staticmethod def randomAGA(): l = len(EnumAGA) idx = np.random.randint(l) aga = EnumAGA(idx) return aga ``` ### AGA 提供具体AGA算子函数，包括GA，AGA等5种AGA算子函数 如 ```python @staticmethod def AGA(k1, k2, f_max, f_avg, f, target): if f < f_avg: return k2 else: return k1 * (f_max - f) / (f_max - f_avg) ``` ## tvm.autotvm.MYGA.sa 增加模拟退火算法（SA），搭配GA使用，包含枚举类EnumAGA，具体算法见[tvm.autotvm.MYGA.ga.](#tvm.autotvm.MYGA.ga) ### EnumSA 继承Enum，包含已配置SA算子种类，并可返回一随机SA算子 ```python class EnumSA(Enum): GA = 0 GA_SA1 = 1 GA_SA2 = 2 @staticmethod def randomSA(): l = len(EnumSA) idx = np.random.randint(l) sa = EnumSA(idx) return sa ``` ## tvm.autotvm.MYGA.config 包含GA各参数对应配置类，包括SelectionConfig，CrossoverConfig，MutationConfig，ExtraConfig，PopulationConfig，MultipopulationConfig ### SelectionConfig 设置选择算子，其类型必须是EnumSelection或其name对应的str ### CrossoverConfig 设置交叉算子，其类型必须是EnumCrossover或其name对应的str 设置交叉率及最大最小值 ### MutationConfig 设置变异算子，其类型必须是EnumMutation或其name对应的str 设置变异率及最大最小值 ### ExtraConfig 设置AGA，其类型必须是EnumAGA或其name对应的str 设置SA，其类型必须是EnumSA或其name对应的str ### PopulationConfig 设置种群参数，包括种群规模，精英数量，SelectionConfig，CrossoverConfig，MutationConfig，ExtraConfig 输入为dict类型，例 ```python config = { "pop_size": 200, "elite_num": 3, "selection": { "op": "dup" # "dup" also can be EnumSelection.dup }, "crossover": { "op": "tpc", # "tpc" also can be EnumCrossover.tpc "rate": 0.6 }, "mutation": { "op": "im", # "im" also can be EnumMutation.im "rate": 0.001 } } ``` ### MultipopulationConfig 设置多种群参数，输入为dict类型，包含各种群对应参数，例 ```python config = { "config1" : { "pop_size": 100, "elite_num": 3, "selection": { "op": "dup" }, "crossover": { "op": "tpc", "rate": 0.6 }, "mutation": { "op": "im2", "rate": 0.02 } }, "config2": { "pop_size": 200, "elite_num": 4, "selection": { "op": "rws" }, "crossover": { "op": "spc", "rate": 0.6 }, "mutation": { "op": "rm", "rate": 0.001 } }, } ``` ## tvm.autotvm.MYGA.ga 用以进行遗传算法的具体计算，包括类Solution，Population，Multipopulation ### Solution 种群个体，包括基因及适应度 ```python class Solution: def __init__(self, gene, score=0.0): self.gene = gene self.score = score ``` ### Population 核心类，单种群GA，下面具体介绍其函数 #### Population.\_\_init\_\_ ```python def __init__(self, tunerConfig, popConfig : PopulationConfig): self.tunerConfig = tunerConfig # tuner config,don't need to care self.popConfig = popConfig # pop config,set selection op and other op self.space = len(self.tunerConfig.space) # get search space size self.popConfig.setSize(min(self.popConfig.getSize(), self.space)) # pop size can't exceed the search space size self.solutions = [] # pop individuals self.elites = [] # pop elites self.genes = [] # individual's gene self.scores = [] # individual's score self.tmp_solution = [] self.next_solution = [] self.next_batch = [] self.next_id = -1 # index of next individual ``` 其中tunerConfig为搜索空间参数，tuner自动生成，无需关注 popConfig为种群配置类，包括该种群的各具体参数 #### Popolation.init 产生足够的随机 ```python def init(self): # random initialization for _ in range(self.popConfig.getSize()): tmp_gene = point2knob(np.random.randint(self.space), self.tunerConfig.dims) tmp_solution = Solution(tmp_gene) self.solutions.append(tmp_solution) ``` #### Population.update ```python def update(self): # check whether to