462 lines
17 KiB
Python
462 lines
17 KiB
Python
|
#!/usr/bin/env python3
|
|||
|
|
"""
|
|||
|
|
遗传算法参数优化 - 使用DEAP库实现
|
|||
|
|
针对TTA-E神经解码的参数搜索优化
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
import os
|
|||
|
|
import sys
|
|||
|
|
import numpy as np
|
|||
|
|
import pickle
|
|||
|
|
import argparse
|
|||
|
|
import time
|
|||
|
|
import json
|
|||
|
|
from typing import Dict, List, Tuple
|
|||
|
|
import random
|
|||
|
|
|
|||
|
|
# DEAP库
|
|||
|
|
from deap import algorithms, base, creator, tools
|
|||
|
|
|
|||
|
|
# GPU加速支持
|
|||
|
|
try:
|
|||
|
|
import cupy as cp
|
|||
|
|
GPU_AVAILABLE = True
|
|||
|
|
print("GPU acceleration available with CuPy")
|
|||
|
|
except ImportError:
|
|||
|
|
import numpy as cp
|
|||
|
|
GPU_AVAILABLE = False
|
|||
|
|
print("Using CPU computation with NumPy")
|
|||
|
|
|
|||
|
|
# 设置随机种子
|
|||
|
|
random.seed(42)
|
|||
|
|
np.random.seed(42)
|
|||
|
|
|
|||
|
|
def to_cpu(x):
|
|||
|
|
"""将CuPy数组转换为NumPy数组"""
|
|||
|
|
if GPU_AVAILABLE and hasattr(x, 'get'):
|
|||
|
|
return x.get()
|
|||
|
|
return x
|
|||
|
|
|
|||
|
|
def load_base_predictions(cache_file='base_predictions_cache.pkl'):
|
|||
|
|
"""加载预计算的基础预测结果"""
|
|||
|
|
if os.path.exists(cache_file):
|
|||
|
|
print(f"Loading base predictions from {cache_file}")
|
|||
|
|
with open(cache_file, 'rb') as f:
|
|||
|
|
return pickle.load(f)
|
|||
|
|
else:
|
|||
|
|
print(f"Cache file {cache_file} not found. Generating mock data...")
|
|||
|
|
return generate_mock_predictions()
|
|||
|
|
|
|||
|
|
def generate_mock_predictions():
|
|||
|
|
"""生成模拟预测数据用于测试"""
|
|||
|
|
print("Generating mock base predictions for testing...")
|
|||
|
|
|
|||
|
|
n_trials = 20 # 模拟20个试验
|
|||
|
|
seq_len = 40 # 序列长度
|
|||
|
|
vocab_size = 31 # 词汇表大小
|
|||
|
|
n_tta = 5 # TTA样本数
|
|||
|
|
|
|||
|
|
base_predictions = []
|
|||
|
|
|
|||
|
|
np.random.seed(42) # 固定种子获得一致的结果
|
|||
|
|
|
|||
|
|
for trial in range(n_trials):
|
|||
|
|
# 生成真实字符序列
|
|||
|
|
true_chars = np.random.randint(0, vocab_size, seq_len)
|
|||
|
|
|
|||
|
|
# 模拟GRU和LSTM的概率预测 - 加入一些真实性
|
|||
|
|
gru_probs = np.random.rand(n_tta, seq_len, vocab_size)
|
|||
|
|
lstm_probs = np.random.rand(n_tta, seq_len, vocab_size)
|
|||
|
|
|
|||
|
|
# 让概率偏向真实答案，提高一些准确性
|
|||
|
|
for i in range(seq_len):
|
|||
|
|
for tta in range(n_tta):
|
|||
|
|
# 给真实答案更高的概率
|
|||
|
|
gru_probs[tta, i, true_chars[i]] += 2.0
|
|||
|
|
lstm_probs[tta, i, true_chars[i]] += 2.0
|
|||
|
|
|
|||
|
|
# 归一化为概率分布
|
|||
|
|
gru_probs = gru_probs / np.sum(gru_probs, axis=2, keepdims=True)
|
|||
|
|
lstm_probs = lstm_probs / np.sum(lstm_probs, axis=2, keepdims=True)
|
|||
|
|
|
|||
|
|
base_predictions.append({
|
|||
|
|
'gru_probs': gru_probs,
|
|||
|
|
'lstm_probs': lstm_probs,
|
|||
|
|
'true_chars': true_chars
|
|||
|
|
})
|
|||
|
|
|
|||
|
|
return base_predictions
|
|||
|
|
|
|||
|
|
def calculate_per(predicted, true):
|
|||
|
|
"""计算音素错误率 (PER)"""
|
|||
|
|
if len(predicted) == 0 and len(true) == 0:
|
|||
|
|
return 0.0
|
|||
|
|
if len(predicted) == 0 or len(true) == 0:
|
|||
|
|
return 1.0
|
|||
|
|
|
|||
|
|
# 简单的字符级编辑距离
|
|||
|
|
n, m = len(predicted), len(true)
|
|||
|
|
dp = np.zeros((n + 1, m + 1))
|
|||
|
|
|
|||
|
|
for i in range(n + 1):
|
|||
|
|
dp[i][0] = i
|
|||
|
|
for j in range(m + 1):
|
|||
|
|
dp[0][j] = j
|
|||
|
|
|
|||
|
|
for i in range(1, n + 1):
|
|||
|
|
for j in range(1, m + 1):
|
|||
|
|
if predicted[i-1] == true[j-1]:
|
|||
|
|
dp[i][j] = dp[i-1][j-1]
|
|||
|
|
else:
|
|||
|
|
dp[i][j] = 1 + min(dp[i-1][j], dp[i][j-1], dp[i-1][j-1])
|
|||
|
|
|
|||
|
|
return dp[n][m] / max(n, m)
|
|||
|
|
|
|||
|
|
def evaluate_individual(individual, base_predictions_data):
|
|||
|
|
"""评估个体适应度（返回PER错误率，越小越好）"""
|
|||
|
|
try:
|
|||
|
|
# 解析个体参数
|
|||
|
|
tta_weights = individual[:5] # 前5个参数是TTA权重
|
|||
|
|
gru_weight = individual[5] # 第6个参数是GRU权重
|
|||
|
|
|
|||
|
|
# 转换为GPU数组（如果可用）
|
|||
|
|
if GPU_AVAILABLE:
|
|||
|
|
tta_weights = cp.array(tta_weights)
|
|||
|
|
else:
|
|||
|
|
tta_weights = np.array(tta_weights)
|
|||
|
|
|
|||
|
|
total_per = 0.0
|
|||
|
|
total_chars = 0
|
|||
|
|
|
|||
|
|
for trial_data in base_predictions_data:
|
|||
|
|
gru_probs = trial_data['gru_probs'] # shape: (5, seq_len, vocab_size)
|
|||
|
|
lstm_probs = trial_data['lstm_probs'] # shape: (5, seq_len, vocab_size)
|
|||
|
|
true_chars = trial_data['true_chars']
|
|||
|
|
|
|||
|
|
# 转换为GPU数组
|
|||
|
|
if GPU_AVAILABLE:
|
|||
|
|
gru_probs = cp.asarray(gru_probs)
|
|||
|
|
lstm_probs = cp.asarray(lstm_probs)
|
|||
|
|
|
|||
|
|
# 计算TTA加权平均
|
|||
|
|
tta_sum = cp.sum(tta_weights)
|
|||
|
|
if tta_sum > 0:
|
|||
|
|
tta_weights_norm = tta_weights / tta_sum
|
|||
|
|
else:
|
|||
|
|
tta_weights_norm = cp.ones_like(tta_weights) / len(tta_weights)
|
|||
|
|
|
|||
|
|
# 对每个TTA样本加权
|
|||
|
|
gru_weighted = cp.sum(gru_probs * tta_weights_norm[:, None, None], axis=0)
|
|||
|
|
lstm_weighted = cp.sum(lstm_probs * tta_weights_norm[:, None, None], axis=0)
|
|||
|
|
|
|||
|
|
# 模型集成
|
|||
|
|
ensemble_probs = gru_weighted * gru_weight + lstm_weighted * (1 - gru_weight)
|
|||
|
|
|
|||
|
|
# 解码预测
|
|||
|
|
if GPU_AVAILABLE:
|
|||
|
|
predicted_chars = cp.argmax(ensemble_probs, axis=1)
|
|||
|
|
predicted_chars = to_cpu(predicted_chars)
|
|||
|
|
else:
|
|||
|
|
predicted_chars = np.argmax(ensemble_probs, axis=1)
|
|||
|
|
|
|||
|
|
# 计算PER
|
|||
|
|
per = calculate_per(predicted_chars, true_chars)
|
|||
|
|
total_per += per * len(true_chars)
|
|||
|
|
total_chars += len(true_chars)
|
|||
|
|
|
|||
|
|
avg_per = total_per / total_chars if total_chars > 0 else 1.0
|
|||
|
|
return (avg_per,) # DEAP需要返回元组
|
|||
|
|
|
|||
|
|
except Exception as e:
|
|||
|
|
print(f"Error in evaluate_individual: {e}")
|
|||
|
|
return (1.0,) # 返回最差分数
|
|||
|
|
|
|||
|
|
class GeneticOptimizer:
|
|||
|
|
"""基于DEAP的遗传算法优化器"""
|
|||
|
|
|
|||
|
|
def __init__(self,
|
|||
|
|
bounds,
|
|||
|
|
population_size=50,
|
|||
|
|
generations=100,
|
|||
|
|
crossover_prob=0.8,
|
|||
|
|
mutation_prob=0.2,
|
|||
|
|
tournament_size=3,
|
|||
|
|
elite_size=2):
|
|||
|
|
|
|||
|
|
self.bounds = bounds
|
|||
|
|
self.population_size = population_size
|
|||
|
|
self.generations = generations
|
|||
|
|
self.crossover_prob = crossover_prob
|
|||
|
|
self.mutation_prob = mutation_prob
|
|||
|
|
self.tournament_size = tournament_size
|
|||
|
|
self.elite_size = elite_size
|
|||
|
|
|
|||
|
|
# 参数边界
|
|||
|
|
self.param_names = list(bounds.keys())
|
|||
|
|
self.n_params = len(self.param_names)
|
|||
|
|
self.lower_bounds = [bounds[name][0] for name in self.param_names]
|
|||
|
|
self.upper_bounds = [bounds[name][1] for name in self.param_names]
|
|||
|
|
|
|||
|
|
self.setup_deap()
|
|||
|
|
|
|||
|
|
def setup_deap(self):
|
|||
|
|
"""设置DEAP框架"""
|
|||
|
|
# 清理之前的类定义
|
|||
|
|
if hasattr(creator, "FitnessMin"):
|
|||
|
|
del creator.FitnessMin
|
|||
|
|
if hasattr(creator, "Individual"):
|
|||
|
|
del creator.Individual
|
|||
|
|
|
|||
|
|
# 创建适应度类（最小化PER）
|
|||
|
|
creator.create("FitnessMin", base.Fitness, weights=(-1.0,))
|
|||
|
|
creator.create("Individual", list, fitness=creator.FitnessMin)
|
|||
|
|
|
|||
|
|
# 创建工具箱
|
|||
|
|
self.toolbox = base.Toolbox()
|
|||
|
|
|
|||
|
|
# 注册属性生成器 - 确保在范围内
|
|||
|
|
def create_bounded_uniform(low, high):
|
|||
|
|
def bounded_uniform():
|
|||
|
|
return random.uniform(low, high)
|
|||
|
|
return bounded_uniform
|
|||
|
|
|
|||
|
|
for i, (name, (low, high)) in enumerate(zip(self.param_names, zip(self.lower_bounds, self.upper_bounds))):
|
|||
|
|
self.toolbox.register(f"attr_{name}", create_bounded_uniform(low, high))
|
|||
|
|
|
|||
|
|
# 注册个体和种群生成器
|
|||
|
|
self.toolbox.register("individual", tools.initCycle, creator.Individual,
|
|||
|
|
[getattr(self.toolbox, f"attr_{name}") for name in self.param_names], n=1)
|
|||
|
|
self.toolbox.register("population", tools.initRepeat, list, self.toolbox.individual)
|
|||
|
|
|
|||
|
|
# 注册遗传操作
|
|||
|
|
self.toolbox.register("mate", tools.cxBlend, alpha=0.3) # 混合交叉，减小alpha
|
|||
|
|
self.toolbox.register("mutate", self.gaussian_mutation) # 高斯变异
|
|||
|
|
self.toolbox.register("select", tools.selTournament, tournsize=self.tournament_size) # 锦标赛选择
|
|||
|
|
self.toolbox.register("evaluate", self.evaluate_wrapper)
|
|||
|
|
self.toolbox.register("clone", tools.clone)
|
|||
|
|
|
|||
|
|
def evaluate_wrapper(self, individual):
|
|||
|
|
"""评估函数包装器"""
|
|||
|
|
return evaluate_individual(individual, self.base_predictions)
|
|||
|
|
|
|||
|
|
def gaussian_mutation(self, individual):
|
|||
|
|
"""高斯变异操作"""
|
|||
|
|
for i in range(len(individual)):
|
|||
|
|
if random.random() < self.mutation_prob:
|
|||
|
|
# 高斯变异，标准差为参数范围的10%
|
|||
|
|
sigma = (self.upper_bounds[i] - self.lower_bounds[i]) * 0.1
|
|||
|
|
individual[i] += random.gauss(0, sigma)
|
|||
|
|
# 边界处理 - 硬约束
|
|||
|
|
individual[i] = max(self.lower_bounds[i], min(self.upper_bounds[i], individual[i]))
|
|||
|
|
return (individual,)
|
|||
|
|
|
|||
|
|
def constraint_repair(self, individual):
|
|||
|
|
"""约束修复 - 确保所有参数在合理范围内"""
|
|||
|
|
for i in range(len(individual)):
|
|||
|
|
individual[i] = max(self.lower_bounds[i], min(self.upper_bounds[i], individual[i]))
|
|||
|
|
return individual
|
|||
|
|
|
|||
|
|
def optimize(self, base_predictions):
|
|||
|
|
"""运行遗传算法优化"""
|
|||
|
|
self.base_predictions = base_predictions
|
|||
|
|
|
|||
|
|
print(f"Starting Genetic Algorithm Optimization")
|
|||
|
|
print(f"Population size: {self.population_size}")
|
|||
|
|
print(f"Generations: {self.generations}")
|
|||
|
|
print(f"Crossover probability: {self.crossover_prob}")
|
|||
|
|
print(f"Mutation probability: {self.mutation_prob}")
|
|||
|
|
print(f"Tournament size: {self.tournament_size}")
|
|||
|
|
|
|||
|
|
start_time = time.time()
|
|||
|
|
|
|||
|
|
# 创建初始种群
|
|||
|
|
population = self.toolbox.population(n=self.population_size)
|
|||
|
|
|
|||
|
|
# 确保初始种群在边界内
|
|||
|
|
for individual in population:
|
|||
|
|
self.constraint_repair(individual)
|
|||
|
|
|
|||
|
|
# 评估初始种群
|
|||
|
|
fitnesses = list(map(self.toolbox.evaluate, population))
|
|||
|
|
for ind, fit in zip(population, fitnesses):
|
|||
|
|
ind.fitness.values = fit
|
|||
|
|
|
|||
|
|
# 统计信息
|
|||
|
|
stats = tools.Statistics(lambda ind: ind.fitness.values)
|
|||
|
|
stats.register("avg", np.mean)
|
|||
|
|
stats.register("min", np.min)
|
|||
|
|
stats.register("max", np.max)
|
|||
|
|
stats.register("std", np.std)
|
|||
|
|
|
|||
|
|
# 名人堂（保存最佳个体）
|
|||
|
|
hall_of_fame = tools.HallOfFame(maxsize=10)
|
|||
|
|
|
|||
|
|
# 记录历史
|
|||
|
|
history = []
|
|||
|
|
|
|||
|
|
print(f"Initial population evaluated. Best fitness: {min(fitnesses)[0]:.6f}")
|
|||
|
|
|
|||
|
|
# 进化过程
|
|||
|
|
for generation in range(self.generations):
|
|||
|
|
# 选择下一代的父代
|
|||
|
|
offspring = self.toolbox.select(population, len(population))
|
|||
|
|
offspring = list(map(self.toolbox.clone, offspring))
|
|||
|
|
|
|||
|
|
# 交叉和变异
|
|||
|
|
for child1, child2 in zip(offspring[::2], offspring[1::2]):
|
|||
|
|
if random.random() < self.crossover_prob:
|
|||
|
|
self.toolbox.mate(child1, child2)
|
|||
|
|
# 约束修复
|
|||
|
|
self.constraint_repair(child1)
|
|||
|
|
self.constraint_repair(child2)
|
|||
|
|
del child1.fitness.values
|
|||
|
|
del child2.fitness.values
|
|||
|
|
|
|||
|
|
for mutant in offspring:
|
|||
|
|
if random.random() < self.mutation_prob:
|
|||
|
|
self.toolbox.mutate(mutant)
|
|||
|
|
# 约束修复
|
|||
|
|
self.constraint_repair(mutant)
|
|||
|
|
del mutant.fitness.values
|
|||
|
|
|
|||
|
|
# 评估无效个体
|
|||
|
|
invalid_ind = [ind for ind in offspring if not ind.fitness.valid]
|
|||
|
|
fitnesses = map(self.toolbox.evaluate, invalid_ind)
|
|||
|
|
for ind, fit in zip(invalid_ind, fitnesses):
|
|||
|
|
ind.fitness.values = fit
|
|||
|
|
|
|||
|
|
# 精英策略：保留最好的个体
|
|||
|
|
combined = population + offspring
|
|||
|
|
combined.sort(key=lambda x: x.fitness.values[0])
|
|||
|
|
population = combined[:self.population_size]
|
|||
|
|
|
|||
|
|
# 更新统计信息和名人堂
|
|||
|
|
hall_of_fame.update(population)
|
|||
|
|
record = stats.compile(population)
|
|||
|
|
|
|||
|
|
# 记录历史
|
|||
|
|
best_individual = hall_of_fame[0]
|
|||
|
|
best_params = {self.param_names[i]: best_individual[i] for i in range(self.n_params)}
|
|||
|
|
history.append((best_params.copy(), best_individual.fitness.values[0]))
|
|||
|
|
|
|||
|
|
if generation % 10 == 0 or generation == self.generations - 1:
|
|||
|
|
print(f"Generation {generation+1:3d}: "
|
|||
|
|
f"Min={record['min']:.6f}, "
|
|||
|
|
f"Avg={record['avg']:.6f}, "
|
|||
|
|
f"Std={record['std']:.6f}")
|
|||
|
|
|
|||
|
|
elapsed_time = time.time() - start_time
|
|||
|
|
|
|||
|
|
# 返回最佳结果
|
|||
|
|
best_individual = hall_of_fame[0]
|
|||
|
|
best_params = {self.param_names[i]: best_individual[i] for i in range(self.n_params)}
|
|||
|
|
best_score = best_individual.fitness.values[0]
|
|||
|
|
|
|||
|
|
return {
|
|||
|
|
'best_params': best_params,
|
|||
|
|
'best_score': best_score,
|
|||
|
|
'history': history,
|
|||
|
|
'total_evaluations': self.population_size * (1 + self.generations),
|
|||
|
|
'elapsed_time': elapsed_time,
|
|||
|
|
'hall_of_fame': [
|
|||
|
|
{
|
|||
|
|
'params': {self.param_names[i]: ind[i] for i in range(self.n_params)},
|
|||
|
|
'score': ind.fitness.values[0]
|
|||
|
|
}
|
|||
|
|
for ind in hall_of_fame
|
|||
|
|
]
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
def save_results(result, output_file='genetic_optimization_results.json'):
|
|||
|
|
"""保存优化结果"""
|
|||
|
|
print(f"Saving results to {output_file}")
|
|||
|
|
|
|||
|
|
with open(output_file, 'w') as f:
|
|||
|
|
json.dump(result, f, indent=2)
|
|||
|
|
|
|||
|
|
print(f"Results saved successfully")
|
|||
|
|
|
|||
|
|
def main():
|
|||
|
|
parser = argparse.ArgumentParser(description='Genetic Algorithm for TTA-E Parameter Optimization')
|
|||
|
|
parser.add_argument('--cache_file', type=str, default='base_predictions_cache.pkl',
|
|||
|
|
help='Base predictions cache file')
|
|||
|
|
parser.add_argument('--output_file', type=str, default='genetic_optimization_results.json',
|
|||
|
|
help='Output file for results')
|
|||
|
|
|
|||
|
|
# 遗传算法参数
|
|||
|
|
parser.add_argument('--population_size', type=int, default=50,
|
|||
|
|
help='Population size')
|
|||
|
|
parser.add_argument('--generations', type=int, default=100,
|
|||
|
|
help='Number of generations')
|
|||
|
|
parser.add_argument('--crossover_prob', type=float, default=0.8,
|
|||
|
|
help='Crossover probability')
|
|||
|
|
parser.add_argument('--mutation_prob', type=float, default=0.2,
|
|||
|
|
help='Mutation probability')
|
|||
|
|
parser.add_argument('--tournament_size', type=int, default=3,
|
|||
|
|
help='Tournament selection size')
|
|||
|
|
parser.add_argument('--elite_size', type=int, default=2,
|
|||
|
|
help='Elite size for preservation')
|
|||
|
|
|
|||
|
|
args = parser.parse_args()
|
|||
|
|
|
|||
|
|
print(f"{'='*60}")
|
|||
|
|
print("TTA-E Genetic Algorithm Parameter Optimization")
|
|||
|
|
print(f"{'='*60}")
|
|||
|
|
|
|||
|
|
# 加载基础预测数据
|
|||
|
|
base_predictions = load_base_predictions(args.cache_file)
|
|||
|
|
|
|||
|
|
# 定义参数搜索空间
|
|||
|
|
bounds = {
|
|||
|
|
'tta_weight_0': (0.0, 2.0), # 原始样本权重
|
|||
|
|
'tta_weight_1': (0.0, 2.0), # 噪声增强权重
|
|||
|
|
'tta_weight_2': (0.0, 2.0), # 缩放增强权重
|
|||
|
|
'tta_weight_3': (0.0, 2.0), # 偏移增强权重
|
|||
|
|
'tta_weight_4': (0.0, 2.0), # 平滑增强权重
|
|||
|
|
'gru_weight': (0.0, 1.0) # GRU模型权重
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
print(f"Parameter search space:")
|
|||
|
|
for param, (low, high) in bounds.items():
|
|||
|
|
print(f" {param}: [{low}, {high}]")
|
|||
|
|
print()
|
|||
|
|
|
|||
|
|
# 创建遗传算法优化器
|
|||
|
|
optimizer = GeneticOptimizer(
|
|||
|
|
bounds=bounds,
|
|||
|
|
population_size=args.population_size,
|
|||
|
|
generations=args.generations,
|
|||
|
|
crossover_prob=args.crossover_prob,
|
|||
|
|
mutation_prob=args.mutation_prob,
|
|||
|
|
tournament_size=args.tournament_size,
|
|||
|
|
elite_size=args.elite_size
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 运行优化
|
|||
|
|
result = optimizer.optimize(base_predictions)
|
|||
|
|
|
|||
|
|
# 显示结果
|
|||
|
|
print(f"\n{'='*60}")
|
|||
|
|
print("OPTIMIZATION RESULTS")
|
|||
|
|
print(f"{'='*60}")
|
|||
|
|
print(f"Best Score (PER): {result['best_score']:.6f}")
|
|||
|
|
print(f"Best Parameters:")
|
|||
|
|
for param, value in result['best_params'].items():
|
|||
|
|
print(f" {param}: {value:.4f}")
|
|||
|
|
print(f"Total Evaluations: {result['total_evaluations']:,}")
|
|||
|
|
print(f"Elapsed Time: {result['elapsed_time']:.2f} seconds")
|
|||
|
|
print()
|
|||
|
|
|
|||
|
|
print("Top 5 Solutions from Hall of Fame:")
|
|||
|
|
for i, solution in enumerate(result['hall_of_fame'][:5], 1):
|
|||
|
|
print(f"{i}. Score: {solution['score']:.6f}")
|
|||
|
|
params_str = ", ".join([f"{k}={v:.3f}" for k, v in solution['params'].items()])
|
|||
|
|
print(f" Parameters: {params_str}")
|
|||
|
|
|
|||
|
|
# 保存结果
|
|||
|
|
save_results(result, args.output_file)
|
|||
|
|
|
|||
|
|
print(f"\nOptimization completed successfully!")
|
|||
|
|
|
|||
|
|
if __name__ == '__main__':
|
|||
|
|
main()
|