观看main函数能够看出框架，框架要简单，比如训练不给它细分，数据流向关注转为哪个数据，而不是关注维度，维度在调试的时候才关注
1、=>表示数据流向
2、# ============================================================================ #包围的表示框架
3、# 表示普通的框架内的注释
4、# -----补充：-----表示补充的注释
5、总体框架以及流向以及说明写在文件最开头
python
"""
基于PSO的BP神经网络预测 - PyTorch版本
框架：
数据准备（设备设置|数据加载）→数据转换→参数设置→训练→结果展示以及预测
数据流向：
# =>加载数据P、T、P_test、T_test、cur_season=>数据转换P_tensor、T_tensor、P_test_tensor、T_test_tensor、cur_season_tensor
# =>particle粒子=>particle粒子、cost、GlobalBest、BestCost=>预测的结果值prob
"""
import numpy as np
import matplotlib.pyplot as plt
from scipy.special import softmax
import time
import torch
import torch.nn as nn
import torch.optim as optim
from bp_function import BpFunction
np.random.seed(42)
torch.manual_seed(42)
# -----补充：粒子类定义-----
class Particle:
def __init__(self, position, velocity, cost, best_position, best_cost):
self.Position = position # 粒子位置
self.Velocity = velocity # 粒子速度
self.Cost = cost # 粒子当前适应度
self.Best = {'Position': best_position, 'Cost': best_cost} # 粒子个体最优
def main():
start_time = time.time()
# ============================================================================ #
# 设备设置------最终数据都需要经过这个设备来进行计算
# ============================================================================ #
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"使用设备: {device}")
# =>加载数据P、T、P_test、T_test、cur_season
# 数据加载
print("加载数据...")
P = np.loadtxt('P_train.txt')
T = np.loadtxt('T_train.txt').reshape(-1, 1)
P_test = np.loadtxt('P_test.txt')
T_test = np.loadtxt('T_test.txt').reshape(-1, 1)
cur_season = np.loadtxt('2022.txt')
# =>加载数据P、T、P_test、T_test、cur_season=>数据转换P_tensor、T_tensor、P_test_tensor、T_test_tensor、cur_season_tensor
# ============================================================================ #
# # 数据转换------转换到PyTorch张量并移动到设备上
# ============================================================================ #
P_tensor = torch.FloatTensor(P).to(device)
T_tensor = torch.FloatTensor(T).to(device)
P_test_tensor = torch.FloatTensor(P_test).to(device)
T_test_tensor = torch.FloatTensor(T_test).to(device)
cur_season_tensor = torch.FloatTensor(cur_season).to(device)
# ============================================================================ #
# # 参数设置------训练参数设置1-BP参数
# ============================================================================ #
inputnum = P.shape[1] # 输入层神经元个数
hiddennum = 2 * inputnum + 1 # 初始隐层神经元个数
outputnum = T.shape[1] # 输出层神经元个数
w1num = inputnum * hiddennum # 输入层到隐层的权值个数
w2num = outputnum * hiddennum # 隐层到输出层的权值个数
N = w1num + hiddennum + w2num + outputnum # 待优化的变量个数
print(f"网络结构: {inputnum} -> {hiddennum} -> {outputnum}")
print(f"待优化变量个数: {N}")
# ============================================================================ #
# # 参数设置------训练参数设置2-PSO参数设置
# ============================================================================ #
nVar = N
VarMin = -0.5
VarMax = 0.5
MaxIt = 200
nPop = 40
w = 1.0
wdamp = 0.99
c1 = 1.5
c2 = 2.0
VelMax = 0.1 * (VarMax - VarMin)
VelMin = -VelMax
# -----补充：粒子变量保存以及最佳解保存-----
particles = []
GlobalBest = {'Position': None, 'Cost': np.inf}
# =>加载数据P、T、P_test、T_test、cur_season=>数据转换P_tensor、T_tensor、P_test_tensor、T_test_tensor、cur_season_tensor
# =>particle粒子（获得第一次BP网络训练之后的结果，得到当前的最佳）
# ============================================================================ #
# 训练
# ============================================================================ #
print("初始化粒子群...")
for i in range(nPop):
position = np.random.uniform(VarMin, VarMax, nVar)
velocity = np.zeros(nVar)
cost, _ = BpFunction(position, P_tensor, T_tensor, hiddennum,
P_test_tensor, T_test_tensor, device)
best_position = position.copy()
best_cost = cost
particle = Particle(position, velocity, cost, best_position, best_cost)
particles.append(particle)
if best_cost < GlobalBest['Cost']:
GlobalBest['Position'] = best_position.copy()
GlobalBest['Cost'] = best_cost
# 初始化变量，之后用来记录最优成本
BestCost = np.zeros(MaxIt)
# =>加载数据P、T、P_test、T_test、cur_season=>数据转换P_tensor、T_tensor、P_test_tensor、T_test_tensor、cur_season_tensor
# =>particle粒子=>particle粒子、cost、GlobalBest、BestCost（PSO训练）
print("开始PSO优化...")
for it in range(MaxIt):
for i in range(nPop):
# 更新速度
r1 = np.random.rand(nVar)
r2 = np.random.rand(nVar)
particles[i].Velocity = (w * particles[i].Velocity +
c1 * r1 * (particles[i].Best['Position'] - particles[i].Position) +
c2 * r2 * (GlobalBest['Position'] - particles[i].Position))
# 速度限制
particles[i].Velocity = np.clip(particles[i].Velocity, VelMin, VelMax)
# 更新位置
particles[i].Position = particles[i].Position + particles[i].Velocity
# 位置边界处理（反弹）
IsOutside = (particles[i].Position < VarMin) | (particles[i].Position > VarMax)
particles[i].Velocity[IsOutside] = -particles[i].Velocity[IsOutside]
particles[i].Position = np.clip(particles[i].Position, VarMin, VarMax)
# 评估适应度
cost, _ = BpFunction(particles[i].Position, P_tensor, T_tensor, hiddennum,
P_test_tensor, T_test_tensor, device)
particles[i].Cost = cost
# 更新个体最优
if cost < particles[i].Best['Cost']:
particles[i].Best['Position'] = particles[i].Position.copy()
particles[i].Best['Cost'] = cost
# 更新全局最优
if cost < GlobalBest['Cost']:
GlobalBest['Position'] = particles[i].Position.copy()
GlobalBest['Cost'] = cost
BestCost[it] = GlobalBest['Cost']
print(f"Iteration {it+1}: Best Cost = {BestCost[it]:.6f}")
w = w * wdamp
# ============================================================================ #
# # 结果展示
# ============================================================================ #
plt.figure(figsize=(10, 6))
plt.semilogy(BestCost, linewidth=2)
plt.xlabel('迭代次数', fontsize=12)
plt.ylabel('误差的变化', fontsize=12)
plt.title('进化过程', fontsize=14)
plt.grid(True)
plt.savefig('evolution_pso.png', dpi=300, bbox_inches='tight')
plt.show()
# =>加载数据P、T、P_test、T_test、cur_season=>数据转换P_tensor、T_tensor、P_test_tensor、T_test_tensor、cur_season_tensor
# =>particle粒子=>particle粒子、cost、GlobalBest、BestCost=>预测的结果值prob
# ============================================================================ #
# # 预测
# ============================================================================ #
print(f"\n最优初始权值和阈值数量: {len(GlobalBest['Position'])}")
print(f"最小误差: {GlobalBest['Cost']:.6f}")
print("\n预测今年总冠军概率...")
_, bestCur_sim = BpFunction(GlobalBest['Position'], P_tensor, T_tensor, hiddennum,
cur_season_tensor, None, device)
prob = softmax(bestCur_sim.flatten())#将预测结果映射为和为1的概率
print(f"勇士队获得2022年NBA总冠军概率为: {prob[0]:.4f}")
print(f"凯尔特人队获得2022年NBA总冠军概率为: {prob[1]:.4f}")
elapsed_time = time.time() - start_time
print(f"\n总耗时: {elapsed_time:.2f} 秒")
if __name__ == "__main__":
main()