GAN-BP生成数据

单特征生成

代码主体部分如下（不含预处理部分）

import torch
import torch.nn as nn
import numpy as np
import matplotlib.pyplot as plt
import random

BATCH_SIZE = 1
LR_G = 0.0008           # learning rate for generator
LR_D = 0.0004           # learning rate for discriminator
N_IDEAS = 200             # think of this as number of ideas for generating an art work (Generator)
# k = random.randint(0,len(datas)1)
k = 2#第k个样本
ART_COMPONENTS = len(list(datas[k-2]))    # it could be total point G can draw in the canvas
PAINT_POINTS = np.vstack([np.linspace(-1, 1, ART_COMPONENTS) for _ in range(BATCH_SIZE)])


def artist_works(datas):     # painting from the famous artist (real target)
    paintings = []
    painting = list(datas[k-2].T)
    paintings.append(painting)
    paintings = torch.from_numpy(np.array(paintings))
    return paintings

# 生成器
G = nn.Sequential(                      
    nn.Linear(N_IDEAS, 150),           
    nn.ReLU(),
    nn.Linear(150, ART_COMPONENTS),    
)
# 判别器
D = nn.Sequential(                      
    nn.Linear(ART_COMPONENTS, 100),     
    nn.ReLU(),
    nn.Linear(100, 1),
    nn.Sigmoid(),                      
)
#优化器
opt_D = torch.optim.Adam(D.parameters(), lr=LR_D)
opt_G = torch.optim.Adam(G.parameters(), lr=LR_G)

plt.ion() 
p=[]
for i in range(ART_COMPONENTS):
    p.append(i)
for step in range(2000):
    artist_paintings = artist_works(datas)  # 真实数据
    artist_paintings = artist_paintings.to(torch.float32)
    G_ideas = torch.randn(BATCH_SIZE, N_IDEAS) 
    
    #Generator先通过随机点创造图，并把图给Discriminator
    G_paintings = G(G_ideas)                    # 生成假数据
    prob_artist1 = D(G_paintings)               # D试图减少这种问题
    G_loss = torch.mean(torch.log(1. - prob_artist1))  
    opt_G.zero_grad()
    G_loss.backward()
    opt_G.step()
    
    #Discriminator比较真实图和生成图，找出两者的区别并减小这种区别
    prob_artist0 = D(artist_paintings)          # D试图增加这个概率
    prob_artist1 = D(G_paintings.detach())      # D试图减少这种问题
    D_loss = - torch.mean(torch.log(prob_artist0) + torch.log(1. - prob_artist1))
    opt_D.zero_grad()
    D_loss.backward(retain_graph=True) 
    opt_D.step()
    
    if (step+1) % 50 == 0:  # plotting
        scaler = joblib.load("D:/PYthon代码/归一化/scalar")#加载
        G_paintings = scaler.fit_transform(G_paintings.T.detach().numpy())#生成数据
        scaler = joblib.load("D:/PYthon代码/归一化/scalar")#加载
        G_paintings = scaler.inverse_transform(G_paintings)
       
        scaler = joblib.load("D:/PYthon代码/归一化/scalar")#加载
        artist_paintings = artist_paintings.permute(0, 2, 1)
        artist_paintings = artist_paintings[-1,:,:]
        artist_paintings = scaler.inverse_transform(artist_paintings.T)#反归一化 
        
        plt.plot( p[::30], list(G_paintings)[::30], lw = 1,label='Generated painting')
        plt.plot( p[::30], list(artist_paintings[0])[::30], lw = 1, label='data')
        plt.legend()
        plt.ioff()
        plt.show()
        
        pt = []
        for i in G_paintings:
            pt.append(i[0])
        rpt = []
        for i in artist_paintings[0]:
            rpt.append(i)
        g_s_m = pd.Series(pt) #利用Series将列表转换成新的、pandas可处理的数据
        g_a_d = pd.Series(rpt)
        print("第{}次训练".format(step+1))
        print("真数据和合成数据的相关系数：{:.4f}".format(g_s_m.corr(g_a_d)))
        print("真数据概率：{:.4f}%".format(float(prob_artist0)))
        print("合成数据概率：{:.4f}%".format(float(prob_artist1)))