Day54

知识点回顾：

对抗生成网络的思想：关注损失从何而来
生成器、判别器
nn.sequential容器：适合于按顺序运算的情况，简化前向传播写法
leakyReLU介绍：避免relu的神经元失活现象
ps；如果你学有余力，对于gan的损失函数的理解，建议去找找视频看看，如果只是用，没必要学

作业：对于心脏病数据集，对于病人这个不平衡的样本用GAN来学习并生成病人样本，观察不用GAN和用GAN的F1分数差异。

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset
import numpy as np
import pandas as pd
from sklearn.preprocessing import MinMaxScaler
from sklearn.datasets import load_iris
import matplotlib.pyplot as plt
import warnings
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
import time
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score # 用于评估分类器性能的指标
warnings.filterwarnings("ignore")
# 设置中文字体支持
plt.rcParams["font.family"] = ["SimHei"]
plt.rcParams['axes.unicode_minus'] = False  # 解决负号显示问题
 
# 检查GPU是否可用
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"使用设备: {device}")
 
LATENT_DIM = 10     # 潜在空间的维度，这里根据任务复杂程度任选
EPOCHS = 10000      # 训练的回合数，一般需要比较长的时间
BATCH_SIZE = 32     # 每批次训练的样本数
LR = 0.0002         # 学习率
BETA1 = 0.5         # Adam优化器的参数
 
# 检查是否有可用的GPU，否则使用CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
 
# --- 2. 加载并预处理数据 ---
 
data = pd.read_csv('heart.csv')    #读取数据
X = data.drop(['target'], axis=1)  # 特征，axis=1表示按列删除
y = data['target'] # 标签
# 按照8:2划分训练集和测试集
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)  # 80%训练集，20%测试集
#生病样本165，未生病的只有138
# 只选择未生病人样本(类别 0)
X_patient = X[y == 0]
 
# 数据缩放到 [-1, 1]
scaler = MinMaxScaler(feature_range=(-1, 1)) 
X_scaled = scaler.fit_transform(X_patient) 
 
# 转换为 PyTorch Tensor 并创建 DataLoader
# 注意需要将数据类型转为 float
real_data_tensor = torch.from_numpy(X_scaled).float() 
dataset = TensorDataset(real_data_tensor)
dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True)
 
# (A) 生成器 (Generator)
class Generator(nn.Module):
    def __init__(self):
        super(Generator, self).__init__()
        self.model = nn.Sequential(
            nn.Linear(LATENT_DIM, 16),
            nn.ReLU(),
            nn.Linear(16, 32),
            nn.ReLU(),
            nn.Linear(32, 13),# 最后的维度只要和目标数据对齐即可
            nn.Tanh() # 输出范围是 [-1, 1]
        )
 
    def forward(self, x):
        return self.model(x) # 因为没有像之前一样做定义x=某些东西，所以现在可以直接输出模型
# (B) 判别器 (Discriminator)
class Discriminator(nn.Module):
    def __init__(self):
        super(Discriminator, self).__init__()
        self.model = nn.Sequential(
            nn.Linear(13, 32),
            nn.LeakyReLU(0.2), # LeakyReLU 是 GAN 中的常用选择
            nn.Linear(32, 16),
            nn.LeakyReLU(0.2), # 负斜率参数为0.2
            nn.Linear(16, 1), # 这里最后输出1个神经元,所以用sigmoid激活函数
            nn.Sigmoid() # 输出 0 到 1 的概率
        )
 
    def forward(self, x):
        return self.model(x)
# 实例化模型并移动到指定设备
generator = Generator().to(device)
discriminator = Discriminator().to(device)
 
# --- 4. 定义损失函数和优化器 ---
 
criterion = nn.BCELoss() # 二元交叉熵损失
 
# 分别为生成器和判别器设置优化器
g_optimizer = optim.Adam(generator.parameters(), lr=LR, betas=(BETA1, 0.999))
d_optimizer = optim.Adam(discriminator.parameters(), lr=LR, betas=(BETA1, 0.999))
# --- 5. 执行训练循环 ---
 
print("\n--- 开始训练 ---")
for epoch in range(EPOCHS):
    for i, (real_data,) in enumerate(dataloader):
        # 将数据移动到设备
        real_data = real_data.to(device)
        current_batch_size = real_data.size(0)
 
        # 创建真实和虚假的标签
        real_labels = torch.ones(current_batch_size, 1).to(device)
        fake_labels = torch.zeros(current_batch_size, 1).to(device)
 
        # ---------------------
        #  训练判别器
        # ---------------------
        d_optimizer.zero_grad() # 梯度清零
 
        # (1) 用真实数据训练
        real_output = discriminator(real_data)
        d_loss_real = criterion(real_output, real_labels)
        
        # (2) 用假数据训练
        noise = torch.randn(current_batch_size, LATENT_DIM).to(device)
        # 使用 .detach() 防止在训练判别器时梯度流回生成器，这里我们未来再说
        fake_data = generator(noise).detach() 
        fake_output = discriminator(fake_data)
        d_loss_fake = criterion(fake_output, fake_labels)
        
        # 总损失并反向传播
        d_loss = d_loss_real + d_loss_fake
        d_loss.backward()
        d_optimizer.step()
 
        # ---------------------
        #  训练生成器
        # ---------------------
        g_optimizer.zero_grad() # 梯度清零
 
        # 生成新的假数据，并尝试"欺骗"判别器
        noise = torch.randn(current_batch_size, LATENT_DIM).to(device)
        fake_data = generator(noise)
        fake_output = discriminator(fake_data)
        
        # 计算生成器的损失，目标是让判别器将假数据误判为真(1)
        g_loss = criterion(fake_output, real_labels)
        
        # 反向传播并更新生成器
        g_loss.backward()
        g_optimizer.step()
 
    # 每 1000 个 epoch 打印一次训练状态
    if (epoch + 1) % 1000 == 0:
        print(
            f"Epoch [{epoch+1}/{EPOCHS}], "
            f"Discriminator Loss: {d_loss.item():.4f}, "
            f"Generator Loss: {g_loss.item():.4f}"
        )
 
print("--- 训练完成 ---")

@浙大疏锦行