当前位置：首页 > news >正文

基于 TensorFlow 2 的 WGAN来生成表格数据、数值数据和序列数据。 WGAN生成对抗网络。代码仅供参考

news 来源：原创 2025/6/7 13:59:06

在这里插入图片描述
WGAN生成对抗网络
基于python tensorflow2以下框架，环境应该如何自己配置？
表格数据、数值数据，序列数据生成
里面是一个标准函数的数据，读取自己数据需要自己改。可以随意替换纬度，需要自己手动改. 在这里插入图片描述

如何配置环境并编写一个基于 TensorFlow 2 的 WGAN（Wasserstein Generative Adversarial Network）来生成表格数据、数值数据和序列数据。我们将使用标准函数，并提供一个模板，你可以根据自己的数据进行调整。

代码仅供参考

环境配置

首先，确保你的 Python 环境已经安装了必要的库。你可以使用以下命令来安装所需的库：

pip install tensorflow numpy pandas matplotlib scikit-learn

代码实现

我们将编写一个通用的 WGAN 模型，可以处理不同类型的数值数据。这个模型将包括以下几个部分：

数据加载和预处理
构建 WGAN 模型
训练 WGAN 模型
生成数据

1. 数据加载和预处理

我们将编写一个函数来加载和预处理数据。这里我们假设你有一个 CSV 文件作为输入数据。

import os
import numpy as np
import pandas as pd
from sklearn.preprocessing import StandardScalerdef load_and_preprocess_data(file_path):# Load data from CSV filedata = pd.read_csv(file_path)# Separate features and labels if neededX = data.values# Normalize the datascaler = StandardScaler()X_scaled = scaler.fit_transform(X)return X_scaled, scaler# Example usage
data_path = 'your_data.csv'
X_scaled, scaler = load_and_preprocess_data(data_path)

2. 构建 WGAN 模型

我们将定义生成器和判别器网络。

import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense, Reshape, Flatten
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.constraints import Constraintclass ClipConstraint(Constraint):def __init__(self, clip_value):self.clip_value = clip_valuedef __call__(self, weights):return tf.clip_by_value(weights, -self.clip_value, self.clip_value)def build_generator(latent_dim, output_shape):model = Sequential()model.add(Dense(128, activation='relu', input_dim=latent_dim))model.add(Dense(256, activation='relu'))model.add(Dense(512, activation='relu'))model.add(Dense(output_shape, activation='tanh'))return modeldef build_discriminator(input_shape):model = Sequential()model.add(Flatten(input_shape=input_shape))model.add(Dense(512, activation='relu', kernel_constraint=ClipConstraint(0.01)))model.add(Dense(256, activation='relu', kernel_constraint=ClipConstraint(0.01)))model.add(Dense(1))return modeldef wasserstein_loss(y_true, y_pred):return tf.reduce_mean(y_true * y_pred)def gradient_penalty_loss(y_true, y_pred, averaged_samples, weight):gradients = tf.gradients(y_pred, averaged_samples)[0]gradients_sqr = tf.square(gradients)gradient_penalty = tf.reduce_mean(tf.reduce_sum(gradients_sqr, axis=np.arange(1, len(gradients_sqr.shape))))return weight * gradient_penalty# Define parameters
latent_dim = 100
output_shape = X_scaled.shape[1]generator = build_generator(latent_dim, output_shape)
discriminator = build_discriminator((output_shape,))discriminator.compile(loss=wasserstein_loss, optimizer=Adam(lr=0.0001, beta_1=0.5), metrics=['accuracy'])discriminator.trainable = Falsegan_input = Input(shape=(latent_dim,))
generated_signal = generator(gan_input)
validity = discriminator(generated_signal)combined = Model(gan_input, validity)
combined.compile(loss=wasserstein_loss, optimizer=Adam(lr=0.0001, beta_1=0.5))

3. 训练 WGAN 模型

我们将编写训练循环来训练 WGAN 模型。

def train(generator, discriminator, combined, X_train, latent_dim, epochs=10000, batch_size=64, sample_interval=1000):valid = -np.ones((batch_size, 1))fake = np.ones((batch_size, 1))g_loss_list = []d_loss_list = []for epoch in range(epochs):idx = np.random.randint(0, X_train.shape[0], batch_size)real_signals = X_train[idx]noise = np.random.normal(0, 1, (batch_size, latent_dim))generated_signals = generator.predict(noise)d_loss_real = discriminator.train_on_batch(real_signals, valid)d_loss_fake = discriminator.train_on_batch(generated_signals, fake)alpha = np.random.random(size=(batch_size, 1, X_train.shape[1]))interpolated_signals = (alpha * real_signals) + ((1 - alpha) * generated_signals)validity_interpolated = discriminator.predict(interpolated_signals)gp_loss = gradient_penalty_loss(None, validity_interpolated, interpolated_signals, 10)d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)[0] + gp_lossnoise = np.random.normal(0, 1, (batch_size, latent_dim))g_loss = combined.train_on_batch(noise, valid)if epoch % sample_interval == 0:print(f"{epoch} [D loss: {d_loss}] [G loss: {g_loss}]")g_loss_list.append(g_loss)d_loss_list.append(d_loss)return g_loss_list, d_loss_list# Train the model
g_loss_list, d_loss_list = train(generator, discriminator, combined, X_scaled, latent_dim)

4. 生成数据

我们将编写一个函数来生成新的数据样本。

def generate_data(generator, latent_dim, num_samples, scaler):noise = np.random.normal(0, 1, (num_samples, latent_dim))generated_signals = generator.predict(noise)generated_signals_rescaled = scaler.inverse_transform(generated_signals)return generated_signals_rescaled# Generate new data samples
num_samples = 1000
generated_data = generate_data(generator, latent_dim, num_samples, scaler)# Save generated data to a CSV file
pd.DataFrame(generated_data).to_csv('generated_data.csv', index=False)