前言
本篇主要是结合手写数字例子,结合PyTorch 介绍一下Gan 实战
第一轮训练效果
第20轮训练效果,已经可以生成数字了
68 轮
目录：
- 谷歌云服务器（Google Colab）
- 整体训练流程
- Python 代码
一 谷歌云服务器（Google Colab）
个人用的一直是联想小新笔记本,虽然非常稳定方便。但是现在跑深度学习,性能确实有点跟不上.
1.1 打开谷歌云服务器（Google Colab）
https://colab.research.google.com/
1. 2 新建笔记
1
1.4 选择T4GPU
1.5 点击运行按钮
可以看到当前硬件的情况
二 整体训练流程
三 PyTorch 例子
# -*- coding: utf-8 -*-
"""
Created on Fri Mar 1 13:27:49 2024@author: chengxf2
"""
import torch.optim as optim #优化器
import numpy as np
import matplotlib.pyplot as plt
import torchvision
from torchvision import transforms
import torch
import torch.nn as nn#第一步加载手写数字集
def loadData():#同时归一化数据集(-1,1)style = transforms.Compose([transforms.ToTensor(), #0-1 归一化0-1， channel,height,widthtransforms.Normalize(mean=0.5, std=0.5) #变成了-1,1 ])trainData = torchvision.datasets.MNIST('data',train=True,transform=style,download=True)dataloader = torch.utils.data.DataLoader(trainData,batch_size= 16,shuffle=True)imgs,_ = next(iter(dataloader))#torch.Size([64, 1, 28, 28])print("\n imgs shape ",imgs.shape)return dataloaderclass Generator(nn.Module):'''定义生成器输入：z 随机噪声[batch, input_size]输出：x： 图片 [batch, height, width, channel]'''def __init__(self,input_size):super(Generator,self).__init__()self.net = nn.Sequential(nn.Linear(in_features = input_size , out_features =256),nn.ReLU(),nn.Linear(in_features = 256 , out_features =512),nn.ReLU(),nn.Linear(in_features = 512 , out_features =28*28),nn.Tanh())def forward(self, z):# z 随机输入[batch, dim]x = self.net(z)#[batch, height, width, channel]#print(x.shape)x = x.view(-1,28,28,1)return xclass Discriminator(nn.Module):'''定义鉴别器输入：x： 图片 [batch, height, width, channel]输出：y: 二分类图片的概率： BCELoss 计算交叉熵损失'''def __init__(self):super(Discriminator,self).__init__()#开始的维度和终止的维度，默认值分别是1和-1self.flatten = nn.Flatten()self.net = nn.Sequential(nn.Linear(in_features = 28*28 , out_features =512),nn.LeakyReLU(), #负值的时候保留梯度信息nn.Linear(in_features = 512 , out_features =256),nn.LeakyReLU(),nn.Linear(in_features = 256 , out_features =1),nn.Sigmoid())def forward(self, x):x = self.flatten(x)#print(x.shape)out =self.net(x)return outdef gen_img_plot(model, epoch, test_input):out = model(test_input).detach().cpu()out = out.numpy()imgs = np.squeeze(out)fig = plt.figure(figsize=(4,4))for i in range(out.shape[0]):plt.subplot(4,4,i+1)img = (imgs[i]+1)/2.0#[-1,1]plt.imshow(img)plt.axis('off')plt.show()def train():#1 初始化参数device ='cuda' if torch.cuda.is_available() else 'cpu'#2 加载训练数据dataloader = loadData()test_input = torch.randn(16,100,device=device)#3 超参数maxIter = 20 #最大训练次数input_size = 100batchNum = 16input_size =100#4 初始化模型gen = Generator(100).to(device)dis = Discriminator().to(device)#5 优化器,损失函数d_optim = torch.optim.Adam(dis.parameters(), lr=1e-4)g_optim = torch.optim.Adam(gen.parameters(),lr=1e-4)loss_fn = torch.nn.BCELoss()#6 loss 变化列表D_loss =[]G_loss= []for epoch in range(0,maxIter):d_epoch_loss = 0.0g_epoch_loss =0.0#count = len(dataloader)for step ,(realImgs, _) in enumerate(dataloader):realImgs = realImgs.to(device)random_noise = torch.randn(batchNum, input_size).to(device)#先训练判别器d_optim.zero_grad()real_output = dis(realImgs)d_real_loss = loss_fn(real_output, torch.ones_like(real_output))d_real_loss.backward()#不要训练生成器，所以要生成器detachfake_img = gen(random_noise)fake_output = dis(fake_img.detach())d_fake_loss = loss_fn(fake_output, torch.zeros_like(fake_output))d_fake_loss.backward()d_loss = d_real_loss+d_fake_lossd_optim.step()#优化生成器g_optim.zero_grad()fake_output = dis(fake_img.detach())g_loss = loss_fn(fake_output, torch.ones_like(fake_output))g_loss.backward()g_optim.step()with torch.no_grad():d_epoch_loss+= d_lossg_epoch_loss+= g_losscount = 16 with torch.no_grad():d_epoch_loss/=countg_epoch_loss/=countD_loss.append(d_epoch_loss)G_loss.append(g_epoch_loss)gen_img_plot(gen, epoch, test_input)print("Epoch: ",epoch)print("-----finised-----")if __name__ == "__main__":train()参考：
10.完整课程简介_哔哩哔哩_bilibili
理论【PyTorch][chapter 19][李宏毅深度学习]【无监督学习][ GAN]【理论】-CSDN博客
