前言：
这里主要实现： Variational Autoencoders (VAEs) 变分自动编码器
其训练效果如下
训练的过程中要注意调节forward 中的kle ,调参。
整个工程两个文件：
vae.py
main.py
目录：
- vae
- main
一 vae
文件名： vae.py
作用： Variational Autoencoders (VAE)
训练的过程中加入一些限制，使它的latent space规则一点呢。于是就引入了variational autoencoder(VAE)，它被定义为一个有规律地训练以避免过度拟合的Autoencoder，可以确保潜在空间具有良好的属性从而实现内容的生成。
variational autoencoder的架构和Autoencoder差不多，区别在于不再是把输入当作一个点，而是把输入当成一个分布。
# -*- coding: utf-8 -*-
"""
Created on Wed Aug 30 14:19:19 2023
@author: chengxf2
"""
import torch
from torch import nn
#ae: AutoEncoder
class VAE(nn.Module):
def __init__(self,hidden_size=20):
super(VAE, self).__init__()
self.encoder = nn.Sequential(
nn.Linear(in_features=784, out_features=256),
nn.ReLU(),
nn.Linear(in_features=256, out_features=128),
nn.ReLU(),
nn.Linear(in_features=128, out_features=64),
nn.ReLU(),
nn.Linear(in_features=64, out_features=hidden_size),
nn.ReLU()
)
# hidden [batch_size, 10]
h_dim = int(hidden_size/2)
self.hDim = h_dim
self.decoder = nn.Sequential(
nn.Linear(in_features=h_dim, out_features=64),
nn.ReLU(),
nn.Linear(in_features=64, out_features=128),
nn.ReLU(),
nn.Linear(in_features=128, out_features=256),
nn.ReLU(),
nn.Linear(in_features=256, out_features=784),
nn.Sigmoid()
)
def forward(self, x):
'''
param x:[batch, 1,28,28]
return
'''
batchSz= x.size(0)
#flatten
x = x.view(batchSz, 784)
#encoder
h= self.encoder(x)
#在给定维度上对所给张量进行分块,前一半的神经元看作u, 后一般的神经元看作sigma
u, sigma = h.chunk(2,dim=1)
#Reparameterize trick：
#randn_like：产生一个正太分布 ~ N(0,1)
#h.shape [batchSize,self.hDim]
h = u+sigma* torch.randn_like(sigma)
#kld :1e-8 防止sigma 平方为0
kld = 0.5*torch.sum(
torch.pow(u,2)+
torch.pow(sigma,2)-
torch.log(1e-8+torch.pow(sigma,2))-
1
)
#MSE loss 是平均loss, 所以kld 也要算一个平均值
kld = kld/(batchSz*32*32)
xHat = self.decoder(h)
#reshape
xHat = xHat.view(batchSz,1,28,28)
return xHat,kld
二 main
文件名: main.py
作用： 训练，测试数据集
# -*- coding: utf-8 -*-
"""
Created on Wed Aug 30 14:24:10 2023
@author: chengxf2
"""
import torch
from torch.utils.data import DataLoader
from torchvision import transforms, datasets
import time
from torch import optim,nn
from vae import VAE
import visdom
def main():
batchNum = 32
lr = 1e-3
epochs = 20
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
torch.manual_seed(1234)
viz = visdom.Visdom()
viz.line([0],[-1],win='train_loss',opts =dict(title='train acc'))
tf= transforms.Compose([ transforms.ToTensor()])
mnist_train = datasets.MNIST('mnist',True,transform= tf,download=True)
train_data = DataLoader(mnist_train, batch_size=batchNum, shuffle=True)
mnist_test = datasets.MNIST('mnist',False,transform= tf,download=True)
test_data = DataLoader(mnist_test, batch_size=batchNum, shuffle=True)
global_step =0
model =VAE().to(device)
criteon = nn.MSELoss().to(device) #损失函数
optimizer = optim.Adam(model.parameters(),lr=lr) #梯度更新规则
print("\n ----main-----")
for epoch in range(epochs):
start = time.perf_counter()
for step ,(x,y) in enumerate(train_data):
#[b,1,28,28]
x = x.to(device)
x_hat,kld = model(x)
loss = criteon(x_hat, x)
if kld is not None:
elbo = -loss -1.0*kld
loss = -elbo
#backprop
optimizer.zero_grad()
loss.backward()
optimizer.step()
viz.line(Y=[loss.item()],X=[global_step],win='train_loss',update='append')
global_step +=1
end = time.perf_counter()
interval = int(end - start)
print("epoch: %d"%epoch, "\t 训练时间 %d"%interval, '\t 总loss: %4.7f'%loss.item(),"\t KL divergence: %4.7f"%kld.item())
x,target = iter(test_data).next()
x = x.to(device)
with torch.no_grad():
x_hat,kld = model(x)
tip = 'hat'+str(epoch)
viz.images(x,nrow=8, win='x',opts=dict(title='x'))
viz.images(x_hat,nrow=8, win='x_hat',opts=dict(title=tip))
if __name__ == '__main__':
main()
参考：
课时118 变分Auto-Encoder实战-2_哔哩哔哩_bilibili
