import os
import torch
import as nn
import as F
import torchvision
from torchvision import transforms
from import save_image
# 配置GPU或CPU设置
device = ('cuda' if .is_available() else 'cpu')
# 创建目录保存生成的图片
sample_dir = 'samples'
if not (sample_dir):
(sample_dir)
# 超参数设置
image_size = 784 #图片大小
h_dim = 400
z_dim = 20
num_epochs = 15 #15个循环
batch_size = 128 #一批的数量
learning_rate = 1e-3 #学习率
# 获取数据集
dataset = (root='./data',
train=True,
transform=(),
download=True)
# 数据加载，按照batch_size大小加载，并随机打乱
data_loader = (dataset=dataset,
batch_size=batch_size,
shuffle=True)
# VAE模型
class VAE():
def __init__(self, image_size=784, h_dim=400, z_dim=20):
super(VAE, self).__init__()
self.fc1 = (image_size, h_dim)
self.fc2 = (h_dim, z_dim)
self.fc3 = (h_dim, z_dim)
self.fc4 = (z_dim, h_dim)
self.fc5 = (h_dim, image_size)
# 编码，学习高斯分布均值与方差
def encode(self, x):
h = (self.fc1(x))
return self.fc2(h), self.fc3(h)
# 将高斯分布均值与方差参数重表示，生成隐变量z 若x~N(mu, var*var)分布,则(x-mu)/var=z~N(0, 1)分布
def reparameterize(self, mu, log_var):
std = (log_var / 2)
eps = torch.randn_like(std)
return mu + eps * std
# 解码隐变量z
def decode(self, z):
h = (self.fc4(z))
return (self.fc5(h))
# 计算重构值和隐变量z的分布参数
def forward(self, x):
mu, log_var = (x) # 从原始样本x中学习隐变量z的分布，即学习服从高斯分布均值与方差
z = (mu, log_var) # 将高斯分布均值与方差参数重表示，生成隐变量z
x_reconst = (z) # 解码隐变量z，生成重构x’
return x_reconst, mu, log_var # 返回重构值和隐变量的分布参数
# 构造VAE实例对象
model = VAE().to(device)
print(model)
"""VAE(
(fc1): Linear(in_features=784, out_features=400, bias=True)
(fc2): Linear(in_features=400, out_features=20, bias=True)
(fc3): Linear(in_features=400, out_features=20, bias=True)
(fc4): Linear(in_features=20, out_features=400, bias=True)
(fc5): Linear(in_features=400, out_features=784, bias=True)
)"""
# 选择优化器，并传入VAE模型参数和学习率
optimizer = ((), lr=learning_rate)
# 开始训练一共15个循环
for epoch in range(num_epochs):
for i, (x, _) in enumerate(data_loader):
# 前向传播
x = (device).view(-1,image_size) # 将batch_size*1*28*28 ---->batch_size*image_size 其中，image_size=1*28*28=784
x_reconst, mu, log_var = model(x) # 将batch_size*748的x输入模型进行前向传播计算,重构值和服从高斯分布的隐变量z的分布参数（均值和方差）
# 计算重构损失和KL散度
# 重构损失
reconst_loss = F.binary_cross_entropy(x_reconst, x, size_average=False)
# KL散度
kl_div = - 0.5 * torch.sum(1 + log_var - mu.pow(2) - log_var.exp())
# 反向传播与优化
# 计算误差(重构误差和KL散度值)
loss = reconst_loss + kl_div
# 清空上一步的残余更新参数值
optimizer.zero_grad()
# 误差反向传播, 计算参数更新值
()
# 将参数更新值施加到VAE model的parameters上
()
# 每迭代一定步骤，打印结果值
if (i + 1) % 10 == 0:
print("Epoch[{}/{}], Step [{}/{}], Reconst Loss: {:.4f}, KL Div: {:.4f}"
.format(epoch + 1, num_epochs, i + 1, len(data_loader), reconst_loss.item(), kl_div.item()))
with torch.no_grad():
# 保存采样值
# 生成随机数 z
z = (batch_size, z_dim).to(device) # z的大小为batch_size * z_dim = 128*20
# 对随机数 z 进行解码decode输出
out = (z).view(-1, 1, 28, 28)
# 保存结果值
save_image(out, (sample_dir, 'sampled-{}.png'.format(epoch + 1)))
# 保存重构值
# 将batch_size*748的x输入模型进行前向传播计算，获取重构值out
out, _, _ = model(x)
# 将输入与输出拼接在一起输出保存 batch_size*1*28*（28+28）=batch_size*1*28*56
x_concat = ([(-1, 1, 28, 28), (-1, 1, 28, 28)], dim=3)
save_image(x_concat, (sample_dir, 'reconst-{}.png'.format(epoch + 1)))