PyTorch 深度学习 || 专题九：PyTorch 全连接自编码网络的无监督学习

PyTorch 全连接自编码网络的无监督学习

深度学习网络用于无监督学习是通过一种称为 自编码(Auto-Encoder, AE)的网络结构实现的。自编码神经网络是一种以重构输入信号为目标的神经网络，它是无监督学习领域中的一类网络，可以自动从无表注的数据中学习特征。

自编码由3个网络层：输入层、隐藏层和输出层。其中输入层的样本也会充当输出层的角色，即这个神经网络就是一个尽可能复线输入层信号的神经网络，具体结构如下

1. 数据去噪

1.1 计算库和数据准备工作

import torch
import torch.nn as nn
import numpy as np
import matplotlib.pyplot as plt

from sklearn.datasets import make_moons

SEED = 2617
N = 500
SIG = 0.1
X, _ = make_moons(n_samples=N, noise=SIG, random_state=SEED)
plt.scatter(X[:,0],X[:,1])
plt.axis('equal');

XT = torch.tensor(X, dtype=torch.float32)

1.2 构建自编码网络

class NetworkAE(nn.Module):
    
    def __init__(self, d_in, d_enc, d1, d2):
        super().__init__()
        # YOUR CODE HERE
        self.encoder=nn.Sequential(
            nn.Linear(d_in,d1),
            nn.Tanh(),
            nn.Linear(d1,d2),
            nn.ReLU(), 
            nn.Linear(d2,d1), 
            nn.ReLU(), 
            nn.Linear(d1,d_enc)
        )
        
        self.decoder = nn.Sequential(
            nn.Linear(d_enc, d1),
            nn.ReLU(),
            nn.Linear(d1, d2),
            nn.ReLU(),
            nn.Linear(d2, d1),
            nn.Tanh(), 
            nn.Linear(d1, d_in)
        )
        #raise NotImplementedError()
    
    def forward(self, x):
        # YOUR CODE HERE
        enc = self.encoder(x)
        x = self.decoder(enc)
        return x

训练网络

def train_nn(NN, X, learning_rate, max_iter, print_iter):
    # YOUR CODE HERE
    loss_func = nn.MSELoss(reduction='sum')        # loss function，此函数默认均值对结果影响非常大，暂时不知道为什么
    optimizer = torch.optim.SGD(NN.parameters(), lr=learning_rate)
    
    for epoch in range(max_iter):
        # ===================forward=====================
        output = NN(X)
        loss = loss_func(output, X)
        # ===================backward====================
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        # ===================log========================
        
        if epoch % print_iter == 0:
            print('epoch [{}/{}], loss:{:.4f}'.format(epoch + 1, max_iter, loss.item()))

1.3 调用主函数

learning_rate = 5e-4
max_iter = 10000
d1 = 10  
d2 = 25
netenc = NetworkAE(2, 1,d1,d2)
train_nn(netenc, XT, learning_rate, max_iter,1000)

X_NET = netenc(XT).detach().numpy()

1.4 可视化

plt.scatter(X[:,0], X[:,1])
plt.scatter(X_NET[:,0], X_NET[:,1])
plt.axis('equal');

for i in range(1000):
    X_NET = torch.tensor(X_NET, dtype=torch.float32)
    X_NET = netenc(X_NET).detach().numpy()

plt.scatter(X[:,0], X[:,1])
plt.scatter(X_NET[:,0], X_NET[:,1])
plt.axis('equal');

2. 数据的重建与降维

2.1 计算模块和数据的准备

import numpy as np 
import pandas as pd 
import matplotlib.pyplot as plt 
from mpl_toolkits.mplot3d import Axes3D # 三维数据可视化
import hiddenlayer as hl 

import torch
import torch.nn as nn
import torch.nn.functional as F 
import torch.utils.data as Data 
import torch.optim as optim
from torchvision import transforms
from torchvision.datasets import MNIST
from torchvision.utils import make_grid

# 模型加载选择GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# device = torch.device('cpu')
print(device)
print(torch.cuda.device_count())
print(torch.cuda.get_device_name(0))

cuda
1
NVIDIA GeForce RTX 2080 SUPER

2.2 自编码网络数据准备

# 使用手写体数据，准备训练数据集
train_data = MNIST(
    root = './data/MNIST',
    train = True, # 只使用训练数据集
    transform = transforms.ToTensor(),
    download = False
)

# 将图像数据转为向量数据
train_data_x = train_data.data.type(torch.FloatTensor) / 255.0
train_data_x = train_data_x.reshape(train_data_x.shape[0], -1)
train_data_y = train_data.targets
# 定义一个数据加载器
train_loader = Data.DataLoader(
    dataset = train_data_x,
    batch_size = 64,
    shuffle = True,
    num_workers = 2 # Windows需要设置为0
)

# 对测试数据集导入
test_data = MNIST(
    root = './data/MNIST',
    train = False, # 不使用训练数据集
    transform = transforms.ToTensor(),
    download = False
)

# 为测试数据添加一个通道维度，获取测试数据的X和Y
test_data_x = test_data.data.type(torch.FloatTensor) / 255.0
test_data_x = test_data_x.reshape(test_data_x.shape[0], -1)
test_data_y = test_data.targets
print('训练数据集:', train_data_x.shape)
print('测试数据集:', test_data_x.shape)

数据的显示

# 可视化训练数据集中一个batch的图像内容，以观察手写体图像的情况，程序如下：
for step, b_x in enumerate(train_loader):
    if step > 0:
        break
# 可视化一个batch
# make_grid将[batch, channel, height, width]形式的batch图像转为图像矩阵，便于对多张图像的可视化
im = make_grid(b_x.reshape((-1, 1, 28, 28)))
im = im.data.numpy().transpose((1, 2, 0))
plt.figure()
plt.imshow(im)
plt.axis('off')
plt.show()

2.3 自编码网络的构建

class EnDecoder(nn.Module):
    def __init__(self):
        super(EnDecoder, self).__init__()
        # 定义Encoder
        self.Encoder = nn.Sequential(
            nn.Linear(784, 512),
            nn.Tanh(),
            nn.Linear(512, 256),
            nn.Tanh(),
            nn.Linear(256, 128),
            nn.Tanh(),
            nn.Linear(128, 3),
            nn.Tanh(),
        )
        # 定义Decoder
        self.Decoder = nn.Sequential(
            nn.Linear(3, 128),
            nn.Tanh(),
            nn.Linear(128, 256),
            nn.Tanh(),
            nn.Linear(256, 512),
            nn.Tanh(),
            nn.Linear(512, 784),
        )
    
    def forward(self, x):
        encoder = self.Encoder(x)
        decoder = self.Decoder(encoder)
        return encoder, decoder

# 定义自编码网络edmodel
myedmodel = EnDecoder().to(device)

from torchsummary import summary

summary(myedmodel, input_size=(1, 784))

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Linear-1               [-1, 1, 512]         401,920
              Tanh-2               [-1, 1, 512]               0
            Linear-3               [-1, 1, 256]         131,328
              Tanh-4               [-1, 1, 256]               0
            Linear-5               [-1, 1, 128]          32,896
              Tanh-6               [-1, 1, 128]               0
            Linear-7                 [-1, 1, 3]             387
              Tanh-8                 [-1, 1, 3]               0
            Linear-9               [-1, 1, 128]             512
             Tanh-10               [-1, 1, 128]               0
           Linear-11               [-1, 1, 256]          33,024
             Tanh-12               [-1, 1, 256]               0
           Linear-13               [-1, 1, 512]         131,584
             Tanh-14               [-1, 1, 512]               0
           Linear-15               [-1, 1, 784]         402,192
          Sigmoid-16               [-1, 1, 784]               0
================================================================
Total params: 1,133,843
Trainable params: 1,133,843
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.04
Params size (MB): 4.33
Estimated Total Size (MB): 4.37
----------------------------------------------------------------

2.4 自编码网络的训练

optimizer = optim.Adam(myedmodel.parameters(), lr = 0.003)
loss_func = nn.MSELoss().to(device)
# 记录训练过程的指标
historyl = hl.History()
# 使用Canvas进行可视化
canvasl = hl.Canvas()
train_num = 0
val_num = 0
# 对模型迭代训练
for epoch in range(10):
    train_loss_epoch = 0
    # 对训练数据的加载器进行迭代计算
    for step, b_x in enumerate(train_loader):
        # 使用每个batch进行训练模型
        b_x = b_x.to(device)
        _, output = myedmodel(b_x) # 前向传递输出
        loss = loss_func(output, b_x) # 均方根误差
        optimizer.zero_grad() # 每个迭代步的梯度初始化为0
        loss.backward() # 损失的后向传播，计算梯度
        optimizer.step() # 使用梯度进行优化
        train_loss_epoch += loss.item() * b_x.size(0)
        train_num = train_num + b_x.size(0)
    # 计算一个epoch的损失
    train_loss = train_loss_epoch / train_num
    # 保存每个epoch上的输出loss
    historyl.log(epoch, train_loss = train_loss)
    # 可视网络训练的过程
    with canvasl:
        canvasl.draw_plot(historyl['train_loss'])

2.5 自编码网络的数据重建

# 预测测试集前100张图像的输出
myedmodel.eval()
_, test_decoder = myedmodel(test_data_x[0: 100, :].to(device))
# 可视化原来的图像
plt.figure(figsize = (6, 6))
for ii in range(test_decoder.shape[0]):
    plt.subplot(10, 10, ii + 1)
    im = test_data_x[ii, :]
    im = im.cpu().data.numpy().reshape(28, 28)
    plt.imshow(im, cmap = plt.cm.gray)
    plt.axis('off')
plt.show()
# 可视化编码后的图像
plt.figure(figsize = (6, 6))
for ii in range(test_decoder.shape[0]):
    plt.subplot(10, 10, ii + 1)
    im  = test_decoder[ii, :]
    im = im.cpu().data.numpy().reshape(28, 28)
    plt.imshow(im, cmap = plt.cm.gray)
    plt.axis('off')
plt.show()

2.6 自编码网络的编码特征可视化

# 获取前500个样本的自编码后的特征，并对数据进行可视化
myedmodel.eval()
TEST_num = 500
test_encoder, _ = myedmodel(test_data_x[0: TEST_num, :].to(device))
print('test_encoder.shape:', test_encoder.shape)

%config InlineBackend.print_figure_kwargs = {'bbox_inches': None}
# 将3个维度的特征进行可视化
test_encoder_arr = test_encoder.cpu().data.numpy()
fig = plt.figure(figsize = (12, 8))
ax1 = Axes3D(fig)
X = test_encoder_arr[:, 0]
Y = test_encoder_arr[:, 1]
Z = test_encoder_arr[:, 2]
ax1.set_xlim([min(X), max(X)])
ax1.set_ylim([min(Y), max(Y)])
ax1.set_zlim([min(Z), max(Z)])
for ii in range(test_encoder.shape[0]):
    text = test_data_y.data.numpy()[ii]
    ax1.text(X[ii], Y[ii], Z[ii], str(text), fontsize = 8, bbox = dict(boxstyle = 'round', facecolor = plt.cm.Set1(text), alpha = 0.7))
plt.show()