【Pytorch】循环神经网络实现手写体识别

1 数据集加载

import seaborn as sns
sns.set(font_scale=1.5,style="white") 
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import time
import copy

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

---

## 准备训练数据集Minist
train_data  = torchvision.datasets.MNIST(root='./data',  
                            train=True,   
                            transform=transforms.ToTensor(),  
                            download=True) 
## 定义一个数据加载器
train_loader = Data.DataLoader(
    dataset = train_data, ## 使用的数据集
    batch_size=64, # 批处理样本大小
    shuffle = True, # 每次迭代前打乱数据
    num_workers = 2, # 使用两个进程 
)


##  可视化训练数据集的一个batch的样本来查看图像内容
for step, (b_x, b_y) in enumerate(train_loader):  
    if step > 0:
        break
## 输出训练图像的尺寸和标签的尺寸，都是torch格式的数据
print(b_x.shape)
print(b_y.shape)
train_data

输出

torch.Size([64, 1, 28, 28])
torch.Size([64])
Dataset MNIST
    Number of datapoints: 60000
    Root location: ./data
    Split: Train
    StandardTransform
Transform: ToTensor()

---

## 准备需要使用的测试数据集
test_data  = torchvision.datasets.MNIST(root='./data', 
                           train=False, 
                           transform=transforms.ToTensor())
## 定义一个数据加载器
test_loader = Data.DataLoader(
    dataset = test_data, ## 使用的数据集
    batch_size=64, # 批处理样本大小
    shuffle = True, # 每次迭代前打乱数据
    num_workers = 2, # 使用两个进程 
)


##  可视化训练数据集的一个batch的样本来查看图像内容
for step, (b_x, b_y) in enumerate(train_loader):  
    if step > 0:
        break
## 输出训练图像的尺寸和标签的尺寸，都是torch格式的数据
print(b_x.shape)
print(b_y.shape)
test_data

输出

torch.Size([64, 1, 28, 28])
torch.Size([64])
Dataset MNIST
    Number of datapoints: 10000
    Root location: ./data
    Split: Test
    StandardTransform
Transform: ToTensor()

2 搭建RNN模型

class RNNimc(nn.Module):
    def __init__(self, input_dim, hidden_dim, layer_dim, output_dim):
        """
        input_dim:输入数据的维度(图片每行的数据像素点)
        hidden_dim: RNN神经元个数
        layer_dim: RNN的层数
        output_dim:隐藏层输出的维度(分类的数量)
        """
        super(RNNimc, self).__init__()
        self.hidden_dim = hidden_dim ## RNN神经元个数
        self.layer_dim = layer_dim ## RNN的层数
        # RNN
        self.rnn = nn.RNN(input_dim, hidden_dim, layer_dim,
                          batch_first=True, nonlinearity='relu')
        
        # 连接全连阶层
        self.fc1 = nn.Linear(hidden_dim, output_dim)
    def forward(self, x):
        # x:[batch, time_step, input_dim]
        # 本例中time_step＝图像所有像素数量／input_dim
        # out:[batch, time_step, output_size]
        # h_n:[layer_dim, batch, hidden_dim]
        out, h_n = self.rnn(x, None) # None表示h0会使用全0进行初始化
        # 选取最后一个时间点的out输出
        out = self.fc1(out[:, -1, :]) 
        return out

---

## 模型的调用
input_dim=28   # 图片每行的像素数量
hidden_dim=128  # RNN神经元个数
layer_dim = 1  # RNN的层数
output_dim=10  # 隐藏层输出的维度(10类图像)
MyRNNimc = RNNimc(input_dim, hidden_dim, layer_dim, output_dim)
print(MyRNNimc)

输出

RNNimc(
  (rnn): RNN(28, 128, batch_first=True)
  (fc1): Linear(in_features=128, out_features=10, bias=True)
)

3 训练模型

## 对模型进行训练
optimizer = torch.optim.RMSprop(MyRNNimc.parameters(), lr=0.0003)  
criterion = nn.CrossEntropyLoss()   # 损失函数
train_loss_all = []
train_acc_all = []
test_loss_all = []
test_acc_all = []
num_epochs = 30
for epoch in range(num_epochs):
    print('Epoch {}/{}'.format(epoch, num_epochs - 1))
    MyRNNimc.train() ## 设置模型为训练模式
    corrects = 0
    train_num  = 0
    for step,(b_x, b_y) in enumerate(train_loader):
        # input :[batch, time_step, input_dim]
        xdata = b_x.view(-1, 28, 28)
        output = MyRNNimc(xdata)     
        pre_lab = torch.argmax(output,1)
        loss = criterion(output, b_y) 
        optimizer.zero_grad()        
        loss.backward()       
        optimizer.step()  
        loss += loss.item() * b_x.size(0)
        corrects += torch.sum(pre_lab == b_y.data)
        train_num += b_x.size(0)
    ## 计算经过一个epoch的训练后在训练集上的损失和精度
    train_loss_all.append(loss / train_num)
    train_acc_all.append(corrects.double().item()/train_num)
    print('{} Train Loss: {:.4f}  Train Acc: {:.4f}'.format(
        epoch, train_loss_all[-1], train_acc_all[-1]))
    ## 设置模型为验证模式
    MyRNNimc.eval()
    corrects = 0
    test_num  = 0
    for step,(b_x, b_y) in enumerate(test_loader):
        # input :[batch, time_step, input_dim]
        xdata = b_x.view(-1, 28, 28)
        output = MyRNNimc(xdata)     
        pre_lab = torch.argmax(output,1)
        loss = criterion(output, b_y) 
        loss += loss.item() * b_x.size(0)
        corrects += torch.sum(pre_lab == b_y.data)
        test_num += b_x.size(0)
    ## 计算经过一个epoch的训练后在测试集上的损失和精度
    test_loss_all.append(loss / test_num)
    test_acc_all.append(corrects.double().item()/test_num)
    print('{} Test Loss: {:.4f}  Test Acc: {:.4f}'.format(
        epoch, test_loss_all[-1], test_acc_all[-1]))

4 模型保存和加载

# 保存
torch.save(MyRNNimc, 'rnn.pkl')

---

model = torch.load('rnn.pkl')
print(model)

输出

RNNimc(
  (rnn): RNN(28, 128, batch_first=True)
  (fc1): Linear(in_features=128, out_features=10, bias=True)
)

模型测试

import cv2
import matplotlib.pyplot as plt

# 第一步：读取图片
img = cv2.imread('./data/test/8.png') 
print(img.shape)

# 第二步：将图片转为灰度图
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
print(img.shape)
plt.imshow(img,cmap='Greys')

# 第三步：将图片的底色和字的颜色取反
img = cv2.bitwise_not(img)
plt.imshow(img,cmap='Greys')

# 第四步：将底变成纯白色，将字变成纯黑色
img[img<=144]=0
img[img>140]=255  # 130

# 显示图片
plt.imshow(img,cmap='Greys')

# 第五步：将图片尺寸缩放为输入规定尺寸
img = cv2.resize(img,(28,28))

# 第六步：将数据类型转为float32
img = img.astype('float32')

# 第七步：数据正则化
img /= 255

img = img.reshape(1,784)
# 第八步：增加维度为输入的规定格式

_img = torch.from_numpy(img).float()
# _img = torch.from_numpy(img).unsqueeze(0)

输出

(234, 182, 3)
(234, 182)

model.eval()
_img = _img.view(-1, 28, 28)
# 第九步：预测
outputs = model(_img)

# 第十步：输出结果
print(outputs.argmax())