学习Pytorch+Python之MNIST手写字体识别

import matplotlib.pyplot as plt
import torch
import torch.nn as nn
import numpy as np
import torchvision.utils
from torchvision import datasets, transforms
from torch.autograd import Variable
import torch.utils.data

判断是否能用GPU，如果能就用GPU，不能就用CPU

use_gpu = torch.cuda.is_available()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

数据转换，Pytorch的底层是tensor(张量)，所有用来训练的图像均需要转换成tensor

transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])])

下载数据集

data_train = datasets.MNIST(root="./data/", transform=transform, train=True, download=True)
data_test = datasets.MNIST(root="./data/", transform=transform, train=False)

加载数据集，批次大小为64，shuffle表示乱序

data_loader_train = torch.utils.data.DataLoader(dataset=data_train, batch_size=64, shuffle=True)
data_loader_test = torch.utils.data.DataLoader(dataset=data_test, batch_size=64, shuffle=True)

创建模型即网络架构

class Model(nn.Module):

def __init__(self):
    super(Model, self).__init__()
    #创建二维卷积
    self.conv1 = nn.Sequential(
        #输入特征数量为1，输出特征数量为64，卷积核大小为3x3，步长为1，边缘填充为1，保证了卷积后的特征尺寸与原来一样
        nn.Conv2d(1, 64, kernel_size=3, stride=1, c=1),
        nn.ReLU(),
        nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1),
        nn.ReLU(),
        #最大池化，特征数量不变，尺寸减半[(input-kernel_size)/stride + 1]
        nn.MaxPool2d(stride=2, kernel_size=2)
    )
    #创建全连接
    self.dense = nn.Sequential(
        nn.Linear(14*14*128, 1024),
        nn.ReLU(),
        #随机丢弃部分结点，防止过拟合
        nn.Dropout(p=0.5),
        nn.Linear(1024, 10)
    )
#创建好网络结构后，建立前向传播
def forward(self, x):
    #对数据进行卷积操作
    x = self.conv1(x)
    #改变特征形状
    x = x.c(-1, 14*14*128)
    #对特征进行全连接
    x = self.dense(x)
    return x

类实例化

model = Model()

指定数据训练次数

epochs = 5

设置学习率，即梯度下降的权重，其值越大收敛越快，越小收敛越慢

learning_rate = 0.0001

选用参数优化器，这里使用Adam

optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

选用损失函数，利率期货这里使用交叉熵函数，来判定实际的输出与期望的输出的接近程度

criterion = nn.CrossEntropyLoss()

判断是否使用GPU训练

if(use_gpu):

model = model.cuda()
loss_f = criterion.cuda()

用for循环的方式完成数据的批次训练

for epoch in range(epochs):

#定义并初始化训练过程的损失以及正确率
running_loss = 0
running_correct = 0
for data in data_loader_train:
    x_train, y_train = data
    x_train, y_train =x_train.cuda(), y_train.cuda()
    x_train = x_train.to(device)
    y_train = y_train.to(device)
    #将预处理好的数据加载到实例化好的model模型中，进行训练得到输出
    outputs = model(x_train)
    _, pred = torch.max(outputs.data, 1)
    #每次循环中，梯度必须清零，防止梯度堆叠
    optimizer.zero_grad()
    #调用设定的损失
    loss = criterion(outputs, y_train)
    #反向传播损失
    loss.backward()
    #参数更新
    optimizer.step()
    #更新损失
    running_loss += loss.item()
    #更新正确率
    running_correct += torch.sum(pred == y_train.data)
testing_correct = 0
#查看每轮训练后，测试数据集中的正确率
for data in data_loader_test:
    x_test, y_test = data
    x_test, y_test = Variable(x_test), Variable(y_test)
    x_test = x_test.to(device)
    y_test = y_test.to(device)
    outputs = model(x_test)
    _, pred = torch.max(outputs.data, 1)
    testing_correct += torch.sum(pred == y_test.data)
    print("Loss is {}, Training Accuray is {}%, Test Accurray is {}".format(running_loss/len(data_train), 100*running_correct/len(data_train), 100*testing_correct/len(data_test)))

测试训练好的模型

随机加载4个手写数字

data_loader_test = torch.utils.data.DataLoader(dataset=data_test, batch_size=4, shuffle=True)

函数next相关 https://www.runoob.com/python...

函数iter相关 https://www.runoob.com/python...

x_test,y_test = next(iter(data_loader_test))
inputs = Variable(x_test)
inputs = inputs.to(device)
pred = model(inputs)

_为输出的最大值，pred为最大值的索引值

_,pred = torch.max(pred, 1)
print('Predict Label is :', [i for i in pred.data])
print('Real Label is:', [ i for i in y_test] )
img = torchvision.utils.make_grid(x_test)
img = img.numpy().transpose(1, 2, 0)
std = [0.5]
mean = [0.5]
img = img*std+mean
plt.imshow(img)
plt.show()