机器学习——利用神经网络进行微波图像分类

目录

第五篇博客感言

任务描述

模型选择和代码实现

源码和报告

---

第五篇博客感言

勤能补拙。

---

任务描述

如题，利用神经网络进行微波图像分类，数据集是七类战斗车辆的微波图像（已分类），利用这些数据集训练好网络后，对测试集（未分类）进行分类，将分类结果上传到kaggle与正确结果对比得到准确率。要求准确率尽可能高。部分数据集内容如下：

---

模型选择和代码实现

总共测试了三个经典的图像分类网络模型：LeNet、VGG和AlexNet。前两者准确率均在70%左右，考虑调整网络参数或者增强数据集，但都没有很好的效果，遂放弃，进而选择AlexNet，达到了至少94%的分类准确率。

本来以为写一个神经网络训练模型的代码很困难，毕竟我是电子信息出身，这方面实在知之甚少。但是多亏了pytorch，现在可以直接调用这些网络模型。所以需要掌握的是如何去编写整个训练和测试的流程，而这件事确实是有套路的：

1）数据提取、转换和增强

2）定义网络结构并设定参数

3）训练模型

4）测试

以AlexNet网络为例：

from torchvision import models
import pandas as pd
from torchsummary import summary
import torch
import torch.nn as nn
from torchvision import transforms
from torchvision.datasets import ImageFolder
from torch.utils.data import DataLoader

#数据提取、转换和增强
transform = transforms.Compose([
        transforms.Resize(224),
        transforms.RandomHorizontalFlip(),       #随机翻转
        transforms.ToTensor(),                          # 转为 Tensor
        transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5)),# 归一化
])

test_transform = transforms.Compose([
        transforms.Resize(224),
        transforms.ToTensor(),                          # 转为 Tensor
        transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5)),# 归一化
])

train_data_dir = "train"
train_data = ImageFolder(train_data_dir, transform=transform)
train_data_loader = DataLoader(train_data, batch_size=16, shuffle=True,num_workers=0)

val_data_dir = "val"
val_data = ImageFolder(val_data_dir, transform=transform)
val_data_loader = DataLoader(val_data, batch_size=16, shuffle=True,num_workers=0)

test_data_dir = "test"
test_data = ImageFolder(test_data_dir, transform=test_transform)
test_data_loader = DataLoader(test_data)

#定义网络结构
classes = 7
alexnet = models.alexnet(pretrained=True)
alexnet.classifier[6] = nn.Linear(alexnet.classifier[6].in_features, classes)
alexnet.classifier[5] = nn.Sequential(
    alexnet.classifier[5]
)
alexnet.classifier[6] = nn.Sequential(
    nn.Linear(alexnet.classifier[6].in_features, classes)
)

#各参数设置
net = alexnet
optimizer = torch.optim.Adam(net.parameters(),lr=0.0001, weight_decay=1e-5)
loss_func = nn.CrossEntropyLoss()
num_epoch = 1

#训练模型
for epoch in range(num_epoch):  # 设置训练的迭代次数
        print(f"epoch{epoch + 1}\n-------------------")
        loss, current, n = 0.0, 0.0, 0
        #训练集
        net.train()
        for i, data in enumerate(train_data_loader, 0):
                # 输入数据
                inputs, labels = data
                # forward + backward
                outputs = net(inputs)
                # 计算损失值
                #loss = criterion(outputs, labels)
                cur_loss = loss_func(outputs, labels)
                _, pred = torch.max(outputs, axis=1)
                # 计算每批次的准确率， output.shape[0]为该批次的多少
                cur_acc = torch.sum(labels == pred) / outputs.shape[0]
                # 反向传播
                optimizer.zero_grad()
                cur_loss.backward()
                optimizer.step()
                # 取出loss值和精度值
                loss += cur_loss.item()
                current += cur_acc.item()
                n = n + 1
        print('train_loss' + str(loss / n) + '  ' + 'train_acc' + str(current / n))
        #验证集
        net.eval()
        for i, data in enumerate(val_data_loader, 0):
                # 输入数据
                inputs, labels = data
                # forward + backward
                outputs = net(inputs)
                # 计算损失值
                #loss = criterion(outputs, labels)
                cur_loss = loss_func(outputs, labels)
                _, pred = torch.max(outputs, axis=1)
                # 计算每批次的准确率， output.shape[0]为该批次的多少
                cur_acc = torch.sum(labels == pred) / outputs.shape[0]
                # 取出loss值和精度值
                loss += cur_loss.item()
                current += cur_acc.item()
                n = n + 1
        print('val_loss' + str(loss / n) + '  ' + 'val_acc' + str(current / n))
#测试集
pred, target = [], []
net.eval()
for idx_, (image_, _) in enumerate(test_data_loader):  # For every batch of data...
        # noinspection PyCallingNonCallable
        output_ = net(image_)  # Generate our predictions
        pred.append(output_.argmax(dim=1).item())

result = pd.DataFrame(pred, columns=['category'])
result.index.name = 'id'
result.to_csv('result_alexnet.csv')

print('Finished Training')

net = net.to("cuda:0")
summary(net,input_size=(3,224,224))

---

源码和报告

所有代码、汇报ppt和项目报告均在下面的链接中。

https://withbreeze.lanzouv.com/iyvBT257jtib

密码:cbvq

---

欢迎交流讨论！