最近需要开始入门深度学习相关的内容，因此先选择了小土堆的pytorch教程，了解了pytorch的基本使用方法，并且跟着视频搭建了一个自己的神经网络，并在数据集上得到了验证。本文是对看完这系列教程之后的总结，仅作分享，如有谬误欢迎指正。

土堆教程链接：PyTorch深度学习快速入门教程（绝对通俗易懂！）【小土堆】_哔哩哔哩_bilibili

概述

就我个人的理解而言，我们在日常项目种常说到的训练一个模型，其实质是通过建立一个模型，让其在训练集上进行参数的优化，最后再应用到验证集中得以验证。

因此，我将这个过程归纳为了以下几步，同时也作为本文后续的小标题：

1. 加载数据

2. 定义模型

3. 训练与验证

4. 完整的代码

1. 加载数据

加载数据相关的代码如下：

# 加载数据
train_set = torchvision.datasets.CIFAR10(root='./dataset', train=True,download=True, transform=torchvision.transforms.ToTensor())
test_set = torchvision.datasets.CIFAR10(root='./dataset', train=False,download=True, transform=torchvision.transforms.ToTensor())

# dataloader
train_dataloader = DataLoader(train_set, batch_size=64)
test_dataloader = DataLoader(test_set, batch_size=64)

这部分代码中涉及以下几个点：

1. 采用torchvision的datasets方法加载数据集并保存在指定路径下，并用transform.Totensor对其格式做初步的处理。

2. 使用dataloader将数据分为64个一组，方面在后续的训练和验证中批量载入。

2. 定义模型

定义模型相关的代码如下：

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.model = nn.Sequential(
            nn.Conv2d(3, 32, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(32, 32, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(32, 64, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(kernel_size=2),
            nn.Flatten(),
            nn.Linear(64*4*4, 64),
            nn.Linear(64, 10)
        )

    def forward(self, x):
        x = self.model(x)
        return x

模型的定义都遵循这样的模板，在init中定义模型结构，在forward中定义输入与输出变换。

其中，在init部分，有这几个点需要注意：

1. Conv2d：是卷积方法，其作用是丰富数据的可观测指标

2. MaxPool2d:池化层，用于减少数据维度，加快训练

3. Linear：线性层，起到归纳结论的作用

3. 训练与验证

for i in range(epoch):
    print("-----第{}轮训练开始-----".format(i))
    net.train()
    for train_data in train_dataloader:
        imgs, targets = train_data
        imgs = imgs.to(device)
        targets = targets.to(device)
        outputs = net(imgs)
        loss = loss_fn(outputs, targets)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        total_train_step += 1
        if total_train_step % 100 == 0:
            end_time = time.time()
            print("已完成{}次训练,loss:{},用时:{}".format(total_train_step,loss.item(), end_time - start_time))
            writer.add_scalar('train_loss', loss.item(), total_train_step)
    # 测试步骤开始
    total_test_loss = 0
    total_test_accuracy = 0
    net.eval()
    with torch.no_grad():
        for test_data in test_dataloader:
            imgs, targets = test_data
            imgs = imgs.to(device)
            targets = targets.to(device)
            outputs = net(imgs)
            loss = loss_fn(outputs, targets)
            total_test_loss += loss.item()
            accuracy = (outputs.argmax(1) == targets).sum()
            total_test_accuracy += accuracy.item()
    total_test_step += 1
    # print("整体测试集上的Loss:{}".format(total_test_loss))
    print("整体测试集上的正确率:{}".format(total_test_accuracy / test_length))
    writer.add_scalar('test_loss', total_test_loss, total_test_step)

在训练时，遵循这样的流程：取数据-送入模型得到输出-得到结果与理想值之间的差异-清空上一轮的梯度数据-反向传播-优化参数

验证时，则不再需要关注梯度，而是单纯使用经过训练后的模型对验证集上的初始数据做处理，最后得到输出，进而算出相关指标。

4. 完整的代码

import torch
import torchvision
from torch import optim, nn
from torch.nn import CrossEntropyLoss
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import time

# from p23_model import *

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.model = nn.Sequential(
            nn.Conv2d(3, 32, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(32, 32, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(32, 64, kernel_size=5, stride=1, padding=2),
            nn.MaxPool2d(kernel_size=2),
            nn.Flatten(),
            nn.Linear(64*4*4, 64),
            nn.Linear(64, 10)
        )

    def forward(self, x):
        x = self.model(x)
        return x

# 定义训练设备
device= torch.device("cuda" if torch.cuda.is_available() else "cpu")
start_time = time.time()

# 加载数据
train_set = torchvision.datasets.CIFAR10(root='./dataset', train=True,download=True, transform=torchvision.transforms.ToTensor())
test_set = torchvision.datasets.CIFAR10(root='./dataset', train=False,download=True, transform=torchvision.transforms.ToTensor())

# 计算长度
train_length = len(train_set)
test_length = len(test_set)

print("训练集长度为{},测试集长度为{}".format(train_length,test_length))

# dataloader
train_dataloader = DataLoader(train_set, batch_size=64)
test_dataloader = DataLoader(test_set, batch_size=64)

# 创建模型
net = Net()
net = net.to(device)

# 损失函数
loss_fn = CrossEntropyLoss()
loss_fn = loss_fn.to(device)

# 优化器
learning_rate = 1e-2
optimizer = optim.SGD(net.parameters(), lr=learning_rate)

# 看板
writer = SummaryWriter("p23_board")
# 训练
total_train_step = 0
total_test_step = 0
epoch = 10

for i in range(epoch):
    print("-----第{}轮训练开始-----".format(i))
    net.train()
    for train_data in train_dataloader:
        imgs, targets = train_data
        imgs = imgs.to(device)
        targets = targets.to(device)
        outputs = net(imgs)
        loss = loss_fn(outputs, targets)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        total_train_step += 1
        if total_train_step % 100 == 0:
            end_time = time.time()
            print("已完成{}次训练,loss:{},用时:{}".format(total_train_step,loss.item(), end_time - start_time))
            writer.add_scalar('train_loss', loss.item(), total_train_step)
    # 测试步骤开始
    total_test_loss = 0
    total_test_accuracy = 0
    net.eval()
    with torch.no_grad():
        for test_data in test_dataloader:
            imgs, targets = test_data
            imgs = imgs.to(device)
            targets = targets.to(device)
            outputs = net(imgs)
            loss = loss_fn(outputs, targets)
            total_test_loss += loss.item()
            accuracy = (outputs.argmax(1) == targets).sum()
            total_test_accuracy += accuracy.item()
    total_test_step += 1
    # print("整体测试集上的Loss:{}".format(total_test_loss))
    print("整体测试集上的正确率:{}".format(total_test_accuracy / test_length))
    writer.add_scalar('test_loss', total_test_loss, total_test_step)

writer.close()

我们训练得到了自己的模型，现在我们随便找一张照片，让模型处理一下该照片，看看输出是什么：

from PIL import Image
from torchvision import transforms

from p23_model import *

# 加载待测试的分类图片
img_path = "ship.jpg"
img = Image.open(img_path)
img = img.convert('RGB')

transform = transforms.Compose([
    transforms.Resize((32, 32)),
    transforms.ToTensor(),
])

img = transform(img)
img = torch.reshape(img, (1, 3, 32, 32))

# 加载训练好的模型
model = torch.load("gpu_model_29.pth",map_location=torch.device('cpu'))
model.eval()

# 测试
with torch.no_grad():
    output = model(img)
print(output.argmax(dim=1))