Pytorch自定义CNN网络实现猫狗分类

数据集下载地址：https://www.microsoft.com/en-us/download/confirmation.aspx?id=54765

Dogs vs. Cats(猫狗大战)来源Kaggle上的一个竞赛题，任务为给定一个数据集，设计一种算法中的猫狗图片进行判别。

数据集包括25000张带标签的训练集图片，猫和狗各125000张，标签都是以cat  or  dog命名的。图像为RGB格式jpg图片，size不一样。截图如下：

一. 数据预处理

pytorch的数据预处理部分要写成一个类，这个类继承Dataset类，并必须要实现三个函数。

from torch.utils.data import DataLoader,Dataset
from torchvision import transforms as T
import matplotlib.pyplot as plt
import os
from PIL import Image

class DogCat(Dataset):
    def __init__(self, root, transforms=None, train=True):
        imgs = [os.path.join(root,img) for img in os.listdir(root)]
        imgs_num = len(imgs)

        if train:
            self.imgs = imgs[:int(0.7 * imgs_num)]
        else:
            self.imgs = imgs[int(0.3 * imgs_num):]

        if transforms is None:
            normalize = T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

            self.transforms = T.Compose([
                    T.Resize(224),
                    T.CenterCrop(224),
                    T.ToTensor(),
                    normalize
            ])
        else:
            self.transforms = transforms
             
    def __getitem__(self, index):
        img_path = self.imgs[index]
        # dog label : 1           cat label : 0
        label = 1 if "dog" in img_path.split('/')[-1] else 0
        data = Image.open(img_path)
        data = self.transforms(data)
        return data,label

    def __len__(self):
        return len(self.imgs)

__init__为构造函数，我这里用力定义数据路径，数据集划分，transforms。

__getitem__为迭代函数，用来return单个数据的data和label。

__len__返回数据集的长度。

二. 定义网络

在这个例子中，我们用一个简单的4层卷积，2层全连接，最后跟一个sigmoid输出二分类的概率的CNN网络。

import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F


class ConvNet(nn.Module):
    def __init__(self):
        super(ConvNet, self).__init__()
        self.conv1 = nn.Conv2d(3, 32, 3)
        self.conv2 = nn.Conv2d(32, 64, 3)
        self.conv3 = nn.Conv2d(64, 128, 3)
        self.conv4 = nn.Conv2d(128, 128, 3)
        self.max_pool = nn.MaxPool2d(2)
        self.relu = nn.ReLU()
        self.sigmoid = nn.Sigmoid()
        # 12*12 for size(224,224)    7*7 for size(150,150)
        self.fc1 = nn.Linear(128*12*12, 512)
        self.fc2 = nn.Linear(512, 1)
        
    def forward(self, x):
        in_size = x.size(0)
        x = self.conv1(x)
        x = self.relu(x)
        x = self.max_pool(x)
        x = self.conv2(x)
        x = self.relu(x)
        x = self.max_pool(x)
        x = self.conv3(x)
        x = self.relu(x)
        x = self.max_pool(x)
        x = self.conv4(x)
        x = self.relu(x)
        x = self.max_pool(x)
        # 展开
        x = x.view(in_size, -1)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        x = self.sigmoid(x)
        return x

pytorch定义网络时，必须实现两个函数，构造函数主要定义一些网络块，forward函数实现前向推理过程。且在后续代码中，如果定义对象model: ConvNet和数据image，可以直接通过model(image)来调用froward函数（python真的很神奇，C++出身的我理解这些骚操作好难）

三．训练模型

数据准备好了，模型网络定义好了，下一步当然是训练权重了。

import torch
import torch.nn as nn
from torch.utils.data import DataLoader,Dataset
from dataset import DogCat
from network import ConvNet
from draw import draw_acc,draw_loss


train_data_root = "/home/elvis/workfile/dataset/dataset_kaggledogvscat/train"
batch_size = 256
# 1. prepare dataset
train_data = DogCat(train_data_root, train=True)
val_data = DogCat(train_data_root, train=False)
train_dataloader = DataLoader(train_data,batch_size=batch_size,shuffle=True)
val_dataloader = DataLoader(val_data,batch_size=batch_size,shuffle=True)

# 2. load model
model = ConvNet()
if torch.cuda.is_available():
    model.cuda()

# 3. prepare super parameters
criterion = nn.BCELoss()
learning_rate = 1e-3
# optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, momentum=0.9)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

# 4. train
train_loss_epoch = []
train_acc_epoch = []
val_loss_epoch = []
val_acc_epoch = []
for epoch in range(1, 10):
    model.train()
    train_loss = 0;
    train_acc = 0;
    for batch_idx, (data, target) in enumerate(train_dataloader):
        if torch.cuda.is_available():
            data, target = data.cuda(), target.cuda().float().unsqueeze(-1)
        else:
            data, target = data, target.float().unsqueeze(-1)
        optimizer.zero_grad()
        output = model(data)
        # print(output)
        loss = criterion(output, target)
        train_loss += loss.item();
        pred = torch.tensor([[1] if num[0] >= 0.5 else [0] for num in output]).cuda();
        train_acc += pred.eq(target.long()).sum().item();
        loss.backward()
        optimizer.step()
        if(batch_idx+1)%10 == 0: 
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, (batch_idx+1) * len(data), len(train_dataloader.dataset),
                100. * (batch_idx+1) / len(train_dataloader), loss.item()))
    train_loss_epoch.append(train_loss / len(train_dataloader));
    train_acc_epoch.append(train_acc / len(train_dataloader.dataset));
    print('\nTrain set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)'.format(train_loss / len(train_dataloader), train_acc, len(train_dataloader.dataset),
                                                                                    100. * train_acc / len(train_dataloader.dataset)));

    # val
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for batch_idx, (data, target) in enumerate(val_dataloader):
            if torch.cuda.is_available():
                data, target = data.cuda(), target.cuda().float().unsqueeze(-1)
            else:
                data, target = data, target.float().unsqueeze(-1)
            output = model(data)
            # print(output)
            test_loss += criterion(output, target).item(); #每个批次平均，一个epoch里所有批次求和
            pred = torch.tensor([[1] if num[0] >= 0.5 else [0] for num in output]).cuda()
            correct += pred.eq(target.long()).sum().item()
    print('Valid set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(test_loss/len(val_dataloader), correct, len(val_dataloader.dataset),
                                                                                    100. * correct / len(val_dataloader.dataset)));
    val_loss_epoch.append(test_loss / len(val_dataloader));
    val_acc_epoch.append(correct / len(val_dataloader.dataset));

    # Save model
    val_acc_rate = correct / len(val_dataloader.dataset);
    save = True
    best = "best.pt"
    last = "last.pt"
    if save:
        # Save last, best and delete
        torch.save(model.state_dict(), last)
        if val_acc_rate == max(val_acc_epoch):
            torch.save(model.state_dict(), best)
            print("save epoch {} model".format(epoch))

    
# 5. drawing
draw_loss(train_loss_epoch, val_loss_epoch)
draw_acc(train_acc_epoch,val_acc_epoch)
    

第一步，准备数据。先用我们之前定义的DogCat类来加载数据，但这个类继承自dataset，是加载一条数据的。如果要批量加载数据，还要用pytorch内部的另一个类DataLoader，然后在构造函数里传入batchsize就可以批量加载数据了。注意这里的类对象实际是一个生成器，后续通过循环就可以一直批量的去取数据了。

第二步，定义模型对象，有用显卡就把模型放在显卡上，没有的话就用cpu跑。

第三步，定义一些超参数。因为是二分类，网络最后一层为sigmoid输出类别的概率值，所以选用二分类交叉熵损失函数。再设置一下学习率和优化器。

第四步，训练n个epoch。在每一个epoch里计算训练集准去率，验证集准确率，并保存模型。

最后结果像这样

有条件的可以多训练几个epoch试试。