快捷導航

Pytorch自定義CNN網(wǎng)絡(luò)實現(xiàn)貓狗分類詳解過程

更新時間：2022年12月08日 15:04:30 作者：專業(yè)女神殺手

PyTorch是一個開源的Python機器學習庫，基于Torch，用于自然語言處理等應(yīng)用程序。它不僅能夠?qū)崿F(xiàn)強大的GPU加速，同時還支持動態(tài)神經(jīng)網(wǎng)絡(luò)。本文將介紹PyTorch自定義CNN網(wǎng)絡(luò)實現(xiàn)貓狗分類，感興趣的可以學習一下

前言

數(shù)據(jù)集下載地址：

鏈接: https://pan.baidu.com/s/17aglKyKFvMvcug0xrOqJdQ?pwd=6i7m

Dogs vs. Cats(貓狗大戰(zhàn))來源Kaggle上的一個競賽題，任務(wù)為給定一個數(shù)據(jù)集，設(shè)計一種算法中的貓狗圖片進行判別。

數(shù)據(jù)集包括25000張帶標簽的訓練集圖片，貓和狗各125000張，標簽都是以cat or dog命名的。圖像為RGB格式j(luò)pg圖片，size不一樣。截圖如下：

一. 數(shù)據(jù)預處理

pytorch的數(shù)據(jù)預處理部分要寫成一個類，這個類繼承Dataset類，并必須要實現(xiàn)三個函數(shù)。

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__為構(gòu)造函數(shù)，我這里用力定義數(shù)據(jù)路徑，數(shù)據(jù)集劃分，transforms。

__getitem__為迭代函數(shù)，用來return單個數(shù)據(jù)的data和label。

__len__返回數(shù)據(jù)集的長度。

二. 定義網(wǎng)絡(luò)

在這個例子中，我們用一個簡單的4層卷積，2層全連接，最后跟一個sigmoid輸出二分類的概率的CNN網(wǎng)絡(luò)。

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)
        # 展開
        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定義網(wǎng)絡(luò)時，必須實現(xiàn)兩個函數(shù)，構(gòu)造函數(shù)主要定義一些網(wǎng)絡(luò)塊，forward函數(shù)實現(xiàn)前向推理過程。且在后續(xù)代碼中，如果定義對象model: ConvNet和數(shù)據(jù)image，可以直接通過model(image)來調(diào)用froward函數(shù)（python真的很神奇，C++出身的我理解這些騷操作好難）

三. 訓練模型

數(shù)據(jù)準備好了，模型網(wǎng)絡(luò)定義好了，下一步當然是訓練權(quán)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(); #每個批次平均，一個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，準備數(shù)據(jù)。先用我們之前定義的DogCat類來加載數(shù)據(jù)，但這個類繼承自dataset，是加載一條數(shù)據(jù)的。如果要批量加載數(shù)據(jù)，還要用pytorch內(nèi)部的另一個類DataLoader，然后在構(gòu)造函數(shù)里傳入batchsize就可以批量加載數(shù)據(jù)了。注意這里的類對象實際是一個生成器，后續(xù)通過循環(huán)就可以一直批量的去取數(shù)據(jù)了。

第二步，定義模型對象，有用顯卡就把模型放在顯卡上，沒有的話就用cpu跑。

第三步，定義一些超參數(shù)。因為是二分類，網(wǎng)絡(luò)最后一層為sigmoid輸出類別的概率值，所以選用二分類交叉熵損失函數(shù)。再設(shè)置一下學習率和優(yōu)化器。

第四步，訓練n個epoch。在每一個epoch里計算訓練集準去率，驗證集準確率，并保存模型。

最后結(jié)果像這樣