深度集成学习不均衡样本图像分类

Secede.2025-04-09 13:06

用五个不同的网络，然后对分类概率进行平均，得到分类结果。基本上分类精度可以提升10%

1.导入基本库

python 复制代码 
import torch
import copy
import torch.nn as nn
import torchvision.models as models
from torchvision import datasets
from torchvision import transforms
from tqdm import tqdm
from torch.utils.data import DataLoader
from torch.utils.data import random_split
from transformers import AutoModelForImageClassification,AutoConfig

2.数据集准备

python 复制代码 
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# 数据预处理
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),  
])

train_dataset = datasets.ImageFolder(root='./aug_datasets1', transform=transform)
dataset_size  = len(train_dataset)

train_size = int(0.8 * dataset_size)
val_size = dataset_size - train_size

train_dataset, val_dataset = random_split(train_dataset, [train_size, val_size])


train_dataloader = DataLoader(train_dataset, batch_size=32, shuffle=True)
val_dataloader = DataLoader(val_dataset, batch_size=32, shuffle=False)

3.定义不同模型与对应的训练策略

模型1 ResNet

python 复制代码 
class ResNet(nn.Module):
    def __init__(self, num_classes=21,train=True):
        super(ResNet, self).__init__()
        if(train):
            self.resnet = models.resnet50(weights=torchvision.models.ResNet50_Weights.IMAGENET1K_V1)
        else:
            self.resnet = models.resnet50(weights=None)
        in_features = self.resnet.fc.in_features
        self.resnet.fc = nn.Sequential(
            nn.Linear(in_features, 512),
            nn.ReLU(inplace=True),
            nn.Dropout(0.5),
            nn.Linear(512, num_classes)
        )
        self.resnet.to(device)
    def forward(self, x):
        return self.resnet(x)

    # 训练策略
    def startTrain(self, train_loader, val_loader):
        criterion = nn.CrossEntropyLoss()
        optimizer = torch.optim.AdamW(self.parameters(), lr=1e-4, weight_decay=1e-4)
        scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=50)
        Best_Acc = 0.0
        print("Training ResNet.....")
        for epoch in range(10):  # 训练 10 个 epoch
            self.train()
            train_loss = 0
            for batch in tqdm(train_loader):
                images, labels = batch
                images, labels = images.to(device), labels.to(device)
                
                optimizer.zero_grad()
                # 处理图像并将其传递给模型
                logits = self(images)

                # 计算损失并进行反向传播
                loss = criterion(logits, labels)
                loss.backward()
                optimizer.step()
                

                train_loss += loss.item()
            print(f"Epoch {epoch+1}/{10}, Train Loss: {train_loss/len(train_loader)}")
            scheduler.step()
            self.eval()
            val_loss = 0
            correct = 0
            total = 0

            with torch.no_grad():
                for batch in tqdm(val_loader):
                    images, labels = batch
                    images, labels = images.to(device), labels.to(device)

                    # 处理图像并传递给模型
                    logits = self(images)

                    # 计算损失
                    loss = criterion(logits, labels)
                    val_loss += loss.item()

                    # 计算准确率
                    _, predicted = torch.max(logits, 1)
                    total += labels.size(0)
                    correct += (predicted == labels).sum().item()

            print(f"Validation Loss: {val_loss/len(val_loader)}")
            print(f"Accuracy: {100 * correct / total}%")
            if(100 * correct / total > Best_Acc):
                Best_Acc = 100 * correct / total
                torch.save(self.state_dict(), './saved/resnet/model_weights_{}.pth'.format(Best_Acc))

模型2 EfficientNet

python 复制代码 
class EfficientNet(nn.Module):
    def __init__(self, num_classes=21,train=True):
        super(EfficientNet, self).__init__()
        if(train):
            self.effnet = models.efficientnet_b2(weights=torchvision.models.EfficientNet_B2_Weights.IMAGENET1K_V1)
        else:
            self.effnet = models.efficientnet_b2(weights=None)
        
        in_features = self.effnet.classifier[1].in_features
        self.effnet.classifier = nn.Sequential(
            nn.Linear(in_features, 512),
            nn.ReLU(inplace=True),
            nn.Dropout(0.5),
            nn.Linear(512, num_classes)
        )
        self.effnet.to(device)
    def forward(self, x):
        return self.effnet(x)

    # 训练策略
    def startTrain(self, train_loader, val_loader):
        # 焦点损失，gamma参数增强对少数类的关注
        criterion = nn.CrossEntropyLoss()
        optimizer = torch.optim.AdamW(self.parameters(), lr=1e-4, weight_decay=1e-4)
        scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=5)
        Best_Acc = 0.0
        print("Training EfficientNet.....")
        for epoch in range(10):  # 训练 10 个 epoch
            self.train()
            train_loss = 0
            for batch in tqdm(train_loader):
                images, labels = batch
                images, labels = images.to(device), labels.to(device)

                optimizer.zero_grad()
                # 处理图像并将其传递给模型
                logits = self(images)
                
                # 计算损失并进行反向传播
                loss = criterion(logits, labels)
                loss.backward()
                optimizer.step()
                

                train_loss += loss.item()
            print(f"Epoch {epoch+1}/{10}, Train Loss: {train_loss/len(train_loader)}")
            scheduler.step(train_loss/len(train_loader))
            self.eval()
            val_loss = 0
            correct = 0
            total = 0

            with torch.no_grad():
                for batch in tqdm(val_loader):
                    images, labels = batch
                    images, labels = images.to(device), labels.to(device)

                    # 处理图像并传递给模型
                    logits = self(images)

                    # 计算损失
                    loss = criterion(logits, labels)
                    val_loss += loss.item()

                    # 计算准确率
                    _, predicted = torch.max(logits, 1)
                    total += labels.size(0)
                    correct += (predicted == labels).sum().item()

            print(f"Validation Loss: {val_loss/len(val_loader)}")
            print(f"Accuracy: {100 * correct / total}%")
            if(100 * correct / total > Best_Acc):
                Best_Acc = 100 * correct / total
                torch.save(self.state_dict(), './saved/efficientnet/model_weights_{}.pth'.format(Best_Acc))

模型3 DenseNet

python 复制代码 
class DenseNet(nn.Module):
    def __init__(self, num_classes=21, train=True):
        super(DenseNet, self).__init__()
        self.num_classes = num_classes
        if(train):
            self.densenet = models.densenet121(weights=torchvision.models.DenseNet121_Weights.IMAGENET1K_V1)
        else:
            self.densenet = models.densenet121(weights=None) 
        
        in_features = self.densenet.classifier.in_features
        self.densenet.classifier = nn.Sequential(
            nn.BatchNorm1d(in_features),
            nn.Linear(in_features, 512),
            nn.ReLU(inplace=True),
            nn.Dropout(0.5),
            nn.Linear(512, num_classes)
        )
        self.densenet.to(device)
    def forward(self, x):
        return self.densenet(x)

    # 训练策略
    def startTrain(self, train_loader, val_loader):
        
        criterion = nn.CrossEntropyLoss()
        optimizer = torch.optim.Adam(self.parameters(), lr=1e-4)
        scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=50)
        Best_Acc = 0.0
        print("Training DenseNet.....")
        for epoch in range(10):  # 训练 10 个 epoch
            self.train()
            train_loss = 0
            for batch in tqdm(train_loader):
                images, labels = batch
                images, labels = images.to(device), labels.to(device)

                optimizer.zero_grad()
                # 处理图像并将其传递给模型
                logits = self(images)

                # 计算损失并进行反向传播
                loss = criterion(logits, labels)
                loss.backward()
                optimizer.step()
                

                train_loss += loss.item()
            print(f"Epoch {epoch+1}/{10}, Train Loss: {train_loss/len(train_loader)}")
            scheduler.step()
            self.eval()
            val_loss = 0
            correct = 0
            total = 0

            with torch.no_grad():
                for batch in tqdm(val_loader):
                    images, labels = batch
                    images, labels = images.to(device), labels.to(device)

                    # 处理图像并传递给模型
                    logits = self(images)

                    # 计算损失
                    loss = criterion(logits, labels)
                    val_loss += loss.item()

                    # 计算准确率
                    _, predicted = torch.max(logits, 1)
                    total += labels.size(0)
                    correct += (predicted == labels).sum().item()

            print(f"Validation Loss: {val_loss/len(val_loader)}")
            print(f"Accuracy: {100 * correct / total}%")
            if(100 * correct / total > Best_Acc):
                Best_Acc = 100 * correct / total
                torch.save(self.state_dict(), './saved/densenet/model_weights_{}.pth'.format(Best_Acc))

模型4 ResNeXt

python 复制代码 
class ResNeXt(nn.Module):

    def __init__(self, num_classes=21,train=True):
        super(ResNeXt, self).__init__()
        if(train):
            self.resnext50 = models.resnext50_32x4d(weights=torchvision.models.ResNeXt50_32X4D_Weights.IMAGENET1K_V1)
        else:
            self.resnext50 = models.resnext50_32x4d(weights=None)
        
        in_features = self.resnext50.fc.in_features
        self.resnext50.fc = nn.Sequential(
            nn.BatchNorm1d(in_features),
            nn.Linear(in_features, 512),
            nn.ReLU(inplace=True),
            nn.Dropout(0.5),
            nn.Linear(512, num_classes)
        )
        self.resnext50.to(device)
              
        self.to(device)
    def forward(self, x):
        return self.resnext50(x)

    def startTrain(self, train_loader, val_loader):
        
        optimizer = torch.optim.AdamW(self.parameters(), lr=1e-4)
        scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer, max_lr=5e-4, epochs=30, steps_per_epoch=len(train_loader))        
        criterion = nn.CrossEntropyLoss()
        Best_Acc = 0.0
        print("Training ResNeXt.....")
        for epoch in range(10):  # 训练 10 个 epoch
            self.train()
            train_loss = 0
            for batch in tqdm(train_loader):
                images, labels = batch
                images, labels = images.to(device), labels.to(device)

                optimizer.zero_grad()
                # 处理图像并将其传递给模型
                logits = self(images)

                # 计算损失并进行反向传播
                loss = criterion(logits, labels)
                loss.backward()
                optimizer.step()

                train_loss += loss.item()
            print(f"Epoch {epoch+1}/{10}, Train Loss: {train_loss/len(train_loader)}")
            scheduler.step(train_loss/len(train_loader))
            self.eval()
            val_loss = 0
            correct = 0
            total = 0

            with torch.no_grad():
                for batch in tqdm(val_loader):
                    images, labels = batch
                    images, labels = images.to(device), labels.to(device)

                    # 处理图像并传递给模型
                    logits = self(images)

                    # 计算损失
                    loss = criterion(logits, labels)
                    val_loss += loss.item()

                    # 计算准确率
                    _, predicted = torch.max(logits, 1)
                    total += labels.size(0)
                    correct += (predicted == labels).sum().item()

            print(f"Validation Loss: {val_loss/len(val_loader)}")
            print(f"Accuracy: {100 * correct / total}%")
            if(100 * correct / total > Best_Acc):
                Best_Acc = 100 * correct / total
                torch.save(self.state_dict(), './saved/se-resnext/model_weights_{}.pth'.format(Best_Acc))

模型5 SwinTransformer

python 复制代码 
class SwinTransformer(nn.Module):
    def __init__(self, num_classes=21,train=True):
        super(SwinTransformer, self).__init__()
        if(train):
            self.vit = AutoModelForImageClassification.from_pretrained('./swinv2-tiny-patch4-window16-256/models--microsoft--swinv2-tiny-patch4-window16-256/snapshots/f4d3075206f2ad5eda586c30d6b4d0500f312421/')   
            #这个地方怎么写加载模型
            self.vit.classifier = nn.Sequential(
                nn.Dropout(0.5),
                nn.Linear(self.vit.classifier.in_features, num_classes)
            )
            # 冻结Swin Transformer模型中的所有层
            for param in self.vit.parameters():
                param.requires_grad = False        
            
            # 只解冻最后两个Transformer块和分类头
            for param in self.vit.swinv2.encoder.layers[-4:].parameters():  # 假设你想解冻最后两层
                param.requires_grad = True
            for param in self.vit.classifier.parameters():
                param.requires_grad = True
        else:
            # 先加载 config，然后手动修改 num_labels
            config = AutoConfig.from_pretrained('./saved/swin-transformer/')
            config.num_labels = 21
            self.vit = AutoModelForImageClassification.from_pretrained('./saved/swin-transformer/',config=config)   
        self.vit.to(device)
        
    def forward(self, x):
        return self.vit(x)

    # 训练策略
    def startTrain(self, train_loader, val_loader):
        # 使用标签平滑处理，考虑到类别是连续尺度
        criterion = nn.CrossEntropyLoss()
        # 两阶段训练策略
        # 阶段1: 只训练解冻的层
        num_epochs_stage1 = 10
        num_epochs_stage2 = 10
        optimizer_stage1 = torch.optim.AdamW([p for p in self.parameters() if p.requires_grad], lr=1e-3)

        scheduler_stage1 = torch.optim.lr_scheduler.OneCycleLR(
            optimizer_stage1, max_lr=1e-3, epochs=num_epochs_stage1, steps_per_epoch=len(train_loader)
        )
        best_model_wts = copy.deepcopy(self.state_dict())
        print("Training SwinTransformer.....") 
        print("===== Stage 1 Training =====")
        Best_Acc = 0.0
        for epoch in range(num_epochs_stage1):  # 训练 10 个 epoch
            self.train()
            train_loss = 0
            for batch in tqdm(train_loader):
                images, labels = batch
                images, labels = images.to(device), labels.to(device)

                optimizer_stage1.zero_grad()
                # 处理图像并将其传递给模型
                outputs = self(images)
                logits = outputs.logits

                # 计算损失并进行反向传播
                loss = criterion(logits, labels)
                loss.backward()
                optimizer_stage1.step()


                train_loss += loss.item()
            print(f"Epoch {epoch+1}/{10}, Train Loss: {train_loss/len(train_loader)}")
            scheduler_stage1.step()
            self.eval()
            val_loss = 0
            correct = 0
            total = 0

            with torch.no_grad():
                for batch in tqdm(val_loader):
                    images, labels = batch
                    images, labels = images.to(device), labels.to(device)

                    # 处理图像并传递给模型
                    outputs = self(images)
                    logits = outputs.logits

                    # 计算损失
                    loss = criterion(logits, labels)
                    val_loss += loss.item()

                    # 计算准确率
                    _, predicted = torch.max(logits, 1)
                    total += labels.size(0)
                    correct += (predicted == labels).sum().item()

            print(f"Validation Loss: {val_loss/len(val_loader)}")
            print(f"Accuracy: {100 * correct / total}%")
            if(100 * correct / total > Best_Acc):
                Best_Acc = 100 * correct / total
                best_model_wts = copy.deepcopy(self.state_dict())
                self.vit.save_pretrained('./saved/swin-transformer/', safe_serialization=False)       
        
        # 阶段1结束后加载最佳模型权重
        self.load_state_dict(best_model_wts)    
        Best_Acc = 0.0
        print("===== Stage 2 Training =====")
        # 阶段2: 微调整个网络
        for param in self.parameters():
            param.requires_grad = True
        optimizer_stage2 = torch.optim.Adam(self.parameters(), lr=1e-6)
        scheduler_stage2 = torch.optim.lr_scheduler.OneCycleLR(
            optimizer_stage2, max_lr=5e-6, epochs=num_epochs_stage2, steps_per_epoch=len(train_loader)
        )
        for epoch in range(num_epochs_stage2):  # 训练 10 个 epoch
            self.train()
            train_loss = 0
            for batch in tqdm(train_loader):
                images, labels = batch
                images, labels = images.to(device), labels.to(device)

                optimizer_stage2.zero_grad()
                # 处理图像并将其传递给模型
                outputs = self(images)
                logits = outputs.logits

                # 计算损失并进行反向传播
                loss = criterion(logits, labels)
                loss.backward()
                optimizer_stage2.step()
                

                train_loss += loss.item()
            print(f"Epoch {epoch+1}/{10}, Train Loss: {train_loss/len(train_loader)}")
            scheduler_stage2.step()
            self.eval()
            val_loss = 0
            correct = 0
            total = 0

            with torch.no_grad():
                for batch in tqdm(val_loader):
                    images, labels = batch
                    images, labels = images.to(device), labels.to(device)

                    # 处理图像并传递给模型
                    outputs = self(images)
                    logits = outputs.logits

                    # 计算损失
                    loss = criterion(logits, labels)
                    val_loss += loss.item()

                    # 计算准确率
                    _, predicted = torch.max(logits, 1)
                    total += labels.size(0)
                    correct += (predicted == labels).sum().item()

            print(f"Validation Loss: {val_loss/len(val_loader)}")
            print(f"Accuracy: {100 * correct / total}%")
            if(100 * correct / total > Best_Acc):
                Best_Acc = 100 * correct / total
                self.vit.save_pretrained('./saved/swin-transformer/', safe_serialization=False)

4.分别训练，然后得到权重

python 复制代码 
    swinTransformer= SwinTransformer()
    swinTransformer.startTrain(train_dataloader,val_dataloader)
      
    efficientNet= EfficientNet()
    efficientNet.startTrain(train_dataloader,val_dataloader)

    resNet= ResNet()
    resNet.startTrain(train_dataloader,val_dataloader)
    
    resNeXt= ResNeXt()
    resNeXt.startTrain(train_dataloader,val_dataloader)
    
    denseNet= DenseNet()
    denseNet.startTrain(train_dataloader,val_dataloader)

5.构建集成分类模型

python 复制代码 
import torch
import torchvision.transforms as transforms
import torch.nn as nn
from torchvision import datasets
from torchvision import transforms
from tqdm import tqdm
from torch.utils.data import DataLoader
from torch.utils.data import random_split
from tqdm import tqdm
from PIL import Image

def remove_prefix_from_state_dict(state_dict, prefix='resnext.'):
    return {"resnext50." + k[len(prefix):] if k.startswith(prefix) else k: v for k, v in state_dict.items()}


# 定义集成模型
class EnsembleModel():
    def __init__(self, efficientNet, resNet, resNeXt, denseNet,swinTransformer):
        super(EnsembleModel, self).__init__()

        self.efficientNet= efficientNet.eval()
        self.resNet= resNet.eval()
        self.resNeXt= resNeXt.eval()
        self.denseNet= denseNet.eval()
        self.swinTransformer= swinTransformer.eval()

    def predict(self, x):
        efficientNet_out = torch.softmax(self.efficientNet(x),dim=1)
        resNet_out = torch.softmax(self.resNet(x),dim=1)
        resNeXt_out = torch.softmax(self.resNeXt(x),dim=1)
        denseNet_out = torch.softmax(self.denseNet(x),dim=1)
        swinTransformer_out = torch.softmax(self.swinTransformer(x).logits,dim=1)
        avg_pred = (efficientNet_out + resNet_out + resNeXt_out + denseNet_out + swinTransformer_out ) / 5
        return avg_pred

这样就可以提升性能

相关推荐
xier_ran11 小时前
【infra之路】AWQ 详解：激活感知权重保护，让 W4A16 量化精度超越 GPTQ
线性代数·算法·机器学习·量化·infra
亦皓ai11 小时前
AI时代后端工程（二）：AI把代码写得越来越快，我却越来越不敢让它直接开工了
人工智能·算法·机器学习·搜索引擎·transformer
柳絮飞祭奠13 小时前
Dify Windows Docker Desktop 部署文档
人工智能·深度学习·语言模型·数据分析·transformer·边缘计算·集成学习
咖啡星人k13 小时前
2026 AI Agent 长期记忆：为什么 AI 助手总“聊完就忘“？MonkeyCode 免费上手
人工智能·深度学习·机器学习·语言模型·自然语言处理
Rocky Ding*14 小时前
一文读懂Wan 3.0视频创作大模型核心基础知识
论文阅读·人工智能·深度学习·机器学习·aigc·ai-native·wan 3.0
我叫孙一鸣-专注电子元器件15 小时前
自动驾驶的眼睛正在换代：激光雷达的“小镜子”正在改变未来
人工智能·机器学习·自动驾驶·卫星通讯·压电驱动器
richard_first16 小时前
Transformer 与大语言模型：第10章 Residual （残差连接）
人工智能·深度学习·机器学习·transformer
北山小恐龙17 小时前
使用手势在电脑刷抖音
人工智能·python·深度学习·yolo·机器学习·计算机视觉
格林威18 小时前
C#图像分块处理：图像按行或按块（Tile）切分,多个 CPU 核心同时处理不同的区域
开发语言·图像处理·人工智能·机器学习·计算机视觉·c#·工业相机
Sinclair19 小时前
从神经网络到文件加密：一次关于 SIREN、ARX 与密码安全性的实验探索
人工智能·机器学习
热门推荐
01国内可直接用、免费额度/永久免费的大模型API清单（含 SiliconFlow、火山、阿里、智谱、百度、Kimi、DeepSeek、DMXAPI 等）02GitHub 镜像站点032026 国产 AI 大模型横评：DeepSeek、通义千问、Kimi、文心一言、星火、豆包谁更能打？04历年考研数学一、数学二、数学三真题试卷及答案PDF05微信历史版本含下载地址（ Windows PC | 安卓 | MAC ）及设置微信不更新06Agnes AI 免费 API 接入指南：文本、生图、生视频，一套接口全免费07AI 编程 IDE 全景解析 2026：Agent 全面接管开发链路08VSCode+CodeX扩展，无法启动09如何新建文件夹？ 电脑新建文件夹的4种方法102026年7月AI圈大地震：GPT-5.6被政府限制、Claude入驻Slack、Anthropic自研芯片