不要上来就堆大模型！基线模型 → 进阶模型 → 微调/蒸馏，这条路线让你稳赢90%的项目。
残差网络，ImageNet 经典，简单有效。从 18 层到 152 层
torchvision.models.resnet50()
复合缩放，精度/效率平衡最好，B0-B7 可调
torchvision.models.efficientnet_b4()
把图像切成 patch 序列送入 Transformer，大数据集表现好
vit_patch16_224(pretrained=True)
绝大多数项目不需要从头训练模型！使用预训练模型 + 微调是行业标配：
解冻所有层，全部重训
预训练特征提取器冻结，只训分类头
用预训练模型抽特征，接XGBoost
import torch import torch.nn as nn from torchvision import models, transforms from torch.utils.data import DataLoader, Dataset from PIL import Image import torch.optim as optim # ===== 1. 数据准备 ===== transform = transforms.Compose([ transforms.Resize((224, 224)), transforms.ToTensor(), transforms.Normalize([0.485], [0.229]) ]) class MyDataset(Dataset): def __init__(self, paths, labels): self.paths = paths; self.labels = labels def __getitem__(self, idx): img = Image.open(self.paths[idx]).convert('RGB') return transform(img), self.labels[idx] def __len__(self): return len(self.paths) dataset = MyDataset(image_paths, labels) loader = DataLoader(dataset, batch_size=32, shuffle=True) # ===== 2. 加载预训练模型 ===== model = models.resnet50(pretrained=True) # 方案B：冻结底层 for param in model.parameters(): param.requires_grad = False # 替换分类头 num_classes = 5 model.fc = nn.Linear(model.fc.in_features, num_classes) # ===== 3. 训练 ===== device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') model = model.to(device) criterion = nn.CrossEntropyLoss() optimizer = optim.Adam(model.fc.parameters(), lr=0.001) for epoch in range(10): model.train() for images, labels in loader: images, labels = images.to(device), labels.to(device) optimizer.zero_grad() outputs = model(images) loss = criterion(outputs, labels) loss.backward() optimizer.step() print(ff'Epoch {epoch+1}, Loss: {loss.item():.4f}')