自监督学习

import torch
import torch.nn as nn
from torchvision import transforms
from lightly.loss import NTXentLoss
from lightly.models import ResNetGenerator

# 1. 构建编码器与投影头
encoder = ResNetGenerator('resnet-18')
model = nn.Sequential(encoder, nn.Linear(512, 128))

# 2. 定义对比损失与优化器
criterion = NTXentLoss(temperature=0.5)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)

# 3. 模拟双视图输入（实际应来自数据增强）
x1 = torch.randn(32, 3, 224, 224)
x2 = torch.randn(32, 3, 224, 224)

# 4. 前向传播与损失计算
z1 = model(x1)
z2 = model(x2)
loss = criterion(z1, z2)

# 5. 反向传播
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(f"对比损失: {loss.item():.4f}")

最短闭环：构建编码器 → 生成双视图 → 计算对比损失 → 优化。实际需替换为真实数据增强与分布式训练。

{
  "pretraining": {
    "epoch": 50,
    "loss": 0.412,
    "learning_rate": 0.0008,
    "batch_size": 256
  },
  "downstream_linear_probe": {
    "dataset": "CIFAR-10",
    "accuracy": 0.891,
    "frozen_encoder": true,
    "trainable_params": "linear_head_only"
  },
  "downstream_finetune": {
    "dataset": "CIFAR-10",
    "accuracy": 0.934,
    "frozen_encoder": false,
    "trainable_params": "all_layers"
  }
}