冻结骨干网络，训练任务头

Why train from zero? A model already trained on millions of images has learned useful features. Transfer learning reuses that knowledge for your task. 🧠

A pretrained model has two parts: the backbone that extracts general features, and the head that makes the final prediction for its original task.

The simplest recipe: freeze the backbone so it stays as is, then swap in a fresh head and train only that part on your own labels.

Start with a model that already carries good weights. torchvision hands you popular architectures with one call.

To freeze a layer you set requires_grad to False. PyTorch then skips computing gradients for it, so its weights never change.

Freezing the backbone keeps its hard-won features intact and means far fewer weights to update. That makes training faster and works with little data.

The old head predicts the wrong classes. Replace it with a fresh Linear layer sized for your number of classes.

A brand-new layer starts with requires_grad set to True automatically. So only your head will learn while the frozen backbone watches.

Pass just the head's parameters to your optimizer. Updating only these few weights is what makes this approach so cheap.

With its weights frozen, the backbone simply turns each image into a rich feature vector. Your head learns to read those features.

Freeze-and-train-head is ideal with a small dataset or limited compute. You get strong results in minutes instead of days.

You froze the backbone and replaced the head. Which weights actually update during training?

You learned to borrow a pretrained backbone, freeze it, attach a fresh head, and train just that head. Fast, cheap, and great with small data. 🎯