Image Classification with Transfer Learning

Why transfer learning

Training a vision model from scratch needs huge data and compute. Transfer learning reuses a network pretrained on millions of images and adapts only the final layer to your classes.

import torch, torchvision
model = torchvision.models.resnet18(weights="DEFAULT")
for p in model.parameters():
    p.requires_grad = False           # freeze the backbone
model.fc = torch.nn.Linear(model.fc.in_features, num_classes)  # new head

The recipe

1. Load a pretrained backbone and freeze it.
2. Replace the classifier head for your number of classes.
3. Train just the head on your (small) dataset.
4. Optionally unfreeze and fine-tune the top layers with a low learning rate.

You can get strong results with only a few hundred labeled images.