Transfer Learning with CNNs
Training a large CNN from scratch requires massive data and computation. Transfer learning reuses a pre‑trained network (e.g., on ImageNet) and adapts it to your task, often with very little data.
Transfer learning: take a model trained on a large dataset, freeze early layers (generic features), retrain later layers (task‑specific).
Why Transfer Learning Works
Early CNN layers detect universal features: edges, corners, colors. These are useful for any vision task. Only the final layers need to adapt to your specific classes.
Two Common Approaches
- Feature extraction: Freeze all pre‑trained layers, add new classifier on top. Train only the new layers.
- Fine‑tuning: Unfreeze some top layers and retrain them along with the new classifier. Often yields better accuracy.
Example with PyTorch (ResNet18)
import torchvision.models as models
model = models.resnet18(pretrained=True)
# Freeze all layers
for param in model.parameters():
param.requires_grad = False
# Replace the final fully connected layer
model.fc = nn.Linear(512, num_classes)When to Use Transfer Learning
- Small dataset (hundreds to few thousand images).
- Similar to pre‑training domain (e.g., natural images → medical images may still work).
- Limited compute resources.
Two Minute Drill
- Transfer learning reuses pre‑trained models.
- Early layers learn universal features; later layers task‑specific.
- Feature extraction: freeze pre‑trained layers.
- Fine‑tuning: retrain some top layers.
Need more clarification?
Drop us an email at career@quipoinfotech.com