Image Classification with EfficientNet: Transfer Learning Best Practices

Fine-Tuning Strategy

Phase 1: Train head only (5 epochs)

model = EfficientNet.from_pretrained('efficientnet-b4')
for param in model.parameters():
    param.requires_grad = False
# Unfreeze classifier
for param in model._fc.parameters():
    param.requires_grad = True

Phase 2: Unfreeze top blocks (10 epochs)

for param in model._blocks[-20:].parameters():
    param.requires_grad = True

Phase 3: Full fine-tune with low LR (10 epochs)

optimizer = AdamW(model.parameters(), lr=1e-5)

Label Smoothing

criterion = nn.CrossEntropyLoss(label_smoothing=0.1)

Reduces overconfidence — typically +1-2% accuracy on small datasets.