Establish a frozen baseline
Start with a pretrained representation and train a small task-specific head while the base is frozen. This separates the value of existing features from the effect of fine-tuning. Match input scaling and shape to the original model, and split related examples together before experimentation. A strong frozen baseline may already meet the requirement with less training cost.
Design a baseline for 2,000 labeled product images. Hold out products, not just photographs.
Unfreeze deliberately
Fine-tuning updates pretrained weights on the new task. First let the randomly initialized head settle; then unfreeze selected layers and use a conservative learning rate. In Keras, changing trainable settings requires recompilation. Batch normalization needs special care because updating its statistics can alter behavior even when the dataset is small.
Compare a frozen base with a partially unfrozen base using identical validation examples. Record which layers change.
Recognize negative transfer
Transfer is not automatically beneficial. A source representation can encode irrelevant visual cues or fail to represent the target domain. Inspect errors on rare classes and capture conditions, and compare against a simpler model. If training improves while validation degrades, consider fewer trainable layers, better data, or stronger regularization before adding compute.
A fine-tuned model improves training accuracy by five points but loses three validation points. Name two controlled follow-up experiments.
Keep the release reproducible
Save preprocessing, label mapping, source checkpoint identity, trainable-layer configuration, and the selected weights together. Evaluate the selected candidate once on an untouched holdout. A reproducible export matters as much as the notebook score: test the same sample through the training and deployed input paths.
Create a release record that lets a teammate reconstruct the frozen baseline and the chosen fine-tuned candidate.
Why train the head before unfreezing the base?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.