Unfreeze fine-tuning gradually, not all at once

Unfreeze fine-tuning gradually, not all at once A workflow diagram generated by Archify. 01 / Fine-tuning level EX / Over-unfreezing risks Cheap level first Unfreeze more if needed Lock in, test once Start · pretrained + new data · Fine-tuning level › Cheap level first Start pretrained + new data Feature extraction · train the head only · Fine-tuning level › Cheap level first · cheapest Feature extraction train the head only cheapest Partial fine-tuning · unfreeze last layers · Fine-tuning level › Unfreeze more if needed · middle ground Partial fine-tuning unfreeze last layers middle ground Full fine-tuning · all layers get gradients · Fine-tuning level › Unfreeze more if needed · most expensive Full fine-tuning all layers get gradients most expensive Good enough · keep best checkpoint · Fine-tuning level › Lock in, test once Good enough keep best checkpoint Run test once · the final number · Fine-tuning level › Lock in, test once · no more tuning Run test once the final number no more tuning Overfitting · validation loss rises · Over-unfreezing risks › Unfreeze more if needed Overfitting validation loss rises not there yet still not there target met config locked refreeze some layers dataset too small Legend Agent logic Context / trace Cloud service External system