I have wasted weeks fine-tuning on pretty photos that were not the problem. The problem was ten lookalikes, night-shift lighting, and people shooting from the hip.
Collect failures first. A hundred honest misses with the image, the guess, and the truth will teach you more than another epoch on the original library.
Fine-tune is a scalpel
Use it when a cluster of SKUs is stable and still confused. Do not use it as the default response to a bad week of Wi-Fi.
I keep a baseline model I can roll back to in an afternoon. If the fine-tune hurts easy SKUs, it does not ship, no matter how good the loss curve looks.
Change packaging is a data event
When a vendor changes a label, that is not "the model decaying." It is a new class. Add photos. Do not immediately melt the old weights.
