A building-permit desk reuses a general encoder for work-order tags and cannot decide whether to keep the backbone frozen and train only a small head, or to unfreeze later layers. How should they decide?
Select an answer to reveal the explanation.
Short Explanation
Freeze the backbone and train a small head, or unfreeze later layers. That is the next measured comparison, read on validation, not a taste call. Always-unfreeze and always-freeze are not the associate rule, and a newer CUDA kernel does not settle it.
Full Explanation
Which layers update is an experimental factor, not a taste call and not an implementation detail. A frozen backbone plus a small head and a later unfreeze of upper layers are two treatments to score on the same validation set. Neither “always unfreeze” nor “always freeze” is the associate rule. Professional distributed training and kernel shopping do not settle the comparison.