What should you inspect when an object detector's regression loss becomes NaN?
Computer VisionLevel: IntermediateInvalid box geometry and unstable loss
What should you inspect when an object detector's regression loss becomes NaN?
Hints
- Which operations in box regression require positive dimensions?
- Check transformed annotations before tuning the optimizer.
Answer
Inspect annotations and transformed boxes for nonpositive width or height, invalid coordinates and aggressive augmentations of tiny objects. Box transforms often use division or logarithms, making invalid geometry dangerous. Then inspect learning rate, warm-up, precision and gradient norms. Clipping gradients does not repair broken labels; reproduce the first failing batch before masking symptoms.
Source and licence: OpenMMLab — MMDetection FAQ · Apache-2.0 · answer: Quantoros editorial