What does an 8xb1 segmentation config tell you, and can you reproduce it unchanged on one GPU?
Computer VisionLevel: IntermediateReproducing distributed segmentation recipes
What does an 8xb1 segmentation config tell you, and can you reproduce it unchanged on one GPU?
Hints
- A recipe's hardware count is not always a minimum requirement.
- Which training statistics depend on the global batch?
Answer
It describes the reference recipe's eight GPUs with one sample per GPU, not a proof that the model requires eight devices to run. On one GPU, effective batch, optimization schedule and normalization behavior may change. Account for gradient accumulation and learning-rate choices, and report any differences from the reference recipe rather than claiming exact reproduction.
Source and licence: OpenMMLab — MMSegmentation FAQ · Apache-2.0 · answer: Quantoros editorial