Megvii-BaseDetection / Megvii-BaseDetection/YOLOX
can't reproduce the latency of yolox series
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 10.6k
- Forks
- 2.5k
- PR merge metrics
- No merged PRs in 30d
Description
Hi,
I'm trying to reproduce the latency and mAP of the proposed yolox series with V100 32G. However, using the evaluation scripts in this repo [python -m yolox.tools.eval -n yolox-s -c yolox_s.pth -b 1 -d 1 --conf 0.001 --fp16 --fuse], I've got 12.3ms for yolox-small with single gpu, fp16, batchsize=1, which is slower than the reported results in paper (9.6ms). And I'm also noticed fp32 (10.5ms) is faster than fp16(12.3ms), which is unexpected. I'm wondering, what is the proper setting to reproduce the reported results?
The environment setting of mine case is:
pytorch==1.7.1
cuda==10.1
cudnn==7.6.3
Any replies will be appreciated!
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the reported python -m yolox.tools.eval command for yolox-s, then compare its V100, batch-size, FP16, PyTorch, CUDA, and cuDNN settings with those used for the paper's 9.6ms result. Check why FP32 is faster than FP16 in this environment. Done means identifying the proper settings or explaining the latency discrepancy and mAP reproduction.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100