facebookresearch / facebookresearch/moco
Issue with batch size
Open
- Dominant language
- No language data
- Stars
- 5.1k
- Forks
- 802
- PR merge metrics
- No merged PRs in 30d
Description
Hello @KaimingHe, I am running the MoCo pre-training with the following configurations:
Num of GPUs = 8
GPU type: NVIDIA Quadro RTX 8000
But I am unable to run this with a batch size greater than 8. It throws a PyTorch Spawn error when I set the batch size greater than 8 (ex:16, 32). I confirmed that all GPUs are running at full utilization when I run the pre training with the above mentioned batch size. Do you have any suggestions or thoughts on how I might proceed? With a batch size this small, would it provide any meaningful results?
Contributor guide
Assessment
This issue has not been assessed yet.