[Feature] Can we set the random seed in the mmdetection?

What’s the feature?

I know the random seed can be set in the distributed training by setting the argument seed to any fix number you want. However, there is little information to set the random seed in single GPU training, which may have impact on the model performance (±0.6 AP in my experiments), making hard to compare models in a fair way. I wonder whether to set the random seed in the single GPU training now ? If not, I think it is necessary to implement this to compare models in a fair way.

Any other context?

No response

[cpwan]

I found it in mmengine’s code

python ./tools/train.py /path/to/config.py --cfg-options randomness.seed=0 

The results for the same seed differ by a little.

When I add also randomness.deterministic=True, it gives me a cuda internal error 😦

[Agito555]

how about the 3.x version?

I haven’t used the 3.x version mmdetection yet, and I don’t see the argument seed in the train.py of 3.x mmdetection. I will reopen the issue, hope somebody else will offer some help.