Loss curves on ImageNet about moco HOT 4 CLOSED

Note: if your other dataset is too small and still use the default queue size of 65536, you should revise some code to make things right. Basically, if the dataset is just 5k, the queue contains >10x of the dataset size, which means it contains >10 positive samples. These positive samples should be removed for the loss to make sense. This is not a problem on ImageNet with 1.28M images, as the queue is negligibly small. But this should be handled with some care for smaller sets.

from moco.

The initial increase of loss is because the queue is being filled, and the task is getting harder when real features replace initial noise in the queue. Other than that, we do not have enough information to diagnose your case as you used a different dataset. In the latest version of our arXiv paper (Appendix A.9), there is a curve on ImageNet for your reference, and the ImageNet curves can be obtained if you run our code following our settings.

from moco.

I met same problem, but in a different dataset(not image dataset). But I tried only use pasal voc 07 dataset, only have 5000 images and do not have this problem. So I think this problem is not related to dataset size

from moco.

I met same problem, but in a different dataset(not image dataset). But I tried only use pasal voc 07 dataset, only have 5000 images and do not have this problem. So I think this problem is not related to dataset size

Hi, I tried the same dataset, it didn't work, what was your queue size? and how many epochs did you see convergence?

from moco.