ReID accuracy dropped when training with smaller batch sizes

[tranhoanglam2712]

Dear the authors of PetFace,

I have experienced a significant drop in accuracy (70% -> 28% for cats) when I was training the unified model. The only config that I changed was the batch size (256 -> 64) due to my hardware restrictions. This issue did not appear when I was training on single species with the above config. Also, I have been able to reproduce the results loading the pretrained unified models.

Do you have any suggestions on this issue?

Cheers,
Hoang Lam Tran

[dahlian00]

Hi, could you try a batch size of 256 for debugging?

[dahlian00]

We overlooked that you have hardware restrictions.
We don’t think this significant drop happened only because of the batch size, but we can’t be sure about that.
We hope you can try using a batch size of 256 for debugging.
If we find out any additional information, we will reply again.