result problem

[caibajin]

It is possible to ask whether the results of ViT-B on the SBD training set were trained with 55 epochs or 230 brother epochs.
I ran 55 epochs with the following results

[qinliuliuqin]

@caibajin Hi, all ViT-B models were trained with 55 epochs. It seems your results are much worse than ours. Did you use the same training settings as specified here: https://github.com/uncbiag/SimpleClick/blob/v1.0/models/iter_mask/plainvit_base448_sbd_itermask.py? Meanwhile, our batch size is 140 for ViT-B models. Let me know your training settings.

[caibajin]

device：4090
batch-size：24

[caibajin]

Hi, I'd like to ask if it's related to batch-size?

[qinliuliuqin]

@caibajin It's less likely that using a batch size of 24 instead of 140 would greatly impact the performance, though you can have a try if that's the only difference between your training settings and mine. BTW, I used 4 A6000, but 4090 should be able to reproduce the results.