Redimnet Large Margin not improving the results

[GoldMan6]

Hi, I am trying to reproduce the redimnetB2 results. I trained the model and got on Vox-O: 0.7% (no LM, no ASNORM), 0.61% (no LM, with ASNORM), 0.56 (no LM, with ASNORM and calibration). when I continue to large margin training, the results i get is: 0.68% (with LM, no ASNORM), 0.58% (with LM, with ASNORM), 0.57 (with LM, with ASNORM and calibration)
also in another training i did, the results got worse. anyone had the same issues ?
Thanks a lot.

[leehyun22]

@GoldMan6 , Hello, I have a question.
Did you use the config file written in the repo for the config you used for training?

[GoldMan6]

Hi @leehyun22 thanks for the quick response! yes i did, changed the seed and batch size.

[JiJiJiang]

After LM, no asnorm and with asnorm both get better, while with asnrom and calibration, it gets slightly worse.
I think the results are normal. Overall speaking, it gets better.

[GoldMan6]

Hi @JiJiJiang, when checking the pre-trained checkpoints of redimnet, the results are:
B2:
w/o LM: 0.77 ; w LM: 0.57
B4:
w/o LM: 0.71 ; w LM: 0.51
So it seems that the improvement is a lot bigger, not like the results that i got when training with LM. also on other experiment the results was worse when using the LM training. any thoughts ?
Thanks