Low accuracy on trillionpairs

[tranvanhoa533]

Hello @nttstar
I train LResNet100E-IR with emore + asian dataset on 1080Ti gpu. The accuracy on trillionpairs challenge is very low:

         m         per_batch_size    num_gpu      batch_size     identification         verification
  1     0.3               62           8              496              71%                  70.6%
  2     0.5               62           7              434              55.4%                55%
  3     0.5               180          6              1080             29.7%                27%

(In third experiment, I used mxnet-memonger to decrease memory).

I am very puzzled. Can you point out my mistake, please ? Thank you very much.

[nttstar]

Make sure the merging process is correct. It should get 80%+ easily with m=0.5 arcface loss.

[tranvanhoa533]

I downloaded glint asian and emore dataset from the link which you shared. After that, i run dataset_merge.py to merged them.

[tranvanhoa533]

Hi @nttstar
I did another experiment with your default parameters ( per-batch-size = 128, num-gpu = 4, r100, m=0.5 arcface loss , dataset: emore+asian) but the result on trillionpairs challenge was only 61.19% (Identification) and 59.17% (Verification). I tried to duplicate your experiment but i could not achieve your result (84%). Did you clean the dataset when you merged them ? Did you have other techniques ? Can you share them with me, please ? Thank you very much !

[jeremmyzong]

Make sure the merging process is correct. It should get 80%+ easily with m=0.5 arcface loss.

Hi @nttstar
Did you use pretrained model to deduplicate the dataset in merging process?
Thanks!

[jeremmyzong]

Hi @tranvanhoa533
I met the same problem, and I did experiment with bigger batch_size(per-batch-size = 128, num-gpu = 8), only get 43%(id) and 39%(verification). Have you solved it?

[tranvanhoa533]

Hi @jeremmyzong
I still don't understand why that is. Have you solved it ?

[Talgin]

Hi @nttstar,
Could you provide an example command to run dataset_merge.py?
I'm running it second time, because after first time when trying to merge faces_emore and faces_glint it returned similar to faces_emore .idx, .rec, and property files. I mean, the size of those files were the same and property file said => 85772, 112, 112.

Thank you in advance!

[eguoguo321]

how to use mxnet-memonger，when I enabled that to be true, it met error.
what else should I do?

[maywander]

Did you merge these two datasets successfully?@Talgin

[Talgin]

Hi @maywander ,
As an answer you can look this thread #256
Thank you! :)

[maywander]

Thanks for your reply. I've solved the problem. At 2019-10-09 17:09:48, "Talgin" <notifications@github.com> wrote: Hi @maywander , As an answer you can look this thread #256 Thank you! :) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.