performance on Voxceleb1?

[linjucs]

How about the performance

[linjucs]

Just notice this paper https://arxiv.org/pdf/1812.00271.pdf

[linjucs]

Hi Mirco,

Thanks for sharing this code. Very cool toolkit. Recently I am working on Sincnet on voxceleb1. But the performance is not well. Could you please give me some suggestions? I removed the silence using webrtcvad.

Thanks. Lin

epoch 0, loss_tr=6.254185 err_tr=0.976379 loss_te=6.414922 err_te=0.987285 err_te_snt=0.973458 epoch 8, loss_tr=3.098647 err_tr=0.626692 loss_te=6.009328 err_te=0.905808 err_te_snt=0.778572 epoch 16, loss_tr=2.121552 err_tr=0.451692 loss_te=5.882671 err_te=0.857070 err_te_snt=0.673373 epoch 24, loss_tr=1.655752 err_tr=0.362991 loss_te=5.679876 err_te=0.823700 err_te_snt=0.611926 epoch 32, loss_tr=1.369494 err_tr=0.305188 loss_te=5.745930 err_te=0.811133 err_te_snt=0.595928 epoch 40, loss_tr=1.167458 err_tr=0.263594 loss_te=5.790279 err_te=0.797424 err_te_snt=0.579203 epoch 48, loss_tr=1.023051 err_tr=0.233562 loss_te=5.812299 err_te=0.780975 err_te_snt=0.566356 epoch 56, loss_tr=0.911310 err_tr=0.210032 loss_te=5.868678 err_te=0.771352 err_te_snt=0.541268 epoch 64, loss_tr=0.828135 err_tr=0.192578 loss_te=5.780817 err_te=0.767120 err_te_snt=0.550358 epoch 72, loss_tr=0.761099 err_tr=0.177930 loss_te=6.150887 err_te=0.759397 err_te_snt=0.538965 epoch 80, loss_tr=0.707083 err_tr=0.166238 loss_te=5.915407 err_te=0.758966 err_te_snt=0.542359

[mravanelli]

Hi, one risk with SincNet is overfitting and apparently you are observing this issue as well. SincNet can tune its filters according to the training conditions and if such conditions are very different from test ones you can see overfitting. A good property, however, is it possible to adapt SincNet very quickly to the new domain if you have adaptation data (this is due to the drastically reduced number of parameters). Unfortunately, this is not possible with VoxCeleb. To make a SincNet system working on this dataset, a lot of efforts should be devoted to avoid overfitting. In our case, we have achieved good performance by coupling SincNet with a semi-supervised strategy that acts as a powerful regularizer (see https://arxiv.org/abs/1812.00271). We are also working on other ideas to help SincNet fighting overfitting.

Best,

Mirco

On Thu, 8 Aug 2019 at 12:35, linjucs notifications@github.com wrote:

Hi Mirco,

Thanks for sharing this code. Very cool toolkit. Recently I am working on Sincnet on voxceleb1. But the performance is not well. Could you please give me some suggestions? I removed the silence using webrtcvad.

Thanks. Lin

epoch 0, loss_tr=6.254185 err_tr=0.976379 loss_te=6.414922 err_te=0.987285 err_te_snt=0.973458 epoch 8, loss_tr=3.098647 err_tr=0.626692 loss_te=6.009328 err_te=0.905808 err_te_snt=0.778572 epoch 16, loss_tr=2.121552 err_tr=0.451692 loss_te=5.882671 err_te=0.857070 err_te_snt=0.673373 epoch 24, loss_tr=1.655752 err_tr=0.362991 loss_te=5.679876 err_te=0.823700 err_te_snt=0.611926 epoch 32, loss_tr=1.369494 err_tr=0.305188 loss_te=5.745930 err_te=0.811133 err_te_snt=0.595928 epoch 40, loss_tr=1.167458 err_tr=0.263594 loss_te=5.790279 err_te=0.797424 err_te_snt=0.579203 epoch 48, loss_tr=1.023051 err_tr=0.233562 loss_te=5.812299 err_te=0.780975 err_te_snt=0.566356 epoch 56, loss_tr=0.911310 err_tr=0.210032 loss_te=5.868678 err_te=0.771352 err_te_snt=0.541268 epoch 64, loss_tr=0.828135 err_tr=0.192578 loss_te=5.780817 err_te=0.767120 err_te_snt=0.550358 epoch 72, loss_tr=0.761099 err_tr=0.177930 loss_te=6.150887 err_te=0.759397 err_te_snt=0.538965 epoch 80, loss_tr=0.707083 err_tr=0.166238 loss_te=5.915407 err_te=0.758966 err_te_snt=0.542359

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/mravanelli/SincNet/issues/56?email_source=notifications&email_token=AEA2ZVTWAB4QDFIX5EAHHUTQDRDM5A5CNFSM4IJCMVQKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD34GDFY#issuecomment-519594391, or mute the thread https://github.com/notifications/unsubscribe-auth/AEA2ZVTBEEEOVRNFHS7EBLTQDRDM5ANCNFSM4IJCMVQA .

[linjucs]

Thanks man. I will try to address overfitting problem.

[brealisty]

Thanks man. I will try to address overfitting problem.

I also met this problem, in my own data, could you give me some advices to address overfitting?

[tommyngx]

@linjucs can you please share your work with voxceleb1? my email is tommyfederation@gmail.com. I am trying to test it with vox1 as well. Do you try verification task? I don't know how to deal with that? thank you very much!