imirzadeh / Teacher-Assistant-Knowledge-Distillation

Using Teacher Assistants to Improve Knowledge Distillation: https://arxiv.org/pdf/1902.03393.pdf
MIT License
256 stars 47 forks source link

Optimized values #7

Closed TeerathChandani closed 5 years ago

TeerathChandani commented 5 years ago

Can you tell me, optimized value of T_student,lambda_student, seed?

Because I am facing issue while running nnctl command?

Thanks

imirzadeh commented 5 years ago

There were many experiments done. For example, you can verify these(-> means distillation):

Dataset: CIFAR-100
Accuracy of Resnet110: 72.38
Accuracy of Resnet20: 67.03
Accuracy of Resnet110 -> Resnet20: 68.11 [seed=55, lambda_student=0.5, T_student=5]
Accuracy of Resnet8: 61.37
Accuracy of Resnet110 -> Resnet8: 61.41 [seed=31, lambda_student=0.2, T_student=2]
Accuracy of Resnet110 -> Resnet20 -> Resnet8: 61.82 [seed=31, lambda_student=0.15, T_student=5]

TeerathKumar142 commented 5 years ago

Thanks for your quick response.

First

accuracy

TeerathChandani commented 5 years ago

Waiting for your response..! I am asking values, only because nnictl command does not work. That is why, I need your help.

Thanks

imirzadeh commented 5 years ago

@TeerathKumar142 The result I mentioned were not for this table and were for cifar100 dataset (table 3, row2 in the paper) For your first question: "It means, you made multiple combinations of those 3 values of each hyper-parameters.. then you have found that which are optimized.? Second" , the answer is Yes and No. Teacher Assistant Knowledge Distillation is a multi step process. The first step is to distillate knowledge from teacher to TA and the second step is to distillate knowledge from TA to the student. So, it is basically doing knowledge distillation 2 times and each time we do it with hyper parameters.

For your second question, I have to check them later(max in 2 days) because I'm currently busy doing another project.

TeerathChandani commented 5 years ago

Can you share your email? I have to discuss regarding this project personally?

TeerathChandani commented 5 years ago

Can you share the weights of the teacher (CIFAR10 and CIFAR100) only? If it is possible?

Or just share all hyperparameter values that you used in the experiement.?

imirzadeh commented 5 years ago

For table 4 of the paper, here are the hyper-parameters for the experiment and the final weights.

First Row(Resnet 26 Teacher, Resnet 20 TA, Resnet 14 student)

step1: resnet26 -> resnet 20 lambda_student: 0.95
seed: 31
T_student: 5

step2: Resnet20 -> Resnet 14:
lambda_student: 0.97
seed: 55
T_student: 10

Weight File of the final student(resnet14 with accuracy 91.23):
https://www.dropbox.com/s/41f4eamebqhh68b/resnet_26_20_14_2572043ab83f4878bb3c17560fa45f3d_best.pth.tar?dl=0

Second Row(Resnet 26 Teacher, Resnet 14 TA, Resnet 8 student)
step1: Resnet 26-> Resnet 14
lambda_student: 0.95
seed: 31
T_student: 5

step2: Resnet 14 -> Resnet 8
lambda_student: 0.49
seed: 55
T_student: 10

Weight file of the final student(resnet 8 with the accuracy of 88.01):
https://www.dropbox.com/s/ulbdftjf43p21c5/resnet_26_14_8_4c3b4d0befca4633bb1265023d14e1ff_best.pth.tar?dl=0

TeerathChandani commented 5 years ago

Thanks you so much..!

On Sun, 4 Aug 2019 at 04:23, Iman Mirzadeh notifications@github.com wrote:

Closed #7 https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7 .

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7?email_source=notifications&email_token=AMF5TEXE4NIHOEED7CKVS33QCXLLJA5CNFSM4IHQOUXKYY3PNVWWK3TUL52HS4DFWZEXG43VMVCXMZLOORHG65DJMZUWGYLUNFXW5KTDN5WW2ZLOORPWSZGOS3RNSIA#event-2531449120, or mute the thread https://github.com/notifications/unsubscribe-auth/AMF5TEQYV6JZC5Y4Y7ZP7J3QCXLLJANCNFSM4IHQOUXA .

TeerathChandani commented 5 years ago

Hello Imirzadeh,

Thanks for your help and quick response. Can you share the weights of table 3?

Thanks, Teerath Kumar

On Sun, 4 Aug 2019 at 23:21, teerath kumar teerathkumar142@gmail.com wrote:

Thanks you so much..!

On Sun, 4 Aug 2019 at 04:23, Iman Mirzadeh notifications@github.com wrote:

Closed #7 https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7 .

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7?email_source=notifications&email_token=AMF5TEXE4NIHOEED7CKVS33QCXLLJA5CNFSM4IHQOUXKYY3PNVWWK3TUL52HS4DFWZEXG43VMVCXMZLOORHG65DJMZUWGYLUNFXW5KTDN5WW2ZLOORPWSZGOS3RNSIA#event-2531449120, or mute the thread https://github.com/notifications/unsubscribe-auth/AMF5TEQYV6JZC5Y4Y7ZP7J3QCXLLJANCNFSM4IHQOUXA .

TeerathChandani commented 5 years ago

Secondly, I can not extract weight files. Please send me .zip instead of .rar It is showing me error of "operation break".

Thanks,

On Tue, 6 Aug 2019 at 09:16, teerath kumar teerathkumar142@gmail.com wrote:

Hello Imirzadeh,

Thanks for your help and quick response. Can you share the weights of table 3?

Thanks, Teerath Kumar

On Sun, 4 Aug 2019 at 23:21, teerath kumar teerathkumar142@gmail.com wrote:

Thanks you so much..!

On Sun, 4 Aug 2019 at 04:23, Iman Mirzadeh notifications@github.com wrote:

Closed #7 https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7 .

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7?email_source=notifications&email_token=AMF5TEXE4NIHOEED7CKVS33QCXLLJA5CNFSM4IHQOUXKYY3PNVWWK3TUL52HS4DFWZEXG43VMVCXMZLOORHG65DJMZUWGYLUNFXW5KTDN5WW2ZLOORPWSZGOS3RNSIA#event-2531449120, or mute the thread https://github.com/notifications/unsubscribe-auth/AMF5TEQYV6JZC5Y4Y7ZP7J3QCXLLJANCNFSM4IHQOUXA .

TeerathChandani commented 5 years ago

Ok.. Solved.. Thanks ..

On Tue, 6 Aug 2019 at 09:28, teerath kumar teerathkumar142@gmail.com wrote:

Secondly, I can not extract weight files. Please send me .zip instead of .rar It is showing me error of "operation break".

Thanks,

On Tue, 6 Aug 2019 at 09:16, teerath kumar teerathkumar142@gmail.com wrote:

Hello Imirzadeh,

Thanks for your help and quick response. Can you share the weights of table 3?

Thanks, Teerath Kumar

On Sun, 4 Aug 2019 at 23:21, teerath kumar teerathkumar142@gmail.com wrote:

Thanks you so much..!

On Sun, 4 Aug 2019 at 04:23, Iman Mirzadeh notifications@github.com wrote:

Closed #7 https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7 .

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7?email_source=notifications&email_token=AMF5TEXE4NIHOEED7CKVS33QCXLLJA5CNFSM4IHQOUXKYY3PNVWWK3TUL52HS4DFWZEXG43VMVCXMZLOORHG65DJMZUWGYLUNFXW5KTDN5WW2ZLOORPWSZGOS3RNSIA#event-2531449120, or mute the thread https://github.com/notifications/unsubscribe-auth/AMF5TEQYV6JZC5Y4Y7ZP7J3QCXLLJANCNFSM4IHQOUXA .

TeerathChandani commented 5 years ago

Hello Imirzadeh,

Thanks for your help and quick response. Can you share the weights of table 3?

Thanks

On Tue, 6 Aug 2019 at 11:35, teerath kumar teerathkumar142@gmail.com wrote:

Ok.. Solved.. Thanks ..

On Tue, 6 Aug 2019 at 09:28, teerath kumar teerathkumar142@gmail.com wrote:

Secondly, I can not extract weight files. Please send me .zip instead of .rar It is showing me error of "operation break".

Thanks,

On Tue, 6 Aug 2019 at 09:16, teerath kumar teerathkumar142@gmail.com wrote:

Hello Imirzadeh,

Thanks for your help and quick response. Can you share the weights of table 3?

Thanks, Teerath Kumar

On Sun, 4 Aug 2019 at 23:21, teerath kumar teerathkumar142@gmail.com wrote:

Thanks you so much..!

On Sun, 4 Aug 2019 at 04:23, Iman Mirzadeh notifications@github.com wrote:

Closed #7 https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7 .

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation/issues/7?email_source=notifications&email_token=AMF5TEXE4NIHOEED7CKVS33QCXLLJA5CNFSM4IHQOUXKYY3PNVWWK3TUL52HS4DFWZEXG43VMVCXMZLOORHG65DJMZUWGYLUNFXW5KTDN5WW2ZLOORPWSZGOS3RNSIA#event-2531449120, or mute the thread https://github.com/notifications/unsubscribe-auth/AMF5TEQYV6JZC5Y4Y7ZP7J3QCXLLJANCNFSM4IHQOUXA .