search

YirongMao / NAN

This is an unofficial PyTorch 1.0.1 implementation of the papr Neural Aggregation Network for Video Face Recognition. CVPR 2017
MIT License
21 stars 5 forks source link

Some questions about NAN #3

Closed JingY0604 closed 3 years ago

JingY0604 commented 4 years ago

hi：     thank you for sharing your code of NAN, but, I have some questions about it:     

  1. Is there any other way to get pre-trained model and data?  like google drive link?     
  2. How train list and test list generated？     
  3. why face features are extracted by ResNet34 with only SoftMax loss, not CosoNet best model? thanks!
YirongMao commented 4 years ago

hi：     thank you for sharing your code of NAN, but, I have some questions about it:     

  1. Is there any other way to get pre-trained model and data?  like google drive link?     
  2. How train list and test list generated？     
  3. why face features are extracted by ResNet34 with only SoftMax loss, not CosoNet best model? thanks!

Thanks for your interest. The purpose of this project is to reproduce the NAN algorithm, NOT to reach top performance. I follow the published protocol of IJB-A to split training and testing dataset. BTW, Sorry for my late response. If you have any other problem, be free to contact me.

JingY0604 commented 4 years ago

hi：     thank you for sharing your code of NAN, but, I have some questions about it:     

  1. Is there any other way to get pre-trained model and data?  like google drive link?
  2. How train list and test list generated？
  3. why face features are extracted by ResNet34 with only SoftMax loss, not CosoNet best model? thanks!

Thanks for your interest. The purpose of this project is to reproduce the NAN algorithm, NOT to reach top performance. I follow the published protocol of IJB-A to split training and testing dataset. BTW, Sorry for my late response. If you have any other problem, be free to contact me.

Thank you for your answer. but，I have another question，why fixed the size of image set during training？

hi：     thank you for sharing your code of NAN, but, I have some questions about it:     

  1. Is there any other way to get pre-trained model and data?  like google drive link?
  2. How train list and test list generated？
  3. why face features are extracted by ResNet34 with only SoftMax loss, not CosoNet best model? thanks!

Thanks for your interest. The purpose of this project is to reproduce the NAN algorithm, NOT to reach top performance. I follow the published protocol of IJB-A to split training and testing dataset. BTW, Sorry for my late response. If you have any other problem, be free to contact me.

Thank you for your answer. but，I have another question，why fixed the size of image set with 8 during training？

YirongMao commented 4 years ago

hi：     thank you for sharing your code of NAN, but, I have some questions about it:     

  1. Is there any other way to get pre-trained model and data?  like google drive link?
  2. How train list and test list generated？
  3. why face features are extracted by ResNet34 with only SoftMax loss, not CosoNet best model? thanks!

Thanks for your interest. The purpose of this project is to reproduce the NAN algorithm, NOT to reach top performance. I follow the published protocol of IJB-A to split training and testing dataset. BTW, Sorry for my late response. If you have any other problem, be free to contact me.

Thank you for your answer. but，I have another question，why fixed the size of image set during training？

hi：     thank you for sharing your code of NAN, but, I have some questions about it:     

  1. Is there any other way to get pre-trained model and data?  like google drive link?
  2. How train list and test list generated？
  3. why face features are extracted by ResNet34 with only SoftMax loss, not CosoNet best model? thanks!

Thanks for your interest. The purpose of this project is to reproduce the NAN algorithm, NOT to reach top performance. I follow the published protocol of IJB-A to split training and testing dataset. BTW, Sorry for my late response. If you have any other problem, be free to contact me.

Thank you for your answer. but，I have another question，why fixed the size of image set with 8 during training？

I tuned the image set size, and found 8 is the best.