Some Clarifications on the Running configurations for Stage1 and Stage2

[varun-krishnaps]

1. For Stage1 : In the paper it is mentioned that batch size used is 256 and learning rate decays by 5% every 5 epochs, but in the code batch size is 300 and learning rate decays by 10% every 5 epochs
2. **For Stage2 ** : In the code "main_train.py" of Stage2, it is mentioned that gates = [1, 3, 3, 5, 6] # Set the gates in each iterations, which is different from our paper because we use stronger augmentation in dataloader

What should be the right the configurations for stage1 and stage2 to reproduce the results?
Currently in the stage1, my EER doesn't go below 8.13

[Tao Ruijie]

1. I train the model for these two settings, they get the same results with EER = 7.36, in my experiments that did not affect the performance.

2. I remember in this code, [1,3,3,5,6] is slightly better so I put in for open-source version after publication.

[varun-krishnaps]

Hi Tao,

Can you please share the model of Stage 2 just like you did for Stage 1 ??

Thank you

[Tao Ruijie]

https://drive.google.com/file/d/1VLTbAPYwDV7YJPkLz-4pJbAPkGre1YvZ/view?usp=sharing

Here is it..

[varun-krishnaps]

Hi Tao,
The model you shared in the link seems to be corrupted....while i load it using the i get the error RuntimeError: PytorchStreamReader failed reading zip archive: failed finding central directory.

Can u check your model file again ??