Question about pre-trained model

[jlian2]

This is an amazing model! However, when I was training using your code given, I could not achieve a good result as yours. I notice your "dev eer" is about 0.0%? But mine is much worse(much higher). I am really interested in this model and I am recently doing another ASV task. Do you have any pretrained model avaliable? Really really amazed at your result.

[wangziqi000]

Hi,

The zero EER happens if we use a fused score calculated on dev dataset. Some of the pretrained model are collected along the way of training that may not have converged. You can continue training on one of our check points, or restart a new model.

[Yolanda-Gao]

Hi Ziqi,

I trained the MFCC model and it achieved around 7% EER for dev dataset on MFCC features and around 4% EER on spectrogram. I did change the feature parameters such as FTT size and use fewer residual blocks, which may be the reason for the performance differences. So just to be sure, the near 0% EER is on the development dataset, right? Do you have any other advice for me to tune the parameters? Thank you!

[wangziqi000]

Hi,
Thank you for your interest. Theoretically, increasing number of FFT filters won't help much, as once the size of FFT is greater than the window length, it is guaranteed that no aliasing happens. I wound suggest you reduce the stride of the convolutional layers(e.g. use 1 or 2) and enable more blocks to see if the results improve. Also, I would like to remind that you can do the training on previous checkpoints so that you can use a gradually decreasing training rate.

Yes, I can confirm we do training on the training set and got very low EER on dev set. But on eval set, the performance is worse due to unknown type attacks.

[ddave25]

Hi.

I am having an issue with the pre-trained model for CQCC features (sample_model/CQCC/cqcc.pth)

When I first load it, I get the following size mismatch error:
size mismatch for fc1.weight: copying a param with shape torch.Size([128, 512]) from checkpoint, the shape in current model is torch.Size([128, 32]).

When I change the inputs of fc1 to 512, I then get the error that there are missing keys for block 10 and block 11. After commenting out these blocks, I get the following size mismatch error:
size mismatch, m1: [32 x 32], m2: [512 x 128]

I am not really sure how to fix this error. If I were to make changes to the architecture, I won't be able to successfully load the pre-trained model.

Has anyone else faced this error and been able to fix it?

[Yolanda-Gao]

Hi.

I am having an issue with the pre-trained model for CQCC features (sample_model/CQCC/cqcc.pth)

When I first load it, I get the following size mismatch error:
size mismatch for fc1.weight: copying a param with shape torch.Size([128, 512]) from checkpoint, the shape in current model is torch.Size([128, 32]).

When I change the inputs of fc1 to 512, I then get the error that there are missing keys for block 10 and block 11. After commenting out these blocks, I get the following size mismatch error:
size mismatch, m1: [32 x 32], m2: [512 x 128]

I am not really sure how to fix this error. If I were to make changes to the architecture, I won't be able to successfully load the pre-trained model.

Has anyone else faced this error and been able to fix it?

I met the same issue. And my own training of the CQCC model gives a much higher EER of ~10% on the eval set...

[potato-asuteed-potato]

Have you solved the problem yet? I ran into this problem when I was training the CQCC model,my own training of the CQCC model gives a much higher EER of ~40% on the eval set...