Hello, I have used the vox2 dataset(including 5994 speakers), and the default parameters.
However, the cost converges at around 19.50, and the speaker verfication result is not good.
I am wondering what's wrong with it, and what is your convergence cost. Thanks!

Excuse me, which paper is this？

Hi there

During training using the GhostVLAD layer, I have experienced that the cluster centres do not change their value, which means that they always stay at the same place during training. Is this an expected behaviour, or is this a bug in my training process?

To give you some context, I used the GhostVLAD layer after a ResNet18 architecture for an audio task. I adapted your exact code to fit my structure. All the initialization, layers and so on are like in your code.

I would appreciate your help 😄.

In the original paper, decaying step of learning rate is every 36 epochs.

But your code, the step is 5000 iterations.

How can I make the same configuration as in the original paper?

Thank you!

Hi taylorlu, I want to know the settings of the pretrained model. (epoch, batch size, learning rate decay...)

I used the model and it seems works nicely.

num of workers: 24

After iteration of 426x, all processes in iterator() function seems to be stopped.

But the next epoch won't start.