problem of the model.

[ooobsidian]

@seongmin-kye Hello, The full connection layer in the model is not defined at initialization time:

meta-SR/model/resnet_256.py

Line 157 in b4c1ea1

x = self.fc(x)

In Youngmoon Jung's repo, I noticed that in similar model code, he defined self.fc, and I guess you and Youngmoon Jung are from the same lab, so did you forget to define it here?

https://github.com/jymsuper/SpeakerRecognition_tutorial/blob/6dce64657632660a791e25624e4257c53c7257d6/model/resnet.py#L117

And in this case, the self.weight defined in the initialization is not used in the forward, where includes num_classes.

meta-SR/model/model.py

Line 31 in b4c1ea1

self.weight = Parameter(torch.Tensor(num_classes, 256))

So I could use any number of n_classes in testing and have poor performance:

meta-SR/EER_full.py

Line 37 in b4c1ea1

n_classes = 5994 if args.data_type == 'vox2' else 1211

Is there anything wrong with what I mentioned?
Looking forward to your reply.