alireza-nasiri / soundclr Goto Github PK

I got a question. You specify dim=0 for every F.normalize(). Is it normalizing along with the batch size axis? The shape of y_rep is B x 2048. If you want to normalize each embedding on a unit hypersphere, should it be through dim=1?

Hello @alireza-nasiri ，
Thank you for sharing such excellent code .
May I ask you a question？
‘ModuleNotFoundError: No module named 'sgmllib'’ error when I run code，Looking up the information, we know that the package above python3 is gone，How to solve this problem？

https://github.com/alireza-nasiri/SoundCLR/blob/main/loss_fn/hybrid_loss.py#L15

def cross_entropy_one_hot(input, target):
_, labels = target.max(dim=1)
return nn.CrossEntropyLoss(weight=class_weights)(input, labels).

What is weight=class_weights for different datasets

The results on US8K with Cross-Entropy Loss is 86.16, but I am getting 78.01. I am training with these specifications:
BASELINE
transfer - augmentation on both waves and specs - 3 channels
US8K
train folds are [2, 3, 4, 5, 6, 7, 8, 9, 10] and test fold is [1]
number of freq masks are 2 and their max length is 32
number of time masks are 1 and their max length is 32

@alireza-nasiri
Hello，
Sorry to disturb you again, I would like to ask if your code can be used on other data sets? (I am very sorry to tell you that my code level is limited, and I have not understood part of your code.) I understand that your data set is converted into a spectrogram for operation. I also have a processed spectrogram here. Form of data set. So what should I do next to run successfully? I would be very grateful if you could put forward some suggestions and methods.