Alphas are not trainable?

[PiotrDabkowski]

I have noticed that alphas is not a ModuleList, but rather a list of parameters, this causes them to be not trainable.

Example: self.alpha1 = [nn.Parameter(torch.ones(1, channels, 1).to(rank)) for i in range(len(self.convs1))]
Is this expected?

[sh-lee-prml]

Thank you...

I have mis-implemented this parameter...

I'll fix it right now.

Thanks again

[sh-lee-prml]

self.alpha1 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs1))])

I changed alphas to ParameterList

https://github.com/sh-lee-prml/BigVGAN/blob/main/models_bigvgan.py#L51
https://github.com/sh-lee-prml/BigVGAN/blob/main/models_bigvgan.py#L52
https://github.com/sh-lee-prml/BigVGAN/blob/main/models_bigvgan.py#L100
https://github.com/sh-lee-prml/BigVGAN/blob/main/models_bigvgan.py#L102
https://github.com/sh-lee-prml/BigVGAN/blob/main/models_bigvgan.py#L108

Now, alpha is trainable 😢

Thank you again👍

[HaiFengZeng]

@sh-lee-prml alpha need to be greater than 0 as a frequency parameter? maybe add a torch.exp(alpha) to make sure it positive?

[sh-lee-prml]

Hi @HaiFengZeng

The official code of snake1d initializes it greater than 0 by abs()

a = torch.zeros_like(x[0]).normal_(mean=0,std=50).abs()

But, I think it does not need to be greater than 0. The negative alpha value just has an inverted phase, and we used the squared value as below

x = (x + (torch.sin(self.a * x) ** 2) / self.a)