How STFT Discriminator Output Logits are used in Hinge Loss?

[steventan0110]

❓ Questions

Hi authors, thanks for open-sourcing the project code as well as your implementation of STFT discriminator. I was able to train a good codec model with your STFT implementation.

However, I see that the output logits from the discriminator have the shape (b c t w), b for batch size, c for #channels, t for timesteps/frames, and w for frequency bins/dimension. In order to compute the hinge loss, D(x) should return a shape of (b c) or (b c t) right, and I am wondering how you aggregate the information from the frequency bins/dimension (the last dimension)?

My current implementation directly sums over the last dimension and I suppose I could also use a nn.Linear layer to map dimension from w->1. I'm wondering if you could provide more details on how the discriminator logits are used during training, thanks!