Periodic Activation Functions about inr-gan

[betterze]

Hi, Ivan,

Thank you for your detailed reply. The literatures you provided are very helpful for me.

Best Wishes,

Alex

[universome]

Hi! We did use them in our original implementation. In the current one, we switched to the manually constructed fourier features instead of the static-random/predictable-from-latent-code ones (as suggested in the SIREN/FourierFeats papers) since 1) they are "guaranteed" to cover all reasonable directions and frequencies and 2) faster because they have much smaller dimensionality (for predictable fourier features, the situation is even worse since they are uncachable even at test time).

However, all the above ways to encode coordinate positions are supported in the current repo: one just needs to specify the desired embedding size for each type of the positional encoding: https://github.com/universome/inr-gan/blob/master/src/training/layers.py#L113-L117

[betterze]

Hi, Ivan,

Thank you for your detailed answer. I am new to this filed, correct me if I am wrong. Siren is not just about encoding coordinate positions in term of fourier feature, but also using activation function as 'sine'. Did you try to use 'sine' activation as mentioned in siren? I am just want to understand if 'sine' activation can work for large dataset (ffhq) for implicit neural representation with hypernetwork.

And a more general question, for implicit neural representation with hypernetwork (or continuous images), what is it advantage of standard generator (stylegan ada) except fast inference and super resolution?

I am really new to this field, your answer is very helpful for me. Thank you again for your answer.

Best Wishes,

Alex

[universome]

Ah, I see what you mean. You are asking whether we tried using sine activations everywhere throughout our INR decoder and not only for positional embeddings. We did try doing this in our preliminary experiments, but found that it does not perform well and also makes the whole pipeline more complicated (it becomes a one more thing you need to keep in mind when debugging things, and we already had enough things to keep in mind) + FourierFeats paper showed that you can have sine activations only for the first layer in INRs (i.e. to encode coordinates). We didn't provide any ablations on the activation function in the final paper, but there is an ablation on this in a parallel work by Anokhin et al (see Table 3) — it decreased the scores for them.

The main advantage of INR-based decoders is that they are supposed to understand the underlying image geometry much better. Check the gifs in the end of this old article here and a recent super cool INR-based GAN by Karras et al. here.

[universome]

Feel free to ask if you'll have any further questions!