yinboc/trans-inr

Transformers as Meta-Learners for Implicit Neural Representations, in ECCV 2022

trans-inr's Issues

Question about hyponet

Hi !

Congratulations on publishing a nice paper! I also found your code to be very useful, and thank you for opening the source code.

I wanted to ask if I want to use a custom hyponet, like a feature grid based INR, would I just add a class file in the models/hyponets folder or do I need to change anything else too? Also while training the Feature grid INR standalone I am using kaolin-wisp trainers, is there a way to integrate custom trainers in the code

Transformer Encoder Output

Hey, thanks for sharing the code.

I am a bit confused regarding the following:

trans-inr/models/trans_inr.py

Line 65 in f4bdc01

w = F.normalize(w * x.repeat(1, 1, w.shape[2] // x.shape[2]), dim=1)

So in essence, you are using Transformer Encoder output (after being passed through a FC layer) to modulate randomly initialized weights of INR, right?

test with SIREN

Nice job! I notice that you still use ReLU MLP with PE. A MLP architecture named SIREN, which replaces the ReLU activation to sine activation, has a better reprenentation ability. Have you even tested SIREN as your hyponet?

Add einops to dependencies

Thanks for great and interesting work.

I think you should add the dependence on einops in the README.md.

Question about the experimental detail

Hi !

Congratulations on publishing a nice paper! I also found your code to be very useful, and thank you for opening the source code.

While reading your paper, I was wondering about the effect of using view direction in NeRF architecture, where it seems like you did not use the view direction for the NeRF architecture in the main experiment (am I correct..?).

During the development, have you had a chance to use the NeRF architecture with view directions?

Thanks again for presenting a nice paper!!

Best,
Jihoon