Ezio77 / DINER

DINER
MIT License
47 stars 3 forks source link

How to interpolate the coords #6

Open XiaoqiangZhou opened 10 months ago

XiaoqiangZhou commented 10 months ago

Thanks for sharing the great work.

A question is that how to test with the interpolated coords. For example, the model is overfitted with an input image with size 100x100. I want to reconstruct the image with 200x200 resolution from the trained models. I find that the hash mode hinders the feasibility. When turn off the hash mode, the performance drops greatly.

So, could you please tell me how to test with an arbitrary image size (different from the training size). Is it a limitation of this work?

Best regards,

grondilu commented 9 months ago

Apparently there are _interp variants to the network architecture classes, though they seem to require their own training, so you have to pick the maximum resolution you intend to use.

That is my understanding anyway.

PS. Also there is a recent paper from the same team which is all about interpolation of hash-table based INR : https://arxiv.org/abs/2309.12642