Customizing hash encoding initialization in PyTorch

[WanqiYuan]

Hi there,

Thanks for releasing such an amazing library.

Is it possible to customize the hash grid embedding in PyTorch? Such as using a customized weight for linear interpolation or use a customized resolution list (like a list with random resolutions for different level)?

Thanks!

[Tom94]

Hi, anything configurable via the JSON config can be configured in PyTorch (by passing the appropriate dict to the constructor of the embedding). But everything beyond that needs modification of the C++/CUDA code.

So for your concrete question: you can change the interpolation to nearest neighbor selection or to a smoothstep variant that has continuous first derivatives, but nothing else. You also can't customize the resolution progression of the levels from within PyTorch.