Penn Jenks

Feature Request: Add default parameter values to docstring

5

Currently the docstrings for `Module` subclasses don't display the default argument values when using `?`. Not sure if there is an automated way to display this, or to update all...

Feature Request: add attention utility functions like flax

3

It would be nice to split up the the [MultiheadAttention](https://github.com/patrick-kidger/equinox/blob/main/equinox/nn/attention.py#L14) Module code into utility functions for calculating attention weights and multihead attention without projection layers. Flax does this -- but...

Rewriting broadcasted functions doesn't work

1

The following works as expected ```julia julia> r = @rule a b log(a / b) --> log(a) - log(b) julia> r(:(log(x / y))) :(log(x) - log(y)) ``` However I'm interested...

Need adjoint for reinterpret SVector

4

Minimum working example: ```julia f(x) = sum(sum(reinterpret(SVector{size(x, 1), eltype(x)}, x))) Zygote.gradient(f, rand(3, 10)) ``` ``` ERROR: Need an adjoint for constructor Base.ReinterpretArray{SVector{3, Float64}, 2, Float64, Matrix{Float64}, false}. Gradient is of...

Lazy Arrays with CUDA.jl (feature request)

5

This is a feature request to get LazyArrays to work with CuArrays (and potentially other GPUArray implementations) Some of the features of LazyArrays works with CuArrays out of the box,...

KNN Implementation

6

I'm interested in a fast GPU implementation for k nearest neighbor queries -- similar to [pytorch3d.ops.knn_points](https://pytorch3d.readthedocs.io/en/latest/modules/ops.html?#pytorch3d.ops.knn_points). Flux3D seems like an appropriate place for this to live. Relevant pytorch3d code: [pytorch3d/csrc/knn](https://github.com/facebookresearch/pytorch3d/tree/master/pytorch3d/csrc/knn)...

Why uniform bias initializer?

14

I noticed that the bias parameters in linear and conv modules use a uniform initializer. Is there a good justification for this? I noticed that PyTorch does this to. I...

Can we get a new release that includes the OpenGL rasterization?

Check out my PyTorch version that implements the `MultiresHashEncoding`: https://github.com/jenkspt/multires-hash-encoding-pytorch And the NeuS fork that replaces the NeRF-like modules with `MultiresHashEncodings`: https://github.com/jenkspt/NeuS/tree/hash And discussion here: https://github.com/Totoro97/NeuS/issues/65

- Adapt jax code from [sarafridov/plenoxels](https://github.com/sarafridov/plenoxels) - Replace the plenoxel sparse grid with `HashArray` of - Replace spherical harmonic `eval_sh` with MLP _The instant neural graphics nerf training is more...