KernelFunctions.jl

KernelFunctions.jl copied to clipboard

Differentiating `FunctionTransform` with Zygote

1

I guess this might be another Zygote related issue: differentiating `FunctionTransform` doesn't work for multidimensional inputs: ```julia julia> x = rand(2, 100); julia> k(θ) = TransformedKernel(TransformedKernel(ExponentialKernel(), FunctionTransform(x->θ[1]*x)), SelectTransform([1])); julia> o(θ)...

molet

The parameter `r` in our `PeriodicKernel` currently governs the size of the circle onto which the inputs are mapped. Would it not be more intuitive to parametrise it in terms...

willtebbutt

https://travis-ci.com/github/JuliaGaussianProcesses/KernelFunctions.jl/jobs/382246276 I am not able to pin-point the issue as I do not use MacOS.

sharanry

We could make a nice connection between https://github.com/ysig/GraKeL and https://github.com/JuliaGraphs/LightGraphs.jl !

theogf

[WIP]Add a generic way to check kernel's input dimensions compatibility.

6

Related PR #111 Please find a generic way to deal with dimensions of inputs to kernels as @willtebbutt and @devmotion had suggested in #111 . I had to remove restrictions...

sharanry

Roadmap to 1.0

5

It would nice to have a first 1.x.x release. With the latest change it feels we have a first stable implementation. Here is what I think is missing so far...

theogf

matrix-valued kernels for structured multi-output

7

we would like matrix-valued kernels so that we can learn vector fields! Good starting points would be the [divergence-free and curl-free kernels](http://preprint.impa.br/FullText/Macedo__Thu_Oct_21_16_38_10_BRDT_2010/macedo-MVRBFs.pdf). There exist [python](https://operalib.github.io/operalib/documentation/auto_examples/plot_ovk_regression_cf.html) and [matlab](https://arxiv.org/pdf/1502.01575.pdf) implementations of this....

gszep

Add ANOVA kernel

1

http://crsouza.com/2010/03/17/kernel-functions-for-machine-learning-applications/#anova

theogf

Provide domain of kernel parameters

1

This is an issue to generalize #20 but in a less restrictive way. I still think that we should provide an hyper-parameter type(s) for the kernels to give informations about...

theogf

Related PR #70 Related Comment [link](https://github.com/JuliaGaussianProcesses/KernelFunctions.jl/pull/70#pullrequestreview-403751395) TODO: Implement efficient way to deal with multi dimensional input for the `NeuralNetworkKernel`.

sharanry