transform(kpca) returns error

[gdkrmr]

using MultivariateStats
data = randn(3, 100)
kpca = fit(KernelPCA, data)
transform(kpca) # error
transform(kpca, data) # works 

[wildart]

transform(kpca) only works with precomputed kernels. There should be more comprehensive error for this call.

[gdkrmr]

How do I get the embedded coordinates of my original data points then?

[wildart]

As you wrote above:

original = randn(3, 100)
model = fit(KernelPCA, original)
embedded = transform(model, original) 

[gdkrmr]

that is quite inefficient, because transform(kpca, X) calculates the kernel matrix again, this could be avoided with:

transform(kpca::KernelPCA) = sqrt.(kpca.λ) .* kpca.α'

[wildart]

You are right, I completely disregarded that. Fix is coming.