Scipy.signal.convolve in Julia

General Usage
1

Hi

I am looking for an implementation of scipy.signal.convolve.
Is there an implementation in Julia?

Or convolution Linear Filtering similar to R stats::filter

Thanks

2

For 1D signals you may use the convolution in DSP.jl. See Convolutions.

For 2D and multi dimensional arrays you may use the convolutions in ImageFiltering.jl implemented by imfilter().
Pay attention that by default it applies correlation. You may change that by applying reflect() on the kernel.

Personally, I wish for a method which is well optimized (Wrapping Intel IPP) which works on arrays in general and not tied to DSP or Images context.
For small kernels I found StaticKernels.jl which is the fastest I could find on the Julia eco system.

If one day NNLib.jl will support the OneDNN backend (See Use oneDNN · Issue #74 · FluxML/NNlib.jl · GitHub), it might become a good CPU based implementation for convolution / correlation.

2 Likes
3

Thanks for your answer. I knew convolutions, but it is not the exact implementation in R or Scipy. Scipy implementation is not too complicated. I will rewrite it in Julia.

4

The DSP methods should be similar to the Scipy ones, see e.g. Translating a 1d convolution from Python to Julia - #4 by rafael.guerra. I think it would be better if you provide some examples of what you’re trying to do and where/why you think they are not behaving the same. Certainly much easier than rewriting your own functions from scratch.

5

I would like to apply linear convolution filtering to univariate time series similar to R’s stats::filter. I could not find such a method in Julia; I will use it for seasonal decomposition similar to R’s stats::decompose or Python’s statsmodels.tsa.seasonal.seasonal_decompose.

6

You can also write your own convolution, if (for whatever reason) you don’t like DSP.jl:

julia> using FFTW

julia> toy_conv(a,b) = begin
           newlen = length(a)+length(b)-1
           pada = [a; zeros(newlen-length(a))]
           padb = [b; zeros(newlen-length(b))]
           real(ifft(fft(pada) .* fft(padb)))
       end
toy_conv (generic function with 1 method)

julia> toy_conv([1,1],[1,1,0])
4-element Vector{Float64}:
 1.0
 2.0
 1.0
 0.0

You can also write your own FFT, if (for whatever reason) you don’t like FFTW…

7

Thank you for your response. I have rewritten the C++ code of R’s stats::filter in Julia. I will share Julia’s implementation of stats::decompose soon. I have seen other Julia developers need it too