[julia-users] Re: 4D interpolation from scattered points
Hi there, xdata (or any other grid) does not need to be on a grid, that's just the way the example is written. but no extrapolation, that's right. On Thursday, 14 May 2015 01:05:55 UTC+1, Yakir Gagnon wrote: Thanks a ton people!!! ExtremelyRandomizedTrees.jl: Might be really good, but errored a lot on version 4. ApproXD.jl: Very cool, I'll come back to that later, but for now: What Lininterp cannot do: Multidimensional extrapolation which I need (and accept is prone to silliness). Plus, according to the examples, the known xdata needs to be on a grid. My x,y,z are not equally spaced that way, they are scattered. compute the distances: I started with that, and in light of these results I might fall back to doing just that. Polyharmonic Splines: does exactly what I want, but not faster than just distances (but probably a lot more accurate). CHull.jl: might be awesome for this, but I didn't quickly see how I can use it to interpolate and extrapolate. I feel like I owe you guys some background: I have some color-map functions that take a value, let's call it X (for some 0X1 and for some -pi/2Xpi/2) and return an RGB. I need to use these to convert images that have colors that are similar (but not always identical) to the RGB values from these color-maps back to the value (X) that would have resulted in said RGB (so it's like the inverse of those color-map functions). After much trial and error, I think that since the images I want to convert this way are all 8-bit, the fastest way would be to construct a lookup table for all 2^(8*3) possible RGB combinations. This table (Dict) will take any RGB value from those images I need to convert and return X. This makes sense because I have only 4 such color-map functions, and potentially endless amounts of images to convert (currently a couple of tera). So I think I'll use simple distances, run that once, save those lookup tables and use them again and again on new images. On Thursday, May 14, 2015 at 4:19:18 AM UTC+10, Luke Stagner wrote: You can use Polyharmonic Splines http://nbviewer.ipython.org/gist/lstagner/04a05b120e0be7de9915 On Wednesday, May 13, 2015 at 5:33:08 AM UTC-7, Yakir Gagnon wrote: I have a bunch (~1000) of x,y,z and a corresponding value, V. One unique V for each x,y,z. I want to interpolate and extrapolate wildly (so I really don't care about how accurate or correct it is). The x,y,z I have are not regularly spaced or anything. They're scattered across some range (they all share similar ranges), and I want to know what value (i.e. new V) I should be getting at new and regularly spaced x,y,z. I think Matlab's griddata would do what I want. I tired following the lower-level functionality from Grid.jl, Interpolations.jl, and Dierckx.jl, but couldn't figure out how to get this to work. I think I need to fit some curve to my scattered points and then use that to find the values at the new xyz (at least that's how I think griddata works)... Any good simple ideas out there?
Re: [julia-users] Re: 4D interpolation from scattered points
On Wednesday, May 13, 2015 06:26:07 AM J Luis wrote: Matlab does that with the interp3 function that, I believe, uses the qhull http://www.qhull.org/C lib. At least it used to use it and Octave does. It would be nice to have a wrapper for this lib. There's a PyCall based wrapper here: https://github.com/davidavdav/CHull.jl --Tim quarta-feira, 13 de Maio de 2015 às 13:33:08 UTC+1, Yakir Gagnon escreveu: I have a bunch (~1000) of x,y,z and a corresponding value, V. One unique V for each x,y,z. I want to interpolate and extrapolate wildly (so I really don't care about how accurate or correct it is). The x,y,z I have are not regularly spaced or anything. They're scattered across some range (they all share similar ranges), and I want to know what value (i.e. new V) I should be getting at new and regularly spaced x,y,z. I think Matlab's griddata would do what I want. I tired following the lower-level functionality from Grid.jl, Interpolations.jl, and Dierckx.jl, but couldn't figure out how to get this to work. I think I need to fit some curve to my scattered points and then use that to find the values at the new xyz (at least that's how I think griddata works)... Any good simple ideas out there?
[julia-users] Re: 4D interpolation from scattered points
You could use an ensemble regression approach - see https://github.com/rened/ExtremelyRandomizedTrees.jl#regression for a 1D to 1D example. For your data you could use this (ndims == 2, for visualization): using ExtremelyRandomizedTrees, FunctionalDataUtils # train model ndim = 2 nsamples = 1000 data = 5+5*randn(ndim, nsamples) targets = sum(data,1) targets += randn(size(targets)) model = ExtraTrees(data, targets, regression = true) # predict a = meshgrid(1:20, 1:30) # a is 2 x 20*30 result = reshape(predict(model, a), 20, 30) using Images convert(Image,asimagesc(result)) Am Mittwoch, 13. Mai 2015 14:33:08 UTC+2 schrieb Yakir Gagnon: I have a bunch (~1000) of x,y,z and a corresponding value, V. One unique V for each x,y,z. I want to interpolate and extrapolate wildly (so I really don't care about how accurate or correct it is). The x,y,z I have are not regularly spaced or anything. They're scattered across some range (they all share similar ranges), and I want to know what value (i.e. new V) I should be getting at new and regularly spaced x,y,z. I think Matlab's griddata would do what I want. I tired following the lower-level functionality from Grid.jl, Interpolations.jl, and Dierckx.jl, but couldn't figure out how to get this to work. I think I need to fit some curve to my scattered points and then use that to find the values at the new xyz (at least that's how I think griddata works)... Any good simple ideas out there?
[julia-users] Re: 4D interpolation from scattered points
Thanks a ton people!!! ExtremelyRandomizedTrees.jl: Might be really good, but errored a lot on version 4. ApproXD.jl: Very cool, I'll come back to that later, but for now: What Lininterp cannot do: Multidimensional extrapolation which I need (and accept is prone to silliness). Plus, according to the examples, the known xdata needs to be on a grid. My x,y,z are not equally spaced that way, they are scattered. compute the distances: I started with that, and in light of these results I might fall back to doing just that. Polyharmonic Splines: does exactly what I want, but not faster than just distances (but probably a lot more accurate). CHull.jl: might be awesome for this, but I didn't quickly see how I can use it to interpolate and extrapolate. I feel like I owe you guys some background: I have some color-map functions that take a value, let's call it X (for some 0X1 and for some -pi/2Xpi/2) and return an RGB. I need to use these to convert images that have colors that are similar (but not always identical) to the RGB values from these color-maps back to the value (X) that would have resulted in said RGB (so it's like the inverse of those color-map functions). After much trial and error, I think that since the images I want to convert this way are all 8-bit, the fastest way would be to construct a lookup table for all 2^(8*3) possible RGB combinations. This table (Dict) will take any RGB value from those images I need to convert and return X. This makes sense because I have only 4 such color-map functions, and potentially endless amounts of images to convert (currently a couple of tera). So I think I'll use simple distances, run that once, save those lookup tables and use them again and again on new images. On Thursday, May 14, 2015 at 4:19:18 AM UTC+10, Luke Stagner wrote: You can use Polyharmonic Splines http://nbviewer.ipython.org/gist/lstagner/04a05b120e0be7de9915 On Wednesday, May 13, 2015 at 5:33:08 AM UTC-7, Yakir Gagnon wrote: I have a bunch (~1000) of x,y,z and a corresponding value, V. One unique V for each x,y,z. I want to interpolate and extrapolate wildly (so I really don't care about how accurate or correct it is). The x,y,z I have are not regularly spaced or anything. They're scattered across some range (they all share similar ranges), and I want to know what value (i.e. new V) I should be getting at new and regularly spaced x,y,z. I think Matlab's griddata would do what I want. I tired following the lower-level functionality from Grid.jl, Interpolations.jl, and Dierckx.jl, but couldn't figure out how to get this to work. I think I need to fit some curve to my scattered points and then use that to find the values at the new xyz (at least that's how I think griddata works)... Any good simple ideas out there?
[julia-users] Re: 4D interpolation from scattered points
You can use Polyharmonic Splines http://nbviewer.ipython.org/gist/lstagner/04a05b120e0be7de9915 On Wednesday, May 13, 2015 at 5:33:08 AM UTC-7, Yakir Gagnon wrote: I have a bunch (~1000) of x,y,z and a corresponding value, V. One unique V for each x,y,z. I want to interpolate and extrapolate wildly (so I really don't care about how accurate or correct it is). The x,y,z I have are not regularly spaced or anything. They're scattered across some range (they all share similar ranges), and I want to know what value (i.e. new V) I should be getting at new and regularly spaced x,y,z. I think Matlab's griddata would do what I want. I tired following the lower-level functionality from Grid.jl, Interpolations.jl, and Dierckx.jl, but couldn't figure out how to get this to work. I think I need to fit some curve to my scattered points and then use that to find the values at the new xyz (at least that's how I think griddata works)... Any good simple ideas out there?
