Hi again,
I found a way to prevent my code from crashing and it's really unexpected.
I modified some global variable which only affects functions that are
called by my emloop function and somehow it works. I have no idea why did
this affect my code though, and especially what it made it segfault instead
of just return an error. So now the serial version works, I'm still
struggling with pmap with plenty of "result shape unspecified" errors but
that's ok.
All the best,
Archibald
On Friday, 24 April 2015 10:36:15 UTC+2, Archibald Pontier wrote:
>
> Hi everyone,
>
> Julia is (sometimes) segfaulting when running this function :
>
> function emloop(img, Sigma, mu, t, model, imgc, bg, FullTs)
> EMiter = 1;
> Check = typemax(Float64);
> GModels, Genes,
> (x, y) = size(img);
> D = zeros(x, y); E = zeros(x, y); B = zeros(x, y);
>
> # And we start the EM loop
> while EMiter <= maxEM && Check > ϵ
> # Calculate the betas & find the optimal transformation for the
> current iteration
> (D, E, B) = fastBetas(model, [], mu, Sigma, img);
> nt = findTransform(model, B, Gray((bg.r + bg.g + bg.b)/3), FullTs);
>
> # Now calculate the optimal parameters for the appearance model
> OldMu = mu; # For termination purposes
> (mu, Sigma) = calcTheta(img, D, E, B);
>
> # Apply the transformation found to the image
> img = RigidTransform(img, nt, bg, true, true);
>
> # Deal with the transformation
> Oldt = t;
> t += nt;
> t.tx %= x;
> t.ty %= y;
>
> # Termination criteria
> NormOldTT = sqrt((Oldt.tx - t.tx)^2 + (Oldt.ty - t.ty)^2 + (Oldt.θ -
> t.θ)^2);
> CheckMu = norm(OldMu.D - mu.D) + norm(OldMu.E - mu.E) + norm(OldMu.B -
> mu.B)
> Check = CheckMu + NormOldTT # Giving lots of weight to the
> transformation
>
> EMiter += 1;
> end
>
> return emdata(D, E, B, img, θ(Sigma, mu), t);
> end
>
> Some context: img is and Image{RGB{Float64}}, Sigma and mu are custom
> types :
> type Σ
> D::Array{Float64, 2}
> E::Array{Float64, 2}
> B::Array{Float64, 2}
> end
>
> type μ
> D::Array{Float64, 1}
> E::Array{Float64, 1}
> B::Array{Float64, 1}
> end
>
> t is another custom type to represent a rigid body transformation
> (translation_x, translation_y, rotation_angle), bg is, for example,
> RGB(0.0, 0.0, 0.0) (black), and FullTs is an array of transformations,
> namely all the transformations we consider.
>
> It's weird for multiple reasons: first of all, this code was running fine
> when not in a separate function returning a custom type (the custom type in
> the return is to ease the access since I want to call this function for an
> array of images with pmap and output all that which is not practical).
> Furthermore, sometimes it crashes with a segfault, sometimes with an
> undefined reference on this line, not always after the same EM iteration:
> CheckMu = norm(OldMu.D - mu.D) + norm(OldMu.E - mu.E) + norm(OldMu.B -
> mu.B)
>
> Thanks in advance for the help,
> Archibald
>
