Stackoverflow with views

[tlienart]

This was reported by an MLJ user here and after some digging it seems that an issue is to do with how a vector y is passed to GLM's glm method; GLM will crash if y is a view. Here's a reproducing example by @olivierlabayle; note that the use of MLJ is just for the sake of getting the same data, everything else is GLM

julia> using Distributions, GLM, MLJ

julia> X, y = @load_boston;

julia> Xmatrix = MLJ.matrix(X);

julia> typeof(X)
NamedTuple{(:Crim, :Zn, :Indus, :NOx, :Rm, :Age, :Dis, :Rad, :Tax, :PTRatio, :Black, :LStat), NTuple{12, SubArray{Float64, 1, Matrix{Float64}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, true}}}

julia> typeof(y)
SubArray{Float64, 1, Matrix{Float64}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, true}

julia> typeof(Xmatrix)
Matrix{Float64} (alias for Array{Float64, 2})

and now this is fine (note the convert)

julia> GLM.glm(Xmatrix, convert(Vector, y), Distributions.Normal(), GLM.IdentityLink())
GeneralizedLinearModel{GLM.GlmResp{Vector{Float64}, Normal{Float64}, IdentityLink}, GLM.DensePredChol{Float64, LinearAlgebra.Cholesky{Float64, Matrix{Float64}}}}:

Coefficients:
──────────────────────────────────────────────────────────────────────
           Coef.  Std. Error      z  Pr(>|z|)   Lower 95%    Upper 95%
──────────────────────────────────────────────────────────────────────
x1   -0.0981569   0.0346913   -2.83    0.0047  -0.166151   -0.0301631
x2    0.0494157   0.0145318    3.40    0.0007   0.020934    0.0778975
x3    0.0166277   0.064685     0.26    0.7971  -0.110153    0.143408
x4   -2.25577     3.38342     -0.67    0.5050  -8.88717     4.37562
x5    5.99805     0.311101    19.28    <1e-82   5.3883      6.60779
x6   -0.00514634  0.0139229   -0.37    0.7117  -0.0324346   0.022142
x7   -0.972546    0.197394    -4.93    <1e-06  -1.35943    -0.585661
x8    0.193115    0.0669889    2.88    0.0039   0.0618192   0.324411
x9   -0.01087     0.00393042  -2.77    0.0057  -0.0185735  -0.00316651
x10  -0.425742    0.110342    -3.86    0.0001  -0.642009   -0.209475
x11   0.015433    0.00271568   5.68    <1e-07   0.0101104   0.0207556
x12  -0.424929    0.051171    -8.30    <1e-15  -0.525222   -0.324635
──────────────────────────────────────────────────────────────────────

but this isn't:

julia> GLM.glm(Xmatrix, y, Distributions.Normal(), GLM.IdentityLink())
ERROR: StackOverflowError:
Stacktrace:
 [1] GLM.GlmResp(y::SubArray{Float64, 1, Matrix{Float64}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, true}, d::Normal{Float64}, l::IdentityLink, off::Vector{Float64}, wts::Vector{Float64}) (repeats 79984 times)
   @ GLM ~/.julia/packages/GLM/5CcRd/src/glmfit.jl:63

The culprit is GlmResp that doesn't seem to dispatch properly and loops to infinity.

(bug) pkg> status
      Status `~/Desktop/jsbx/bug/Project.toml`
  [31c24e10] Distributions v0.25.11
  [38e38edf] GLM v1.5.1
  [add582a8] MLJ v0.16.7

cc @olivierlabayle, @ablaom, @goerch

[nalimilan]

Looks like we need a fix similar to #443.

[ablaom]

@bkamins @nalimilan Been revisiting this issue and wondered if there was anyone with time to push this along?

[bkamins]

I think - assuming he has time - @andreasnoack had spent most time thinking about how to fix these things.

[andreasnoack]

@ablaom Does #446 fix that issue?

[ablaom]

@jbrea would you have a chance to check this? This concerns JuliaAI/MLJGLMInterface.jl#10 .

[jbrea]

Not sure I'm the right person to look at this. But I checked the following:

(@v1.6) pkg> activate --temp

(jl_Aj9lB4) pkg> add Distributions, MLJ, GLM#6059527

julia> using Distributions, GLM, MLJ

julia> X, y = @load_boston;

julia> Xmatrix = MLJ.matrix(X);

julia> GLM.glm(Xmatrix, y, Distributions.Normal(), GLM.IdentityLink())

which runs without errors on julia 1.6.3.

[ablaom]

@jbrea Thanks for checking! I'm so sorry, I pasted the wrong handle 🙄 .

@andreasnoack It looks like we can close this now, thanks.