fit! on a LinearBinaryClassifier fails for Float32 values

[tiemvanderdeure]

Fitting a machine with a LinearBinaryClassifier model errors if the input data has datatype Float32 or Float16.

The same regression using vanilla GLM works. A LinearRegressor also works.

A reproducible example:

using GLM, CategoricalArrays, MLJBase, MLJGLMInterface
X = (a = rand(Float32, 200), b = rand(Float32, 200))
y_categorical = X.a .+ X.b .> rand(200)

mach = MLJBase.machine(LinearBinaryClassifier(), X, categorical(y_categorical))
mach2 = MLJBase.machine(LinearRegressor(), X, X.a .+ X.b .+ rand(200))

fit!(mach) # errors
fit!(mach2) # works

GLM.glm(@formula(c ~ a + b), merge(X, (; c = (y_categorical))), Bernoulli()) # also works

Stacktrace:

[ Info: Training machine(LinearBinaryClassifier(fit_intercept = true, …), …).
┌ Error: Problem fitting the machine machine(LinearBinaryClassifier(fit_intercept = true, …), …).
└ @ MLJBase C:\Users\tsh371\.julia\packages\MLJBase\fEiP2\src\machines.jl:682
[ Info: Running type checks...
[ Info: Type checks okay.
ERROR: MethodError: no method matching GLM.GlmResp(::Vector{…}, ::Bernoulli{…}, ::LogitLink, ::Vector{…}, ::Vector{…}, ::Vector{…}, ::Vector{…})

Closest candidates are:
  GLM.GlmResp(::V, ::D, ::L, ::V, ::V, ::V, ::V, ::V, ::V, ::V) where {V<:(AbstractVector{T} where T<:AbstractFloat), D<:(Distributions.UnivariateDistribution), L<:Link}
   @ GLM C:\Users\tsh371\.julia\packages\GLM\vM20T\src\glmfit.jl:7
  GLM.GlmResp(::V, ::D, ::L, ::V, ::V, ::V, ::V) where {V<:(AbstractVector{T} where T<:AbstractFloat), D, L}
   @ GLM C:\Users\tsh371\.julia\packages\GLM\vM20T\src\glmfit.jl:28
  GLM.GlmResp(::AbstractVector{T} where T<:AbstractFloat, ::Distributions.Distribution, ::Link, 
::AbstractVector{T} where T<:AbstractFloat, ::AbstractVector{T} where T<:AbstractFloat)
   @ GLM C:\Users\tsh371\.julia\packages\GLM\vM20T\src\glmfit.jl:54
  ...

Stacktrace:
  [1] GLM.GlmResp(y::Vector{Float32}, d::Bernoulli{Float64}, l::LogitLink, off::Vector{Float64}, wts::Vector{Float64})
    @ GLM C:\Users\tsh371\.julia\packages\GLM\vM20T\src\glmfit.jl:61
  [2] GLM.GlmResp(y::Vector{Float32}, d::Bernoulli{Float64}, l::LogitLink, off::Vector{Float64}, wts::Vector{Int64})
    @ GLM C:\Users\tsh371\.julia\packages\GLM\vM20T\src\glmfit.jl:69
  [3] fit(::Type{…}, X::Matrix{…}, y::Vector{…}, d::Bernoulli{…}, l::LogitLink; dropcollinear::Bool, dofit::Bool, wts::Vector{…}, offset::Vector{…}, fitargs::@Kwargs{…})
    @ GLM C:\Users\tsh371\.julia\packages\GLM\vM20T\src\glmfit.jl:582
  [4] fit(::Type{…}, ::FormulaTerm{…}, ::Tables.MatrixTable{…}, ::Bernoulli{…}, ::Vararg{…}; contrasts::Dict{…}, kwargs::@Kwargs{…})
    @ StatsModels C:\Users\tsh371\.julia\packages\StatsModels\syVEq\src\statsmodel.jl:88        
  [5] glm(::FormulaTerm{…}, ::Tables.MatrixTable{…}, ::Bernoulli{…}, ::Vararg{…}; kwargs::@Kwargs{…})
    @ GLM C:\Users\tsh371\.julia\packages\GLM\vM20T\src\glmfit.jl:604
  [6] fit(model::LinearBinaryClassifier, verbosity::Int64, X::@NamedTuple{…}, y::CategoricalVector{…}, w::Nothing)
    @ MLJGLMInterface C:\Users\tsh371\.julia\packages\MLJGLMInterface\h8qVE\src\MLJGLMInterface.jl:419
  [7] fit(model::LinearBinaryClassifier, verbosity::Int64, X::@NamedTuple{…}, y::CategoricalVector{…})
    @ MLJGLMInterface C:\Users\tsh371\.julia\packages\MLJGLMInterface\h8qVE\src\MLJGLMInterface.jl:413
  [8] fit_only!(mach::Machine{LinearBinaryClassifier, true}; rows::Nothing, verbosity::Int64, force::Bool, composite::Nothing)
    @ MLJBase C:\Users\tsh371\.julia\packages\MLJBase\fEiP2\src\machines.jl:680
  [9] fit_only!
    @ MLJBase C:\Users\tsh371\.julia\packages\MLJBase\fEiP2\src\machines.jl:606 [inlined]
 [10] #fit!#63
    @ MLJBase C:\Users\tsh371\.julia\packages\MLJBase\fEiP2\src\machines.jl:777 [inlined]       
 [11] fit!(mach::Machine{LinearBinaryClassifier, true})
    @ MLJBase C:\Users\tsh371\.julia\packages\MLJBase\fEiP2\src\machines.jl:774
 [12] top-level scope

Some type information was truncated. Use `show(err)` to see complete types.

[rikhuijzer]

Hi @tiemvanderdeure. Thanks for reporting this issue. Out of curiosity, could you tell how much you would need Float32 computations? Normally, GLM is meant for semi-small datasets, so it should usually be very easy to promote the numbers to the 64-bit representation (Float32 to Float64) without serious consequences for running time.

The second situation with mach2 actually is promoted to Float64 in the background because

julia> X.a .+ X.b .+ rand(200)) |> typeof
Vector{Float64} (alias for Array{Float64, 1})

In this case, I've gave fixing this a shot in #43, but it looks not as easy as I hoped.

[tiemvanderdeure]

I discovered this error because I am working with data extracted for rasters, which is often Float32.

I don't actually need the Float32 computations, so if it's not easy to fix I'll just put in a safeguard that converts to Float64 before the data is passed to machine for now.

Sloppy on my part with the second example. So in reality, LinearRegressorerror errors as well.

using GLM, CategoricalArrays, MLJBase, MLJGLMInterface
X = (a = rand(Float32, 200), b = rand(Float32, 200))
y_categorical = X.a .+ X.b .> rand(200)

mach = MLJBase.machine(LinearBinaryClassifier(), X, categorical(y_categorical))
mach2 = MLJBase.machine(LinearRegressor(), X, X.a .+ X.b .+ rand(Float32, 200))

fit!(mach) # errors
fit!(mach2) # also errors!

[tiemvanderdeure]

Solved by #45