MethodError from sample with Float32 weights

[juliohm]

I think this code should work:

julia> using StatsBase

julia> sample(1:3, StatsBase.Weights(rand(Float32,3)), replace=true)
ERROR: MethodError: no method matching sample(::UnitRange{Int64}, ::Weights{Float32, Float32, Vector{Float32}}; replace=true)
Closest candidates are:
  sample(::AbstractArray, ::AbstractWeights) at /home/juliohm/.julia/packages/StatsBase/DU1bT/src/sampling.jl:561 got unsupported keyword argument "replace"
  sample(::AbstractArray, ::AbstractWeights, ::Integer; replace, ordered) at /home/juliohm/.julia/packages/StatsBase/DU1bT/src/sampling.jl:932
  sample(::AbstractArray, ::AbstractWeights, ::Tuple{Vararg{Int64, N}} where N; replace, ordered) at /home/juliohm/.julia/packages/StatsBase/DU1bT/src/sampling.jl:939

It works fine with Float64.

[nalimilan]

It doesn't seem to work with Float64 either. AFAICT you need to pass a third positional argument giving the number of values you want if you pass replace.

[juliohm]

Sorry I copied the wrong line of code. This is the code that triggers the bug:

using StatsBase

n = 5030
N = 141230

sample(1:n, StatsBase.Weights(rand(Float32,n)), N, replace=true)

If you change to Float64 it works fine.

[nalimilan]

OK. Looking at the problematic make_alias_table function, it seems that its signature could be relaxed without creating issues (this code is quite old).

[juliohm]

This bug is affecting me again in other projects. The make_alias_table function is causing a lot of issues downstream it seems.

[juliohm]

Copying the error message:

ERROR: MethodError: no method matching make_alias_table!(::Weights{Float32, Float32, Vector{Float32}}, ::Float32, ::Vector{Float64}, ::Vector{Int64})
Closest candidates are:
  make_alias_table!(::AbstractVector{Float64}, ::Float64, ::AbstractVector{Float64}, ::AbstractVector{Int64}) at /home/juliohm/.julia/packages/StatsBase/DWW4r/src/sampling.jl:587
Stacktrace:
 [1] alias_sample!(rng::Random._GLOBAL_RNG, a::UnitRange{Int64}, wv::Weights{Float32, Float32, Vector{Float32}}, x::Vector{Int64})
   @ StatsBase ~/.julia/packages/StatsBase/DWW4r/src/sampling.jl:666
 [2] sample!(rng::Random._GLOBAL_RNG, a::UnitRange{Int64}, wv::Weights{Float32, Float32, Vector{Float32}}, x::Vector{Int64}; replace::Bool, ordered::Bool)
   @ StatsBase ~/.julia/packages/StatsBase/DWW4r/src/sampling.jl:914
 [3] #sample#203
   @ ~/.julia/packages/StatsBase/DWW4r/src/sampling.jl:929 [inlined]
 [4] #sample#204
   @ ~/.julia/packages/StatsBase/DWW4r/src/sampling.jl:932 [inlined]
 [5] top-level scope
   @ REPL[10]:1

[nalimilan]

Can you investigate whether widening the type signature seems to be OK?

[LilithHafner]

This bug is no longer present.

julia> using StatsBase

julia> n = 5030
5030

julia> N = 141230
141230

julia> sample(1:n, StatsBase.Weights(rand(Float32,n)), N, replace=true)
141230-element Vector{Int64}:
  247
  862
 2144
    ⋮
 3452
 1185

[devmotion]

This issue was fixed by https://github.com/JuliaStats/StatsBase.jl/pull/499.