Logistic regression fails if y is a string of vectors

[biona001]

From README:

For logistic models, y is either a string vector or a m x 2 matrix

But the following doesn't work

using GLMNet
y = ["M", "B", "M", "B"]
X = rand(4, 10)
glmnet(X, y, Binomial())

MethodError: no method matching glmnet(::Matrix{Float64}, ::Vector{String}, ::Binomial{Float64})
Closest candidates are:
  glmnet(::AbstractMatrix{T} where T, ::AbstractVector{T} where T, ::AbstractVector{T} where T) at /home/users/bbchu/.julia/packages/GLMNet/C8WKF/src/CoxNet.jl:151
  glmnet(::AbstractMatrix{T} where T, ::AbstractVector{T} where T, ::AbstractVector{T} where T, ::CoxPH; kw...) at /home/users/bbchu/.julia/packages/GLMNet/C8WKF/src/CoxNet.jl:151
  glmnet(::Matrix{Float64}, ::Vector{Float64}, ::Distribution; kw...) at /home/users/bbchu/.julia/packages/GLMNet/C8WKF/src/GLMNet.jl:485
  ...

Fortunately if y is a matrix with 2 columns, it does work

y = [1 0; 0 1; 0 1; 1 0]
X = rand(4, 10)
glmnet(X, y, Binomial())

Logistic GLMNet Solution Path (100 solutions for 10 predictors in 833 passes):
────────────────────────────────
       df    pct_dev           λ
────────────────────────────────
  [1]   0  0.0        0.476672
  [2]   1  0.0582906  0.455006
  [3]   1  0.11166    0.434325
  [4]   1  0.160737   0.414585
  [5]   1  0.206039   0.395741
  [6]   1  0.248      0.377754
  [7]   1  0.286986   0.360585
  ...

[JackDunnNZ]

It looks like the method that supports the string-vector input is this one:

GLMNet.jl/src/Multinomial.jl

Lines 191 to 203 in 8eff4c4

	function glmnet(X::AbstractMatrix, y; kw...)
	lev = sort(unique(y))
	if length(lev) >= 2
	y = convert(Matrix{Float64}, [i == j for i in y, j in lev])
	if length(lev) == 2
	glmnet(X, y, Binomial(); kw...)
	else
	glmnet(X, y, Multinomial(); kw...)
	end
	else
	error("y has only one level.")
	end
	end

So this works:

using GLMNet
y = ["M", "B", "M", "B"]
X = rand(4, 10)
glmnet(X, y)

The reason it doesn't need a distribution is because it chooses between Binomial and Multinomial based on the number of unique values in y. This method could probably be extended to support passing a distribution, and I guess throwing an error if the distribution and y are incompatible.

At the very least the README should be updated