Provide a kwarg to turn off data healthy check.

[GiggleLiu]

Great tool!

For some user defined types, the following check would break for not having isinf and isnan.

function check_data_health(xdata, ydata)
    if any(ismissing, xdata) || any(ismissing, ydata)
        error("Data contains `missing` values and a fit cannot be performed")
    end
    if any(isinf, xdata) || any(isinf, ydata) || any(isnan, xdata) || any(isnan, ydata)
        error("Data contains `Inf` or `NaN` values and a fit cannot be performed")
    end
end

An example is StaticArray. By overloading this check (so called prirating), everything works fine.
I guess it would be good to provide a kwarg to allow user to switch off this check?

[pkofod]

Does it not work for a static array?

julia> using StaticArrays

julia> any(isnan, @SVector([1,2,3]))
false

[GiggleLiu]

I mean sometimes, it could be helpful to use a StaticArray as a scalar. isnan and isinf are originally designed for Floating point numbers.

julia> isnan(SMatrix{2,2}(1.0, 2.0, 3.0, 5.0))
ERROR: MethodError: no method matching isnan(::SArray{Tuple{2,2},Float64,2,4})
Closest candidates are:
  isnan(::BigFloat) at mpfr.jl:893
  isnan(::Missing) at missing.jl:100
  isnan(::Float16) at float.jl:537
  ...
Stacktrace:
 [1] top-level scope at REPL[4]:1

julia> isnan(SVector{4}(1.0, 2.0, 3.0, 5.0))
ERROR: MethodError: no method matching isnan(::SArray{Tuple{4},Float64,1,4})
Closest candidates are:
  isnan(::BigFloat) at mpfr.jl:893
  isnan(::Missing) at missing.jl:100
  isnan(::Float16) at float.jl:537
  ...
Stacktrace:
 [1] top-level scope at REPL[5]:1

(@v1.4) pkg> st StaticArrays
Status `~/.julia/environments/v1.4/Project.toml`
  [90137ffa] StaticArrays v0.12.3

[pkofod]

Let's try to fix this :) Are you using something like a static array of static arrays?

[GiggleLiu]

Sounds great, yes, I am using a static vector as the element.

[pkofod]

What are you using as the container?

[GiggleLiu]

A regular array, I should probably write a MWE.

[pkofod]

That would be helpful. Can you use RecursiveArrayTools?

[gustaphe]

In #202 , I used something like

is_unhealthy(x) = ismissing(x) || isnan(x) || isinfinite(x)
is_unhealthy(x::AbstractArray) = any(is_unhealthy, x)

which recursively does the right thing for vectors. Could be interesting to implement for this whether or not #202 happens?

[m-wells]

I ran into this issue. I have a custom LazyTable that I pass as xdata. I only consume certain columns but some unused columns contain NaN. As this is a lazy source the presence of the extra columns do not harm performance. The check_data_health function is eager (and ignorant of the data source). I can disable this check easily enough.

function LsqFit.check_data_health(::MyData, ::Any) end

I imagine there is probably a more flexible way of erroring out such as waiting to do any ismissing, isnan, isfinite such as waiting until after the first evaluation and checking df.

LsqFit.jl/src/levenberg_marquardt.jl

Lines 79 to 100 in 863bbca

	function levenberg_marquardt(
	df::OnceDifferentiable,
	initial_x::AbstractVector{T};
	x_tol::Real=1e-8,
	g_tol::Real=1e-12,
	maxIter::Integer=1000,
	maxTime::Float64=Inf,
	lambda=T(10),
	tau=T(Inf),
	lambda_increase::Real=10.0,
	lambda_decrease::Real=0.1,
	min_step_quality::Real=1e-3,
	good_step_quality::Real=0.75,
	show_trace::Bool=false,
	store_trace::Bool=false,
	lower::AbstractVector{T}=Array{T}(undef, 0),
	upper::AbstractVector{T}=Array{T}(undef, 0),
	avv!::Union{Function,Nothing,Avv}=nothing,
	) where {T}
	# First evaluation
	value_jacobian!!(df, initial_x)