Website moved

Gotchas in Julia

Julia is a fun new language that definitely fills a gap in capabilities between something like Python and something like Fortran, but it certainly has some “gotchas” that I run into a lot, and its error messages can be downright cryptic. Here I’ve put together a list of the ones that got me the most when I was learning, and I’ll update this if I find any more.

Method errors

Because Julia uses multiple dispatch, which basically means it determines which version of a function (which “method”) to call at run time based on the types of the inputs, a large subset of errors are reported as “method errors”, meaning it can’t find a method (i.e. version of a function) that accepts the right set of types. Here’s some of the ways that these errors got me.

No method found for function with nested types in the signature

Let’s say you have a method like this:

function foo(x::AbstractArray{AbstractFloat})
...
end

but, when you call, e.g. foo([1.0, 2.0, 3.0]) you get the infamous “No method found matching” error. The problem is that foo needs defined like this:

function foo(x::AbstractArray{<:AbstractFloat})
...
end

For whatever reason, Julia will automatically match the outermost type to any subtypes, but the has to be explicitly told to match subtypes of inner types. So for a function foo(x::AbstractArray) would match when called with a concrete array of floats, e.g. Array{Float64}, but foo(x::AbstractArray{AbstractFloat}) would only match if called with some type of array containing AbstractFloats and not any subtype of AbstractFloat.

No method found for math operations on arrays

One that gets me a lot is when I’m trying to assign a value to a slice of an array. If I try to do something like:

arr[:,1] = NaN;

I get the error MethodError: no method matching setindex_shape_check. This one is particularly confusing, but all I need to do to fix it is replace = with .=

arr[:,1] .= NaN;

This can lead to expressions which look like they have an awful lot of periods in them. For instance, I was once trying to get a logical index (xx) for elements in two arrays of equal size that were not NaNs in either array, which looked like:

xx = .!isnan.(arr1) .& .!isnan.(arr2)

!, isnan, and & all only work on single elements, so I had to broadcast all of them.

Another favorite is when I’m trying to replace a bunch of values in an array that all meet some criteria, say all negative values should be replaced with NaNs. This looks like:

data[data .< 0] .= NaN

and if you forget the dot in the .= operator, you’ll get a method error about something called setindex_shape_check. Well, now you know what that refers to.

No method found with keyword arguments, even though all the other types are right

If you have a function with keyword arguments, like

function reldiff(a, b; percent=false)
    delta = (b-a)/abs(denom)
    if percent
        delta *= 100
    end
    return delta
end

you’ll get a method error if you misspell the keyword, so reldiff(100., 101., pct=true) will give a method error, even though all the types are fine. Fortunately, it will tell you slightly further down in the error message “got unsupported keyword”.

Syntax errors

Character vs. string literals

Coming from Matlab or Python, you might be used to being able to define strings using single or double quotes. (Yes there’s a slight difference in Matlab between the two types of quote, but 'hello' and "hello" are both valid.)

In Julia, single quotes are for characters and double quotes are for strings. That means that 'a', "a", and "apple" are all valid but 'apple' is not (can’t have more than one letter inside a set of single quotes.) In general, use double quotes in Julia unless you know you need a character specifically, and not just a length-1 string.

Array literals: commas vs. spaces

This is one that’s handy once you know about it, but diabolically confusing if you don’t. In Julia, when you define an array with square brackets, it matters whether you separate the entries with spaces or commas. Commas give you a one dimensional array (a vector):

julia> ndims[1,2,3]
1

julia> size([1,2,3])
(3,)

Spaces give you a 1-by-N array, which actually has two dimensions:

julia> ndims([1 2 3])
2

julia> size([1 2 3])
(1, 3)

This means matrix multiplication works:

# Dot product, 1*1 + 2*2 + 3*3 = 14
julia> [1 2 3] * [1,2,3]
1-element Array{Int64,1}:
 14

But does mean that you can’t freely choose spaces or commas depending on what is easier to type or reads better.

Bounds errors

** Bounds error with logical index **

If you get a BoundsError when trying to index an array with a logical array, check that the logical array is the right size. For example if arr is 5x3, then

xx = [true, false, false, false];
arr[xx,:];

will cause a BoundsError because xx is only four long, but size(arr,1) is 5.

Dependency warnings

Package requires another, but it is not in its dependencies

This is usually just a warning. It usually happens when you are working on a custom package and have inserted a new using <package> or import <package> call somewhere, but not also done ]add <package> while that project is active. If the package you’re trying to import is available on your system, it’ll be imported, but because it hasn’t been “officially” added to your project, the fact that it’s needed is not recorded in the Project.toml file. adding that package with ]add <package> from the Julia REPL is usually enough to fix this.