Re-posted from: https://bkamins.github.io/julialang/2021/04/16/arrays.html
Introduction
DataFrames.jl will have a 1.0 release in a few days. In this post I
want to comment on an issue that I expect might cause the most legacy code
breakage with this release.
The post is tested under Julia 1.6, OrdinaryDiffEq 5.52.3, and
DataFrames.jl 0.22.7 (but I also discuss what changes in 1.0 release).
The past
In the past all matrices could be converted to a DataFrame
like this:
~$ julia --banner=no
(@v1.6) pkg> st DataFrames
Status `~/.julia/environments/v1.6/Project.toml`
[a93c6f00] DataFrames v0.22.7
julia> using DataFrames
julia> mat = [1 2; 3 4]
2×2 Matrix{Int64}:
1 2
3 4
julia> DataFrame(mat)
2×2 DataFrame
Row │ x1 x2
│ Int64 Int64
─────┼──────────────
1 │ 1 2
2 │ 3 4
julia> exit()
However, unfortunately this output is deceptive. Try this:
~$ julia --banner=no --depwarn=error
(@v1.6) pkg> st DataFrames
Status `~/.julia/environments/v1.6/Project.toml`
[a93c6f00] DataFrames v0.22.7
julia> using DataFrames
julia> mat = [1 2; 3 4]
2×2 Matrix{Int64}:
1 2
3 4
julia> DataFrame(mat)
ERROR: `DataFrame(columns::AbstractMatrix)` is deprecated, use `DataFrame(columns, :auto)` instead.
julia> exit()
As you can see DataFrame(mat)
is deprecated. The problem is that
Julia 1.6 is hiding deprecation warnings by default. Under
DataFrames.jl 1.0 DataFrame(mat)
will error always.
Why is DataFrame(mat)
not allowed?
The reason why we decided to disallow DataFrame
constructor with a single
Matrix
argument is that DataFrames.jl now follows the rule:
If
DataFrame
constructor is passed a single positional argument this argument
must be a table that is following Tables.jl API.
The point is that Matrix
does not support this API.
A person knowing the DataFrames.jl API more thoroughly might point out that the
statement above is not true, and DataFrames.jl allows for some exceptions where
a non-table is accepted in a constructor. This is indeed the case. Here are
the offending cases:
~$ julia --banner=no --depwarn=error
(@v1.6) pkg> st DataFrames
Status `~/.julia/environments/v1.6/Project.toml`
[a93c6f00] DataFrames v0.22.7
julia> using DataFrames
julia> DataFrame(:a => 1) # a pair
1×1 DataFrame
Row │ a
│ Int64
─────┼───────
1 │ 1
julia> df = DataFrame([:a => 1]) # a vector of pairs
1×1 DataFrame
Row │ a
│ Int64
─────┼───────
1 │ 1
julia> dfr = df[1, :]
DataFrameRow
Row │ a
│ Int64
─────┼───────
1 │ 1
julia> DataFrame(dfr) # DataFrameRow
1×1 DataFrame
Row │ a
│ Int64
─────┼───────
1 │ 1
julia> gdf = groupby(df, :a)
GroupedDataFrame with 1 group based on key: a
First Group (1 row): a = 1
Row │ a
│ Int64
─────┼───────
1 │ 1
julia> DataFrame(gdf) # GroupedDataFrame
1×1 DataFrame
Row │ a
│ Int64
─────┼───────
1 │ 1
These four cases are deliberately left for convenience because there is a very
low risk that they would cause confusion. Out of these four cases a vector of
pairs is most problematic, as in Tables.jl it would get the following treatment:
julia> DataFrame(Tables.columntable([:a => 1]))
1×2 DataFrame
Row │ first second
│ Symbol Int64
─────┼────────────────
1 │ a 1
However, we have decided that it is extremely unlikely that someone might want
to get this type of a data frame (and in case you wanted it the code above
shows you how to get it reliably).
So why have we not made an exception for matrices? The reason is that there
are many cases where AbtractMatrix
actually supports Tables.jl interface
and requires a special way how it should be converted to a DataFrame
.
To give some specific example have a look at this issue related
to differential equations solving (I have adapted it a bit):
~/Desktop/Dev/DF_dev$ julia --banner=no
(@v1.6) pkg> st DataFrames
Status `~/.julia/environments/v1.6/Project.toml`
[a93c6f00] DataFrames v0.22.7
julia> using OrdinaryDiffEq, DataFrames
julia> function parameterized_lorenz(du, u, p, t)
du[1] = p[1] * (u[2] - u[1])
du[2] = u[1] * (p[2] - u[3]) - u[2]
du[3] = u[1] * u[2] - p[3] * u[3]
end
parameterized_lorenz (generic function with 1 method)
julia> u0 = [1.0, 0.0, 0.0];
julia> tspan = (0.0, 1.0);
julia> p = [10.0, 28.0, 8/3];
julia> prob = ODEProblem(parameterized_lorenz, u0, tspan, p);
julia> sol1 = solve(prob, Rosenbrock23());
julia> DataFrame(sol1)
3×61 DataFrame
Row │ x1 x2 x3 x4 x5 x6 ⋯
│ Float64 Float64 Float64 Float64 Float64 Float64 ⋯
─────┼─────────────────────────────────────────────────────────────────────────
1 │ 1.0 0.999925 0.999178 0.995241 0.989338 0.977794 ⋯
2 │ 0.0 0.000209532 0.00230391 0.0134134 0.0302987 0.0641965
3 │ 0.0 7.83999e-10 9.47873e-8 3.21298e-6 1.63912e-5 7.35689e-5
55 columns omitted
julia> DataFrame(Tables.columntable(sol1))
61×4 DataFrame
Row │ timestamp value1 value2 value3
│ Float64 Float64 Float64 Float64
─────┼────────────────────────────────────────────────────
1 │ 0.0 1.0 0.0 0.0
2 │ 7.48361e-6 0.999925 0.000209532 7.83999e-10
3 │ 8.23197e-5 0.999178 0.00230391 9.47873e-8
4 │ 0.000480311 0.995241 0.0134134 3.21298e-6
5 │ 0.00108852 0.989338 0.0302987 1.63912e-5
6 │ 0.00232154 0.977794 0.0641965 7.35689e-5
7 │ 0.00391981 0.963652 0.107499 0.000206214
⋮ │ ⋮ ⋮ ⋮ ⋮
55 │ 0.735315 -7.14414 -8.67351 25.5598
56 │ 0.771575 -7.66582 -9.02351 25.4699
57 │ 0.812974 -8.20321 -9.44232 25.6821
58 │ 0.859509 -8.74036 -9.79982 26.2593
59 │ 0.908783 -9.18149 -9.8967 27.1128
60 │ 0.960609 -9.41649 -9.59708 28.0144
61 │ 1.0 -9.39804 -9.12529 28.5183
47 rows omitted
Now we can see that DataFrame(sol1)
produces a wrong result because sol1
is an AbstractMatrix
as you can check here:
julia> sol1 isa AbstractMatrix
true
Let us switch to main
branch of DataFrames.jl for the remaining of this post
to test the behavior under DataFrames.jl 1.0 that will be released soon:
~/Desktop/Dev/DF_dev$ julia --banner=no
(@v1.6) pkg> activate --temp
Activating new environment at `/tmp/jl_43Ofes/Project.toml`
(jl_43Ofes) pkg> add DataFrames#main
(jl_43Ofes) pkg> st DataFrames
Status `/tmp/jl_43Ofes/Project.toml`
[a93c6f00] DataFrames v0.22.7 `https://github.com/JuliaData/DataFrames.jl.git#main`
julia> using OrdinaryDiffEq, DataFrames
[ Info: Precompiling OrdinaryDiffEq [1dea7af3-3e70-54e6-95c3-0bf5283fa5ed]
julia> function parameterized_lorenz(du, u, p, t)
du[1] = p[1] * (u[2] - u[1])
du[2] = u[1] * (p[2] - u[3]) - u[2]
du[3] = u[1] * u[2] - p[3] * u[3]
end
parameterized_lorenz (generic function with 1 method)
julia> u0 = [1.0, 0.0, 0.0];
julia> tspan = (0.0, 1.0);
julia> p = [10.0, 28.0, 8/3];
julia> prob = ODEProblem(parameterized_lorenz, u0, tspan, p);
julia> sol1 = solve(prob, Rosenbrock23());
julia> DataFrame(sol1)
61×4 DataFrame
Row │ timestamp value1 value2 value3
│ Float64 Float64 Float64 Float64
─────┼────────────────────────────────────────────────────
1 │ 0.0 1.0 0.0 0.0
2 │ 7.48361e-6 0.999925 0.000209532 7.83999e-10
3 │ 8.23197e-5 0.999178 0.00230391 9.47873e-8
4 │ 0.000480311 0.995241 0.0134134 3.21298e-6
5 │ 0.00108852 0.989338 0.0302987 1.63912e-5
6 │ 0.00232154 0.977794 0.0641965 7.35689e-5
7 │ 0.00391981 0.963652 0.107499 0.000206214
⋮ │ ⋮ ⋮ ⋮ ⋮
56 │ 0.771575 -7.66582 -9.02351 25.4699
57 │ 0.812974 -8.20321 -9.44232 25.6821
58 │ 0.859509 -8.74036 -9.79982 26.2593
59 │ 0.908783 -9.18149 -9.8967 27.1128
60 │ 0.960609 -9.41649 -9.59708 28.0144
61 │ 1.0 -9.39804 -9.12529 28.5183
48 rows omitted
And as you can see this time all worked as expected.
How to move from a matrix to a data frame the old way?
A question is can you get an old behavior easily under DataFrames.jl 1.0?
The answer is yes. It is enough to pass :auto
as a second positional argument
to treat any AbstractMatrix
the old way. The key point here is that :auto
adds a second argument to a constructor, which allows to disambiguate this call
and make sure we do not try a Tables.jl fallback. So continuing our last example
we have:
julia> DataFrame(sol1, :auto)
3×61 DataFrame
Row │ x1 x2 x3 x4 x5 x6 ⋯
│ Float64 Float64 Float64 Float64 Float64 Float64 ⋯
─────┼───────────────────────────────────────────────────────────────────────
1 │ 1.0 0.999925 0.999178 0.995241 0.989338 0.977794 ⋯
2 │ 0.0 0.000209532 0.00230391 0.0134134 0.0302987 0.0641965
3 │ 0.0 7.83999e-10 9.47873e-8 3.21298e-6 1.63912e-5 7.35689e-5
55 columns omitted
Here are some more examples (still on main
):
julia> DataFrame([1 2; 3 4])
ERROR: ArgumentError: `DataFrame` constructor from a `Matrix` requires passing :auto as a second argument to automatically generate column names: `DataFrame(matrix, :auto)`
julia> DataFrame([1 2; 3 4], :auto) # auto generated column names
2×2 DataFrame
Row │ x1 x2
│ Int64 Int64
─────┼──────────────
1 │ 1 2
2 │ 3 4
julia> DataFrame([1 2; 3 4], [:c1, :c2]) # passing column names explicitly
2×2 DataFrame
Row │ c1 c2
│ Int64 Int64
─────┼──────────────
1 │ 1 2
2 │ 3 4
So as you can see the fix is easy.
Finally let me comment that another common similar case is a vector of vectors
being passed to a DataFrame
constructor. It follows the same rules:
julia> DataFrame([1:2, 3:4])
ERROR: ArgumentError: `DataFrame` constructor from a `Vector` of vectors requires passing :auto as a second argument to automatically generate column names: `DataFrame(vecs, :auto)`
julia> DataFrame([1:2, 3:4], :auto)
2×2 DataFrame
Row │ x1 x2
│ Int64 Int64
─────┼──────────────
1 │ 1 3
2 │ 2 4
julia> DataFrame([1:2, 3:4], [:c1, :c2])
2×2 DataFrame
Row │ c1 c2
│ Int64 Int64
─────┼──────────────
1 │ 1 3
2 │ 2 4
Again – using Tables.jl default behavior would get you something unexpected
(unless you are really deep into Tables.jl mechanics 😄):
julia> DataFrame(Tables.columntable([1:2, 3:4]))
2×2 DataFrame
Row │ start stop
│ Int64 Int64
─────┼──────────────
1 │ 1 2
2 │ 3 4
Conclusions
We have tried very hard to make things in DataFrames.jl 1.0 maximally consistent
with the whole Julia package ecosystem while allowing a relatively easy handling
of common data processing tasks.
Conversion from a matrix to a DataFrame
is one of common hard corner cases
affected. I hope this post explains you the rationale behind the design
decisions taken in DataFrames.jl 1.0 release in this area and ways how the
DataFrame
constructor should be used to give you desired results.