improve pretty print for DimensionalArray

[Datseris]

Now from

using Dates, DimensionalData
using DimensionalData: Time, X
timespan = DateTime(2001):Month(1):DateTime(2001,12)
t = Time(timespan)
x = X(10:10:100)
A = DimensionalArray(rand(12,10), (Time(timespan), X(10:10:100)))

one gets:

[Datseris]

I don't know if this is a bug, unwanted behavior, or wanted behavior, but this PR made me realize that:

julia> B = mean(A; dims = Time)
DimensionalArray with dimensions:
  Time: DateTime[2001-01-01T00:00:00]
  X: 10:10:100
and data:
1×10 Array{Float64,2}:
 0.573145  0.576649  0.454423  0.528486  0.639443  …  0.373484  0.464725  0.47384  0.526452

reducing operations keep the dimensions even if reduced over them (in contrast to xarray from Python).

[codecov-io]

Codecov Report

Merging #33 into master will increase coverage by 1.04%.
The diff coverage is 87.5%.

@@            Coverage Diff            @@
##           master     #33      +/-   ##
=========================================
+ Coverage   88.15%   89.2%   +1.04%     
=========================================
  Files          10      11       +1     
  Lines         515     528      +13     
=========================================
+ Hits          454     471      +17     
+ Misses         61      57       -4

Impacted Files	Coverage Δ
src/dimension.jl	91.83% <ø> (-1.27%)	⬇️
src/DimensionalData.jl	100% <ø> (ø)	⬆️
src/array.jl	84.84% <100%> (+16.55%)	⬆️
src/prettyprint.jl	86.2% <86.2%> (ø)
src/methods.jl	95.12% <0%> (+0.06%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 847eb58...5a8ec98. Read the comment docs.

[rafaqz]

That's because Julia keeps them! After a reducing operation you always have a length 1 dim. So we have to keep the dimension type too... I just change the step size and make it cover the whole original dim length.

I'm not sure what python does.

[rafaqz]

But otherwise that looks great. You might want to try it with a large vector of DateTime or something as val and see if it holds up - we might need to just use the start and end however julia does that with arrays.

Also maybe try it with larger grids.

[Datseris]

You might want to try it with a large vector of DateTime or something as val and see if it holds up - we might need to just use the start and end however julia does that with arrays.

Good point and thankfully I know how to fix it since I've already done it in DynamicaSystems.jl. I'll check it tomorrow at the office.

[Datseris]

That's because Julia keeps them! After a reducing operation you always have a length 1 dim. So we have to keep the dimension type too... I just change the step size and make it cover the whole original dim length.

Yeah, you are right (although I have to admit I never liked that Julia decision). Regardless, what I do personally is implement drop versions of these functions, like dropmean or dropsum that just do

dropmean(A; dims = 1) = dropdims(mean(A; dims = dims); dims = dims)

As you can see, for standard Julia arrays this is no big deal. But for DimensionalArray, I would imagine that the processes of "dropping" the singleton dimension is much more involved for the general user. Perhaps I should contribute dropmean/sum/etc. here?

Python's xarray automatically drops singleton dimensions, e.g. if you sum over Lattitude, the final array will have only longitude.

[rafaqz]

dropdims is actually easier with dimensions as you don't need to know the number: dropdims(A; dims=X)

Look in the methods.jl tests to see everything that's implemented with dims. It's quite a lot.

For your dropmean I actually want to minimise custom method exports and focus on covering all the base methods first. DimensionalArrays should feel pretty much the same as Arrays just with typed dims.

[Datseris]

running test/prettyprint.jl now gives:

Unfortunately, even though I've used the limiter function suggested by Sebastian, it works but not "precisely". It limits the vector output to a elements to fill the row, as it should, but its just goes a little overboard in Juno. Regardless, this is unrelated with this PR, and as far as Julia Base is concerned, I used all the correct tools to make it print the dimension values in 1 line. The rest is up to Juno, and I've reported it.

This is good to go from my end.

[rafaqz]

Can we run show on these? We probably don't need an actual test, but it would be good to run the code to make sure it doesn't error.

[rafaqz]

The outputs look great. Thanks for working on the print limiting I'm happy with that.

[Datseris]

The outputs look great. Thanks for working on the print limiting I'm happy with that.

Thanks, you are welcome. Can I ask why you didn't merge it if you liked the PR? Seems pretty useful to have and I'd like to work on other features as well.

[rafaqz]

See the inline comment above! Maybe you missed it. Just add show methods in the tests to make sure they don't error, as a minimal test.

[Datseris]

Ah sorry, yes I missed it. I can add proper tests, hold on.

[rafaqz]

Thanks thats even better