Support categorical split in tree model dump.

[trivialfis]

Support categorical tree node for text dump.

Related #6503 .
Close #6298 .

[codecov-commenter]

Codecov Report

Merging #7036 (b223b53) into master (b56614e) will increase coverage by 0.00%.
The diff coverage is 50.00%.

@@           Coverage Diff           @@
##           master    #7036   +/-   ##
=======================================
  Coverage   81.86%   81.87%           
=======================================
  Files          13       13           
  Lines        3916     3917    +1     
=======================================
+ Hits         3206     3207    +1     
  Misses        710      710           

Impacted Files	Coverage Δ
python-package/xgboost/plotting.py	73.78% <50.00%> (+0.25%)	⬆️

---

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 1faad82...b223b53. Read the comment docs.

[hcho3]

Since name is only needed for the error message, can you put this inside if (!is_categorical) block? Inside, use LOG(FATAL) to throw error

[hcho3]

Also, how come we are not writing anything to result for case FeatureMap::kCategorical:, like all the other cases?

[trivialfis]

I kept the possibility that users don't provide a feature map with categorical type. Categorical is a different split type, all the others are just variations of numerical splits. So no matter what the feature map says it has to go through a dedicated print procedure.

I can restrict it and ensure a correct feature map is provided.

[trivialfis]

I remove this logic, now users either not provide a feature map or it has to be consistent with the training data.

[trivialfis]

@hcho3 @RAMitchell could you please take a look at how the output should look. I'm currently printing the categories as sets, but we don't have an \in operator here, so I used =, which is inaccurate. A possible alternative is something like 0 = 0 || 1 || 3. Here is an example text dump, notice the 0=[0, 14, 32] part:

0:[0={0,14,32}] yes=2,no=1,missing=1,gain=11,cover=2
        1:[f1<1] yes=3,no=4,missing=4,gain=0,cover=0
                3:leaf=0,cover=0
                4:leaf=0,cover=0
        2:[0={0,14,32}] yes=6,no=5,missing=5,gain=11,cover=2
                5:leaf=2,cover=3
                6:leaf=3,cover=4

and here's the graphviz output:

[trivialfis]

Changed to using colon:

Text dump:

0:[0:{0,14,32}] yes=2,no=1,missing=1,gain=11,cover=2
        1:[f1<1] yes=3,no=4,missing=4,gain=0,cover=0
                3:leaf=0,cover=0
                4:leaf=0,cover=0
        2:[0:{0,14,32}] yes=6,no=5,missing=5,gain=11,cover=2
                5:leaf=2,cover=3
                6:leaf=3,cover=4

Graphviz:

[RAMitchell]

LGTM, the format looks fine to me.

[trivialfis]

Hmm .. gpu coordinate descent is failing non-deterministically..