Use DFS to get a valid but smaller new node order

[tmct]

Please see #571 for details

[tmct]

Resolves #571

[hcho3]

Hi, thanks for submitting a pull request. The fix looks good to me.

Can you add a unit test that uses a LightGBM tree with depth >32 ?

[tmct]

Good idea. I've synthesised a model that reproduces - will try to get it into a unit test.

'tree\nversion=v4\nnum_class=1\nnum_tree_per_iteration=1\nlabel_index=0\nmax_feature_idx=0\nobjective=regression\nfeature_names=this\nfeature_infos=[0:100]\ntree_sizes=1119\n\nTree=0\nnum_leaves=32\nnum_cat=0\nsplit_feature=0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0\nsplit_gain=0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1\nthreshold=1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1\ndecision_type=2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2\nleft_child=1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 -31\nright_child=-1 -2 -3 -4 -5 -6 -7 -8 -9 -10 -11 -12 -13 -14 -15 -16 -17 -18 -19 -20 -21 -22 -23 -24 -25 -26 -27 -28 -29 -30 -32\nleaf_value=31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 0 1\nleaf_weight=1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1\nleaf_count=1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1\ninternal_value=0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0\ninternal_weight=1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1\ninternal_count=1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1\nis_linear=0\nshrinkage=1\n\n\nend of trees\n\nfeature_importances:\nthis=31\n\npandas_categorical:null\n'

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 84.66%. Comparing base (27b81f7) to head (8175573).
Report is 5 commits behind head on mainline.

Additional details and impacted files

@@             Coverage Diff              @@
##           mainline     #570      +/-   ##
============================================
+ Coverage     84.58%   84.66%   +0.08%     
============================================
  Files            75       75              
  Lines          6547     6549       +2     
  Branches        528      528              
============================================
+ Hits           5538     5545       +7     
+ Misses         1009     1004       -5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[hcho3]

@tmct Can you post the code that generated the model? It would be nice to not add the model file itself to the git repo.

[tmct]

The model is effectively handmade outside of LightGBM using the LightGBM model format, and I can't share the specific construction code for it, sorry. I think it might be very difficult to produce a small model with this depth property through the natural training dynamics of LightGBM.... You might be able to achieve it through clever use of features such as sample weights and custom objectives...

The model file above is a string less than 2kB - I tried to make it as small as possible but still failing the test.

[hcho3]

I see. Let's include the model file in the repo then. If you need help setting up the test, ping me.

[tmct]

Thank you, will do.

[hcho3]

@tmct Do you mind if I take over this pull request? I'd like to include this as part of the next release of Treelite.

[tmct]

@hcho3 I would be delighted if you could finish the tests for me, please. Sorry that I have not yet found the time.

[hcho3]

Done!

[tmct]

Many thanks!