Faster model import for sklearn tree models

[hcho3]

Closes #263

Implement treelite.sklearn.import_model_v2, a faster method for importing sklearn models. The method is much faster than its predecessor (treelite.sklearn.import_model) because:

• The new method eschews the Model Builder API and builds a treelite::Tree object directly in C++.
• The new method avoids the use of recursive function calls when parsing the sklearn tree model. Instead, it takes in 7 flat arrays per tree.

Preliminary result: The following code snippet completes in 1.1 sec. Since the same model would take 8 hours to load using the previous method, we have achieved 26122x speedup.

import time
import ctypes
import numpy as np
from sklearn.datasets import make_regression
from sklearn.ensemble import RandomForestRegressor
from treelite.sklearn import import_model_v2

X, y = make_regression(n_samples=100000, n_features=100, n_informative=100)
clf = RandomForestRegressor(n_estimators=100, n_jobs=-1)
clf.fit(X, y)

tstart = time.perf_counter()
import_model_v2(clf)
tend = time.perf_counter()

print(f'Time elapsed = {tend - tstart} sec')

TODOs:

• Implement RandomForestRegressor
• Implement RandomForestClassifier
• Implement GradientBoostingRegressor
• Implement GradientBoostingClassifier
• Implement ExtraTreesRegressor
• Implement ExtraTreesClassifier
• Add tests

cc @wphicks @canonizer @JohnZed

[canonizer]

Approved.

Possibly for a different pull request: would it be possible to import other Scikit-Learn tree-based models from https://scikit-learn.org/stable/modules/classes.html#module-sklearn.ensemble as well?

[wphicks]

In general, this looks great! My only significant questions here are around how we're doing type inspection on the Python side.

[codecov-commenter]

Codecov Report

Merging #264 (a5b249d) into mainline (0d98548) will increase coverage by 0.56%.
The diff coverage is 97.01%.

@@              Coverage Diff               @@
##             mainline     #264      +/-   ##
==============================================
+ Coverage       83.62%   84.18%   +0.56%     
  Complexity         44       44              
==============================================
  Files              93       95       +2     
  Lines            6937     7228     +291     
  Branches           42       42              
==============================================
+ Hits             5801     6085     +284     
- Misses           1112     1119       +7     
  Partials           24       24              

Impacted Files	Coverage Δ	Complexity Δ
include/treelite/frontend.h	90.00% <ø> (ø)	0.00 <0.00> (ø)
python/treelite/sklearn/__init__.py	90.38% <84.21%> (-0.53%)	0.00 <0.00> (ø)
python/treelite/sklearn/importer.py	92.20% <92.20%> (ø)	0.00 <0.00> (?)
src/c_api/c_api.cc	96.32% <100.00%> (+0.42%)	0.00 <0.00> (ø)
src/frontend/sklearn.cc	100.00% <100.00%> (ø)	0.00 <0.00> (?)

---

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 0d98548...a5b249d. Read the comment docs.