Add dumps and loads interface

[pavelzw]

Fixes #48

[pavelzw]

with a bunch of isinstance checks

The question is how many isinstance checks we should do... It should work with RandomForestRegressor, DecisionTreeRegressor, Pipeline but also custom regressors that contain a Tree somewhere inside them... I don't think that we will be able to catch all cases by just calling something on the outer object.

Another option would be to just edit Tree and Booster in the dispatch table s.t. the pickler can handle everything but this comes with the drawback of scikit-learn and lightgbm as necessary dependencies.

[github-actions[bot]]

(benchmark 5091527615 / attempt 1) Base results / Our results / Change	Model	Size	Dump Time	Load Time
sklearn rf 20M	20.8 MiB / 3.0 MiB / 6.87 x	0.03 s / 0.04 s / 1.33 x	0.03 s / 0.03 s / 0.91 x
sklearn rf 20M lzma	6.5 MiB / 2.0 MiB / 3.26 x	12.35 s / 1.23 s / 0.10 x	0.56 s / 0.19 s / 0.33 x
sklearn rf 200M	212.3 MiB / 30.6 MiB / 6.94 x	0.22 s / 0.33 s / 1.53 x	0.26 s / 0.36 s / 1.38 x
sklearn rf 200M lzma	47.5 MiB / 14.6 MiB / 3.25 x	100.38 s / 19.79 s / 0.20 x	4.48 s / 1.45 s / 0.32 x
sklearn rf 1G	1157.5 MiB / 166.8 MiB / 6.94 x	1.35 s / 1.68 s / 1.24 x	1.63 s / 1.78 s / 1.09 x
sklearn rf 1G lzma	257.1 MiB / 98.2 MiB / 2.62 x	507.21 s / 108.81 s / 0.21 x	23.86 s / 8.73 s / 0.37 x
sklearn gb 2M	2.2 MiB / 1.1 MiB / 2.08 x	0.03 s / 0.24 s / 7.50 x	0.08 s / 0.17 s / 2.24 x
sklearn gb 2M lzma	0.6 MiB / 0.2 MiB / 3.82 x	0.97 s / 0.44 s / 0.45 x	0.09 s / 0.14 s / 1.60 x
lgbm gbdt 2M	2.6 MiB / 1.0 MiB / 2.78 x	0.09 s / 0.26 s / 2.81 x	0.01 s / 0.15 s / 12.03 x
lgbm gbdt 2M lzma	0.9 MiB / 0.5 MiB / 1.90 x	1.40 s / 0.53 s / 0.38 x	0.08 s / 0.23 s / 2.81 x
lgbm gbdt 5M	5.3 MiB / 1.9 MiB / 2.81 x	0.19 s / 0.51 s / 2.77 x	0.03 s / 0.32 s / 12.18 x
lgbm gbdt 5M lzma	1.7 MiB / 0.8 MiB / 1.96 x	3.85 s / 1.10 s / 0.29 x	0.15 s / 0.39 s / 2.55 x
lgbm gbdt 20M	22.7 MiB / 7.6 MiB / 3.00 x	0.75 s / 2.14 s / 2.85 x	0.12 s / 1.30 s / 11.20 x
lgbm gbdt 20M lzma	6.3 MiB / 3.0 MiB / 2.09 x	19.14 s / 5.23 s / 0.27 x	0.60 s / 1.55 s / 2.57 x
lgbm gbdt 100M	101.1 MiB / 33.0 MiB / 3.06 x	3.27 s / 9.58 s / 2.93 x	0.53 s / 55.34 s / 105.15 x
lgbm gbdt 100M lzma	25.6 MiB / 10.6 MiB / 2.41 x	99.62 s / 26.12 s / 0.26 x	2.58 s / 7.01 s / 2.71 x
lgbm rf 10M	10.9 MiB / 3.2 MiB / 3.46 x	0.37 s / 0.70 s / 1.86 x	0.05 s / 0.55 s / 12.04 x
lgbm rf 10M lzma	0.7 MiB / 0.4 MiB / 1.85 x	2.11 s / 0.98 s / 0.47 x	0.12 s / 0.62 s / 4.93 x

[jonashaag]

I agree. It’s not worth the effort. The current interface is easy enough to use.