lightgbm.cs

public class LightGBMTree
{
	public int[] split_feature;
	public double[] threshold;
	public int[] left_child;
	public int[] right_child;
	public double[] leaf_value;
	
	public double Predict(double[] features)
	{
		var node = 0;
		for (;;) {
			var left = features[split_feature[node]] < threshold[node];
			node = left ? left_child[node] : right_child[node];
			if (node < 0)
				return leaf_value[-node-1];
		}
	}
}

lightgbm.py

import re

_FIELDS = [
    ('split_feature', int),
    ('threshold', float),
    ('decision_type', int),
    ('left_child', int),
    ('right_child', int),
    ('leaf_value', float),
    ('cat_boundaries', int),
    ('cat_threshold', int),
]

def read_model(filename):
    with open('model.txt') as f:
        content = f.read()
        
    trees = []
    for i in [m.start() for m in re.finditer('Tree=', content)]:
        c = content[i:].split('\n\n')[0]

        t = {}
        for field, ctor in _FIELDS:
            q = re.search(field + '=([^\n]+)\n', c)
            t[field] = [ctor(x) for x in q.group(1).split()] if q else None
            
        trees.append(t)
        
    return trees


def _predict(tree, features):
    node = 0
    while True:
        if tree['decision_type'][node] & 1:
            int_fval = int(features[tree['split_feature'][node]])
            cat_idx = int(tree['threshold'][node])
            
            cat_boundaries = tree['cat_boundaries']
            cat_threshold = tree['cat_threshold']
            
            rhs = 32 * (cat_boundaries[cat_idx + 1] - cat_boundaries[cat_idx])
            left = int_fval < rhs
            
            z = (cat_threshold[cat_boundaries[cat_idx] + int_fval // 32] >> (int_fval & 31)) & 1
            left &= z
        else:
            left = features[tree['split_feature'][node]] <= tree['threshold'][node]
    
        node = tree['left_child' if left else 'right_child'][node]
        if node < 0:
            return tree['leaf_value'][-node-1]


def predict(trees, features):
    return sum(_predict(t, features) for t in trees)