XGB model to C++ function

dump_xgb.py

########################################
## # Train BDT and output model .json ##
########################################
import numpy as np
from sklearn.model_selection import train_test_split
import sys
import os
import xgboost as xgb

### Define test and train matrices somehow

### Train BDT
dtrain = xgb.DMatrix( X_train, label=y_train)
dtest = xgb.DMatrix( X_test, label=y_test)
evallist  = [(dtest,'eval'), (dtrain,'train')]
num_round = 800
param = {}
param['max_depth'] = 5
param['eval_metric'] = "auc"
bst = xgb.train( param.items(), dtrain, num_round, evallist, early_stopping_rounds=20 )
y_pred = bst.predict(dtest)

# Dump BDT into model file
with open("model.json", "w") as fhout:
    fhout.write("[\n"+",\n".join(bst.get_dump(dump_format="json"))+"\n]")


########################################
#### # Parse .json in separate file ####
########################################
import json
import pprint
import numpy as np

nfeatures = 17 # for now, have to hardcode number of variables

with open("model.json", "r") as fhin:
    js = json.loads(fhin.read())

def get_leaf(d_features, entry, depth=0):
    # recursively venture into tree and collect leaf values,
    # making a nested set of C++ ternary statements
    if "leaf" in entry:
        return entry["leaf"]
    splitvar = entry["split"]
    splitval = entry["split_condition"]
    yesnode = [c for c in entry["children"] if c["nodeid"] == entry["yes"]][0]
    nonode = [c for c in entry["children"] if c["nodeid"] == entry["no"]][0]
    return "({} < {} ? {} : {})".format(splitvar, splitval, get_leaf(d_features,yesnode, depth=depth+1), get_leaf(d_features,nonode, depth=depth+1))

# Print c++ function to stdout, so will need to redirect
# like: `python blah.py > blah.h`
colnames = ["float f"+str(ic) for ic in range(nfeatures)]
print "float get_prediction({}) {{".format(",".join(colnames))
print "  float w = 0.;"
for j in js:
    print "  w += {};".format(get_leaf(vals, j))
print "  return 1.0/(1.0+exp(-1.0*w));"
print "}"