@harishvk27 Is the model file human-readable or binary?

from treelite.

@harishvk27 Would it be possible to post your model file so that I can look at it?

from treelite.

it's binary.

from treelite.

@harishvk27 Also, can you check if your model can be read by XGBoost Python?

import xgboost
bst = xgboost.Booster()
bst.load_model('test.model')

To my best knowledge, Treelite uses the same loading logic as XGBoost does.

from treelite.

Thanks i could get this working.. great help indeed.

But my predictions aren't comparable to what i am seeing on python... please take a look at the attached example.

my data set file:

Feature-1.0,Feature-2.0,Feature-3.0,Classification-Type
121.0,181500.0,18.0,0.0
175.0,262500.0,26.0,0.0
197.0,295500.0,29.0,0.0
183.0,274500.0,27.0,0.0
106.0,159000.0,15.0,0.0
169.0,253500.0,25.0,0.0
127.0,190500.0,19.0,0.0
190.0,285000.0,28.0,0.0
102.0,153000.0,15.0,0.0
150.0,225000.0,22.0,0.0
1324.0,1986000.0,198.0,1.0
1667.0,2500500.0,250.0,1.0
1393.0,2089500.0,208.0,1.0
1269.0,1903500.0,190.0,1.0
1369.0,2053500.0,205.0,1.0
1285.0,1927500.0,192.0,1.0
1308.0,1962000.0,196.0,1.0
1085.0,1627500.0,162.0,1.0
1855.0,2782500.0,278.0,1.0
1536.0,2304000.0,230.0,1.0
15437.0,23155500.0,2315.0,2.0
11817.0,17725500.0,1772.0,2.0
12270.0,18405000.0,1840.0,2.0
14581.0,21871500.0,2187.0,2.0
11515.0,17272500.0,1727.0,2.0
14503.0,21754500.0,2175.0,2.0
15225.0,22837500.0,2283.0,2.0
13400.0,20100000.0,2010.0,2.0
11399.0,17098500.0,1709.0,2.0
14428.0,21642000.0,2164.0,2.0

my python code to generate model:

import pandas as pd
import sklearn
import numpy as np
import sys
from sklearn import preprocessing
import os

os.chdir('C:/Users/hkulkarni/Documents/data-sets/test')
df = pd.read_csv('test.csv')


# first 3-columns are attributes
# fourth/last column is classification-type
data_x = df[df.columns[0:3]]
data_y = df[df.columns[-1]]

from sklearn.model_selection import train_test_split
train_x, test_x, train_y, test_y = train_test_split(data_x, data_y, test_size=0.11, random_state=42)

import xgboost as xgb
xg_train = xgb.DMatrix(train_x, label=train_y)
xg_test = xgb.DMatrix(test_x, label=test_y)

param = {}
# use softmax multi-class classification
param['objective'] = 'multi:softprob'
# scale weight of positive examples
param['eta'] = 0.1
param['silent'] = 1
param['nthread'] = 1
param['num_class'] = 3
param['max_depth'] = 10
param['colsample_bytree'] = 0.7
param['alpha'] = 0.03
param['subsample'] = 0.8

watchlist = [(xg_train, 'train'), (xg_test, 'test')]

bst = xgb.train(param, xg_train, 300, watchlist, early_stopping_rounds=50)

pred =  bst.predict(xg_test)

predictions = [value for value in pred]
print("test data:",test_x)
print("predictions:",predictions)
bst.save_model('xgBoostMultiClassification.model')

My result from python:
test data: Feature-1.0 Feature-2.0 Feature-3.0
27 13400.0 20100000.0 2010.0
15 1285.0 1927500.0 192.0
23 14581.0 21871500.0 2187.0
17 1085.0 1627500.0 162.0
predictions: [array([ 0.02790421, 0.0505532 , 0.92154264], dtype=float32), array([ 0.0560572 , 0.89143544, 0.05250741], dtype=float32), array([ 0.02790421, 0.0505532 , 0.92154264], dtype=float32), array([ 0.0560572 , 0.89143544, 0.05250741], dtype=float32)]

My result from C-code:

int main(int argc, char **argv)
{
   union Entry my_entry[] = {
                        {.fvalue=13400,.missing=1},
                        {.fvalue=20100000,.missing=1},
                        {.fvalue=2010,.missing=1},
                      };
   float result[3];
   size_t len = predict_multiclass(my_entry,0,result);
   printf("len:[%d]\n",len);
   printf("[%lf][%lf][%lf]\n",result[0],result[1],result[2]);
}

root@ubuntu:/home/hkulkarni/treelite/tests/python/tmp/test_model# ./a.out
len:[3]
[0.910801][0.056217][0.032982]

from treelite.

Thanks! I'll take a look and get back to you.

from treelite.

@harishvk27 The C program should be revised as follows to obtain correct results:

#include <stdio.h>
#include "header.h"

int main(int argc, char **argv) {
   union Entry my_entry[] = {
                        {.fvalue=13400},
                        {.fvalue=20100000},
                        {.fvalue=2010},
                      };   // do not specify missing field, since no feature is missing 
   float result[3];
   size_t len = predict_multiclass(my_entry, 0, result);
   /* Use correct format specifiers in printf() */
   printf("len:[%ld]\n",len);
   printf("[%f][%f][%f]\n", result[0], result[1], result[2]);

   return 0;
}

Hope it helped.

from treelite.

Hats off. thanks this worked.

from treelite.

@harishvk27 how did you get this working after the runtime error? I am getting the same error.

from treelite.