1+ import sys
2+ import unittest
3+ import pytest
4+
5+ import numpy as np
6+ import xgboost as xgb
7+ from hypothesis import given , strategies , assume , settings , note
8+
9+ sys .path .append ("tests/python" )
10+ import testing as tm
11+
12+ rng = np .random .RandomState (1994 )
13+
14+ shap_parameter_strategy = strategies .fixed_dictionaries ({
15+ 'max_depth' : strategies .integers (1 , 11 ),
16+ 'max_leaves' : strategies .integers (0 , 256 ),
17+ 'num_parallel_tree' : strategies .sampled_from ([1 , 10 ]),
18+ }).filter (lambda x : x ['max_depth' ] > 0 or x ['max_leaves' ] > 0 )
19+
20+
21+ class TestOneAPIPredict (unittest .TestCase ):
22+ def test_predict (self ):
23+ iterations = 10
24+ np .random .seed (1 )
25+ test_num_rows = [10 , 1000 , 5000 ]
26+ test_num_cols = [10 , 50 , 500 ]
27+ for num_rows in test_num_rows :
28+ for num_cols in test_num_cols :
29+ dtrain = xgb .DMatrix (np .random .randn (num_rows , num_cols ),
30+ label = [0 , 1 ] * int (num_rows / 2 ))
31+ dval = xgb .DMatrix (np .random .randn (num_rows , num_cols ),
32+ label = [0 , 1 ] * int (num_rows / 2 ))
33+ dtest = xgb .DMatrix (np .random .randn (num_rows , num_cols ),
34+ label = [0 , 1 ] * int (num_rows / 2 ))
35+ watchlist = [(dtrain , 'train' ), (dval , 'validation' )]
36+ res = {}
37+ param = {
38+ "objective" : "binary:logistic_oneapi" ,
39+ "predictor" : "oneapi_predictor" ,
40+ 'eval_metric' : 'logloss' ,
41+ 'tree_method' : 'hist' ,
42+ 'updater' : 'grow_quantile_histmaker_oneapi' ,
43+ 'max_depth' : 1
44+ }
45+ bst = xgb .train (param , dtrain , iterations , evals = watchlist ,
46+ evals_result = res )
47+ assert self .non_increasing (res ["train" ]["logloss" ])
48+ oneapi_pred_train = bst .predict (dtrain , output_margin = True )
49+ oneapi_pred_test = bst .predict (dtest , output_margin = True )
50+ oneapi_pred_val = bst .predict (dval , output_margin = True )
51+
52+ param ["predictor" ] = "cpu_predictor"
53+ bst_cpu = xgb .train (param , dtrain , iterations , evals = watchlist )
54+ cpu_pred_train = bst_cpu .predict (dtrain , output_margin = True )
55+ cpu_pred_test = bst_cpu .predict (dtest , output_margin = True )
56+ cpu_pred_val = bst_cpu .predict (dval , output_margin = True )
57+
58+ np .testing .assert_allclose (cpu_pred_train , oneapi_pred_train ,
59+ rtol = 1e-6 )
60+ np .testing .assert_allclose (cpu_pred_val , oneapi_pred_val ,
61+ rtol = 1e-6 )
62+ np .testing .assert_allclose (cpu_pred_test , oneapi_pred_test ,
63+ rtol = 1e-6 )
64+
65+ def non_increasing (self , L ):
66+ return all ((y - x ) < 0.001 for x , y in zip (L , L [1 :]))
67+
68+ @pytest .mark .skipif (** tm .no_sklearn ())
69+ def test_multi_predict (self ):
70+ from sklearn .datasets import make_regression
71+ from sklearn .model_selection import train_test_split
72+
73+ n = 1000
74+ X , y = make_regression (n , random_state = rng )
75+ X_train , X_test , y_train , y_test = train_test_split (X , y ,
76+ random_state = 123 )
77+ dtrain = xgb .DMatrix (X_train , label = y_train )
78+ dtest = xgb .DMatrix (X_test )
79+
80+ params = {}
81+ params ["tree_method" ] = "hist"
82+ params ["updater" ] = "grow_quantile_histmaker_oneapi"
83+
84+ params ['predictor' ] = "oneapi_predictor"
85+ bst_oneapi_predict = xgb .train (params , dtrain )
86+
87+ params ['predictor' ] = "cpu_predictor"
88+ bst_cpu_predict = xgb .train (params , dtrain )
89+
90+ predict0 = bst_oneapi_predict .predict (dtest )
91+ predict1 = bst_oneapi_predict .predict (dtest )
92+ cpu_predict = bst_cpu_predict .predict (dtest )
93+
94+ assert np .allclose (predict0 , predict1 )
95+ assert np .allclose (predict0 , cpu_predict )
96+
97+ @pytest .mark .skipif (** tm .no_sklearn ())
98+ def test_sklearn (self ):
99+ m , n = 15000 , 14
100+ tr_size = 2500
101+ X = np .random .rand (m , n )
102+ y = 200 * np .matmul (X , np .arange (- 3 , - 3 + n ))
103+ X_train , y_train = X [:tr_size , :], y [:tr_size ]
104+ X_test , y_test = X [tr_size :, :], y [tr_size :]
105+
106+ # First with cpu_predictor
107+ params = {'tree_method' : 'hist' ,
108+ 'predictor' : 'cpu_predictor' ,
109+ 'n_jobs' : - 1 ,
110+ 'seed' : 123 }
111+ m = xgb .XGBRegressor (** params ).fit (X_train , y_train )
112+ cpu_train_score = m .score (X_train , y_train )
113+ cpu_test_score = m .score (X_test , y_test )
114+
115+ # Now with oneapi_predictor
116+ params ['predictor' ] = 'oneapi_predictor'
117+
118+ m = xgb .XGBRegressor (** params ).fit (X_train , y_train )
119+ oneapi_train_score = m .score (X_train , y_train )
120+ m = xgb .XGBRegressor (** params ).fit (X_train , y_train )
121+ oneapi_test_score = m .score (X_test , y_test )
122+
123+ assert np .allclose (cpu_train_score , oneapi_train_score )
124+ assert np .allclose (cpu_test_score , oneapi_test_score )
125+
126+ @given (strategies .integers (1 , 10 ),
127+ tm .dataset_strategy .filter (lambda x : x .name != "empty" ), shap_parameter_strategy )
128+ @settings (deadline = None )
129+ def test_shap (self , num_rounds , dataset , param ):
130+ param .update ({"predictor" : "oneapi_predictor" })
131+ param = dataset .set_params (param )
132+ dmat = dataset .get_dmat ()
133+ bst = xgb .train (param , dmat , num_rounds )
134+ test_dmat = xgb .DMatrix (dataset .X , dataset .y , dataset .w , dataset .margin )
135+ shap = bst .predict (test_dmat , pred_contribs = True )
136+ margin = bst .predict (test_dmat , output_margin = True )
137+ assume (len (dataset .y ) > 0 )
138+ assert np .allclose (np .sum (shap , axis = len (shap .shape ) - 1 ), margin , 1e-3 , 1e-3 )
139+
140+ @given (strategies .integers (1 , 10 ),
141+ tm .dataset_strategy .filter (lambda x : x .name != "empty" ), shap_parameter_strategy )
142+ @settings (deadline = None , max_examples = 20 )
143+ def test_shap_interactions (self , num_rounds , dataset , param ):
144+ param .update ({"predictor" : "oneapi_predictor" })
145+ param = dataset .set_params (param )
146+ dmat = dataset .get_dmat ()
147+ bst = xgb .train (param , dmat , num_rounds )
148+ test_dmat = xgb .DMatrix (dataset .X , dataset .y , dataset .w , dataset .margin )
149+ shap = bst .predict (test_dmat , pred_interactions = True )
150+ margin = bst .predict (test_dmat , output_margin = True )
151+ assume (len (dataset .y ) > 0 )
152+ assert np .allclose (np .sum (shap , axis = (len (shap .shape ) - 1 , len (shap .shape ) - 2 )), margin ,
153+ 1e-3 , 1e-3 )
0 commit comments