44#
55# License: BSD 3 clause
66
7- # import pytest
7+ import pytest
88import numpy .testing as npt
9- # import pandas.testing as pdt
9+ import pandas .testing as pdt
1010import shutil
1111import tempfile
1212
1313import numpy as np
1414import pymc3 as pm
15- # from pymc3 import summary
16- # from sklearn.gaussian_process import \
17- # GaussianProcessRegressor as skGaussianProcessRegressor
18- # from sklearn.model_selection import train_test_split
19- #
20- #
21- # from pmlearn.exceptions import NotFittedError
15+ from pymc3 import summary
16+ from sklearn .gaussian_process import \
17+ GaussianProcessRegressor as skGaussianProcessRegressor
18+ from sklearn .model_selection import train_test_split
19+
20+
21+ from pmlearn .exceptions import NotFittedError
2222from pmlearn .gaussian_process import (GaussianProcessRegressor )
2323# ,
2424# SparseGaussianProcessRegressor,
2525# StudentsTProcessRegressor)
2626
2727
2828class TestGaussianProcessRegressor (object ):
29- """
30- Compare the logp of GPR models in pmlearn to sklearn
31- """
3229
3330 def setup_method (self ):
34- """Setup the data for testing
35- """
3631 self .num_pred = 1
37- self .num_training_samples = 20
38- self .length_scale = 0.1
39- self .signal_variance = 0.01
40- self .noise_variance = 0.01
41- self .X = np .random .randn (self .num_training_samples , self .num_pred )
42- self .y = np .random .randn (self .num_training_samples ) * \
43- self .noise_variance
44- self .Xnew = np .random .randn (50 , self .num_pred )
45- self .pnew = np .random .randn (50 ) * self .noise_variance
46- with pm .Model () as model :
47- cov_func = self .signal_variance ** 2 * \
48- pm .gp .cov .ExpQuad (self .num_pred , self .length_scale )
49- gp = pm .gp .Latent (cov_func = cov_func )
50- f = gp .prior ("f" , self .X , reparameterize = False )
51- p = gp .conditional ("p" , self .Xnew )
52-
53- self .latent_logp = model .logp ({"f" : self .y , "p" : self .pnew })
54- self .plogp = p .logp ({"f" : self .y , "p" : self .pnew })
55-
56- self .test_gpr = GaussianProcessRegressor (kernel = cov_func )
32+ self .num_training_samples = 300
33+
34+ self .length_scale = 1.0
35+ self .signal_variance = 0.1
36+ self .noise_variance = 0.1
37+
38+ X = np .linspace (start = 0 , stop = 10 ,
39+ num = self .num_training_samples )[:, None ]
40+
41+ cov_func = self .signal_variance ** 2 * pm .gp .cov .ExpQuad (
42+ 1 , self .length_scale )
43+ mean_func = pm .gp .mean .Zero ()
44+
45+ f_true = np .random .multivariate_normal (
46+ mean_func (X ).eval (),
47+ cov_func (X ).eval () + 1e-8 * np .eye (self .num_training_samples ),
48+ 1 ).flatten ()
49+ y = f_true + \
50+ self .noise_variance * np .random .randn (self .num_training_samples )
51+
52+ self .X_train , self .X_test , self .y_train , self .y_test = \
53+ train_test_split (X , y , test_size = 0.3 )
54+
55+ self .advi_gpr = GaussianProcessRegressor ()
56+
5757 self .test_dir = tempfile .mkdtemp ()
5858
5959 def teardown_method (self ):
@@ -67,93 +67,83 @@ def test_advi_fit_returns_correct_model(self):
6767 # This print statement ensures PyMC3 output won't overwrite
6868 # the test name
6969 print ('' )
70- self .test_gpr .fit (self .X , self .y )
71-
72- npt .assert_equal (self .num_pred , self .test_gpr .num_pred )
73- npt .assert_almost_equal (self .signal_variance ,
74- int (self .test_GPR .summary ['mean' ]['signal_variance__0' ]),
75- 0 )
76- self .assertAlmostEqual (self .length_scale ,
77- int (self .test_GPR .summary ['mean' ]['length_scale__0_0' ]),
78- 0 )
79- self .assertAlmostEqual (self .noise_variance ,
80- int (self .test_GPR .summary ['mean' ]['noise_variance__0' ]),
81- 0 )
70+ self .advi_gpr .fit (self .X_train , self .y_train ,
71+ inference_args = {"n" : 25000 })
8272
83- # def test_nuts_fit_returns_correct_model(self):
84- # # This print statement ensures PyMC3 output won't overwrite the test name
85- # print('')
86- # self.test_nuts_GPR.fit(self.X_train, self.y_train, inference_type='nuts')
87- #
88- # self.assertEqual(self.num_pred, self.test_nuts_GPR.num_pred)
89- # self.assertAlmostEqual(self.signal_variance,
90- # int(self.test_nuts_GPR.summary['mean']['signal_variance__0']),
91- # 0)
92- # self.assertAlmostEqual(self.length_scale,
93- # int(self.test_nuts_GPR.summary['mean']['length_scale__0_0']),
94- # 0)
95- # self.assertAlmostEqual(self.noise_variance,
96- # int(self.test_nuts_GPR.summary['mean']['noise_variance__0']),
97- # 0)
73+ npt .assert_equal (self .num_pred , self .advi_gpr .num_pred )
74+ npt .assert_almost_equal (
75+ self .signal_variance ,
76+ self .advi_gpr .summary ['mean' ]['signal_variance__0' ],
77+ 0 )
78+ npt .assert_almost_equal (
79+ self .length_scale ,
80+ self .advi_gpr .summary ['mean' ]['length_scale__0_0' ],
81+ 0 )
82+ npt .assert_almost_equal (
83+ self .noise_variance ,
84+ self .advi_gpr .summary ['mean' ]['noise_variance__0' ],
85+ 0 )
9886
9987
100- # class GaussianProcessRegressorPredictTestCase(GaussianProcessRegressorTestCase):
101- # def test_predict_returns_predictions(self):
102- # print('')
103- # self.test_GPR.fit(self.X_train, self.y_train)
104- # preds = self.test_GPR.predict(self.X_test)
105- # self.assertEqual(self.y_test.shape, preds.shape)
106- #
107- # def test_predict_returns_mean_predictions_and_std(self):
108- # print('')
109- # self.test_GPR.fit(self.X_train, self.y_train)
110- # preds, stds = self.test_GPR.predict(self.X_test, return_std=True)
111- # self.assertEqual(self.y_test.shape, preds.shape)
112- # self.assertEqual(self.y_test.shape, stds.shape)
113- #
114- # def test_predict_raises_error_if_not_fit(self):
115- # print('')
116- # with self.assertRaises(NotFittedError) as no_fit_error:
117- # test_GPR = GaussianProcessRegressor()
118- # test_GPR.predict(self.X_train)
119- #
120- # expected = 'Run fit on the model before predict.'
121- # self.assertEqual(str(no_fit_error.exception), expected)
88+ class TestGaussianProcessRegressorPredict (TestGaussianProcessRegressor ):
89+ def test_predict_returns_predictions (self ):
90+ print ('' )
91+ self .advi_gpr .fit (self .X_train , self .y_train ,
92+ inference_args = {"n" : 25000 })
93+ preds = self .advi_gpr .predict (self .X_test )
94+ npt .assert_equal (self .y_test .shape , preds .shape )
95+
96+ def test_predict_returns_mean_predictions_and_std (self ):
97+ print ('' )
98+ self .advi_gpr .fit (self .X_train , self .y_train ,
99+ inference_args = {"n" : 25000 })
100+ preds , stds = self .advi_gpr .predict (self .X_test , return_std = True )
101+ npt .assert_equal (self .y_test .shape , preds .shape )
102+ npt .assert_equal (self .y_test .shape , stds .shape )
103+
104+ def test_predict_raises_error_if_not_fit (self ):
105+ print ('' )
106+ with pytest .raises (NotFittedError ):
107+ advi_gpr = GaussianProcessRegressor ()
108+ advi_gpr .predict (self .X_train )
109+
110+
111+ class TestGaussianProcessRegressorScore (TestGaussianProcessRegressor ):
112+ def test_score_matches_sklearn_performance (self ):
113+ print ('' )
114+ sk_gpr = skGaussianProcessRegressor ()
115+ sk_gpr .fit (self .X_train , self .y_train )
116+ sk_gpr_score = sk_gpr .score (self .X_test , self .y_test )
117+
118+ self .advi_gpr .fit (self .X_train , self .y_train ,
119+ inference_args = {"n" : 25000 })
120+ advi_gpr_score = self .advi_gpr .score (self .X_test , self .y_test )
121+
122+ npt .assert_almost_equal (sk_gpr_score , advi_gpr_score , 1 )
123+
124+
125+ class TestGaussianProcessRegressorSaveAndLoad (TestGaussianProcessRegressor ):
126+ def test_save_and_load_work_correctly (self ):
127+ print ('' )
128+ self .advi_gpr .fit (self .X_train , self .y_train ,
129+ inference_args = {"n" : 25000 })
130+ score1 = self .advi_gpr .score (self .X_test , self .y_test )
131+ self .advi_gpr .save (self .test_dir )
132+
133+ gpr2 = GaussianProcessRegressor ()
134+ gpr2 .load (self .test_dir )
135+
136+ npt .assert_equal (self .advi_gpr .inference_type , gpr2 .inference_type )
137+ npt .assert_equal (self .advi_gpr .num_pred , gpr2 .num_pred )
138+ npt .assert_equal (self .advi_gpr .num_training_samples ,
139+ gpr2 .num_training_samples )
140+ pdt .assert_frame_equal (summary (self .advi_gpr .trace ),
141+ summary (gpr2 .trace ))
142+
143+ score2 = gpr2 .score (self .X_test , self .y_test )
144+ npt .assert_almost_equal (score1 , score2 , 0 )
122145
123146
124- # class GaussianProcessRegressorScoreTestCase(GaussianProcessRegressorTestCase):
125- # def test_score_matches_sklearn_performance(self):
126- # print('')
127- # skGPR = skGaussianProcessRegressor()
128- # skGPR.fit(self.X_train, self.y_train)
129- # skGPR_score = skGPR.score(self.X_test, self.y_test)
130- #
131- # self.test_GPR.fit(self.X_train, self.y_train)
132- # test_GPR_score = self.test_GPR.score(self.X_test, self.y_test)
133- #
134- # self.assertAlmostEqual(skGPR_score, test_GPR_score, 1)
135- #
136- #
137- # class GaussianProcessRegressorSaveAndLoadTestCase(GaussianProcessRegressorTestCase):
138- # def test_save_and_load_work_correctly(self):
139- # print('')
140- # self.test_GPR.fit(self.X_train, self.y_train)
141- # score1 = self.test_GPR.score(self.X_test, self.y_test)
142- # self.test_GPR.save(self.test_dir)
143- #
144- # GPR2 = GaussianProcessRegressor()
145- # GPR2.load(self.test_dir)
146- #
147- # self.assertEqual(self.test_GPR.inference_type, GPR2.inference_type)
148- # self.assertEqual(self.test_GPR.num_pred, GPR2.num_pred)
149- # self.assertEqual(self.test_GPR.num_training_samples, GPR2.num_training_samples)
150- # pd.testing.assert_frame_equal(summary(self.test_GPR.trace),
151- # summary(GPR2.trace))
152- #
153- # score2 = GPR2.score(self.X_test, self.y_test)
154- # self.assertAlmostEqual(score1, score2, 1)
155- #
156- #
157147# class StudentsTProcessRegressorTestCase(unittest.TestCase):
158148#
159149# def setUp(self):
0 commit comments