Coverage for mlprodict/grammar/grammar_sklearn/g_sklearn_linear_model.py: 100%

1# -*- coding: utf-8 -*- 

2""" 

3@file 

4@brief List of interpreted from scikit-learn model. 

5""" 

6import numpy 

7from .g_sklearn_type_helpers import check_type 

8from .grammar.exc import Float32InfError 

9from .grammar.gactions import MLActionCst, MLActionVar, MLActionConcat, MLActionReturn 

10from .grammar.gactions_num import MLActionAdd, MLActionSign 

11from .grammar.gactions_tensor import MLActionTensorDot 

12from .grammar.gmlactions import MLModel 

13 

14 

15def sklearn_logistic_regression(model, input_names=None, output_names=None, **kwargs): 

16 """ 

17 Interprets a `logistic regression <http://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html>`_ 

18 model into a *grammar* model (semantic graph representation). 

19 

20 @param model *scikit-learn* model 

21 @param input_names name of the input features 

22 @param output_names name of the output predictions 

23 @param kwargs additional parameters (*with_loop*) 

24 @return graph model 

25 

26 If *input* is None or *output* is None, default values 

27 will be given to the outputs 

28 ``['Prediction', 'Score']`` for the outputs. 

29 If *input_names* is None, it wil be ``'Features'``. 

30 

31 Additional parameters: 

32 - *with_loop*: False by default, *True* not implemented. 

33 - *dtype*: float32 or float64 

34 """ 

35 if kwargs.get('with_loop', False): 

36 raise NotImplementedError( # pragma: no cover 

37 "Loop version is not implemented.") 

38 if output_names is None: 

39 output_names = ['Prediction', 'Score'] # pragma: no cover 

40 if input_names is None: 

41 input_names = 'Features' 

42 

43 from sklearn.linear_model import LogisticRegression 

44 check_type(model, LogisticRegression) 

45 if len(model.coef_.shape) > 1 and min(model.coef_.shape) != 1: 

46 raise NotImplementedError( # pragma: no cover 

47 "Multiclass is not implemented yet: coef_.shape={0}.".format(model.coef_.shape)) 

48 dtype = kwargs.get('dtype', numpy.float32) 

49 coef_ = model.coef_.ravel() 

50 coef = coef_.astype(dtype) 

51 bias = dtype(model.intercept_[0]) 

52 

53 for i, c in enumerate(coef): 

54 if numpy.isinf(c): 

55 raise Float32InfError( # pragma: no cover 

56 'Unable to convert coefficient {0}: {1}'.format(i, coef[i])) 

57 if numpy.isinf(bias): 

58 raise Float32InfError( # pragma: no cover 

59 'Unable to convert intercept {0}'.format(model.intercept_[0])) 

60 

61 gr_coef = MLActionCst(coef) 

62 gr_var = MLActionVar(coef, input_names) 

63 gr_bias = MLActionCst(bias) 

64 gr_dot = MLActionTensorDot(gr_coef, gr_var) 

65 gr_dist = MLActionAdd(gr_dot, gr_bias) 

66 gr_sign = MLActionSign(gr_dist) 

67 gr_conc = MLActionConcat(gr_sign, gr_dist) 

68 ret = MLActionReturn(gr_conc) 

69 return MLModel(ret, output_names, name=LogisticRegression.__name__) 

70 

71 

72def sklearn_linear_regression(model, input_names=None, output_names=None, **kwargs): 

73 """ 

74 Converts a `linear regression <http://scikit-learn.org/stable/modules/ 

75 generated/sklearn.linear_model.LinearRegression.html>`_ 

76 into a *grammar* model (semantic graph representation). 

77 

78 @param model *scikit-learn* model 

79 @param input_names name of the input features 

80 @param output_names name of the output predictions 

81 @param kwargs additional parameter (*with_loop*) 

82 @return graph model 

83 

84 If *input* is None or *output* is None, default values 

85 will be given to the outputs 

86 ``['Prediction', 'Score']`` for the outputs. 

87 If *input_names* is None, it wil be ``'Features'``. 

88 

89 Additional parameters: 

90 - *with_loop*: False by default, *True* not implemented. 

91 - *dtype*: float32 or float64 

92 """ 

93 if kwargs.get('with_loop', False): # pragma: no cover 

94 raise NotImplementedError("Loop version is not implemented.") 

95 if output_names is None: 

96 output_names = ['Prediction', 'Score'] 

97 if input_names is None: 

98 input_names = 'Features' 

99 

100 from sklearn.linear_model import LinearRegression 

101 check_type(model, LinearRegression) 

102 if len(model.coef_.shape) > 1 and min(model.coef_.shape) != 1: 

103 raise NotImplementedError( # pragma: no cover 

104 "MultiOutput is not implemented yet: coef_.shape={0}.".format(model.coef_.shape)) 

105 

106 dtype = kwargs.get('dtype', numpy.float32) 

107 coef_ = model.coef_.ravel() 

108 coef = coef_.astype(dtype) 

109 bias = dtype(model.intercept_) 

110 

111 for i, c in enumerate(coef): 

112 if numpy.isinf(c): 

113 raise Float32InfError( # pragma: no cover 

114 'Unable to convert coefficient {0}: {1}'.format(i, coef[i])) 

115 if numpy.isinf(bias): 

116 raise Float32InfError( # pragma: no cover 

117 'Unable to convert intercept {0}'.format(model.intercept_)) 

118 

119 gr_coef = MLActionCst(coef) 

120 gr_var = MLActionVar(coef, input_names) 

121 gr_bias = MLActionCst(bias) 

122 gr_dot = MLActionTensorDot(gr_coef, gr_var) 

123 gr_dist = MLActionAdd(gr_dot, gr_bias) 

124 ret = MLActionReturn(gr_dist) 

125 return MLModel(ret, output_names, name=LinearRegression.__name__)