"""
This script was generated by the train.py script in this repository:
https://github.com/ecohealthalliance/geoname-annotator-training
"""
import numpy as np
from numpy import array, int32
HIGH_CONFIDENCE_THRESHOLD = 0.5
GEONAME_SCORE_THRESHOLD = 0.13
base_classifier =\
{
'penalty': 'l1',
'dual': False,
'tol': 0.0001,
'C': 0.1,
'fit_intercept': True,
'intercept_scaling': 1,
'class_weight': None,
'random_state': None,
'solver': 'liblinear',
'max_iter': 100,
'multi_class': 'auto',
'verbose': 0,
'warm_start': False,
'n_jobs': None,
'l1_ratio': None,
'classes_': array([False, True]),
'coef_': array([[
# log_population
0.2833264880250032,
# name_count
0.496850674834566,
# names_used
0.6182820633848923,
# exact_name_match
1.2348109901150883,
# multiple_spans
1.0027275126871056,
# span_length
0.1556028330182338,
# all_acronyms
-1.723855098135833,
# cannonical_name_used
2.4448401433682614,
# loc_NE_portion
1.100324515118662,
# other_NE_portion
-0.006789540638115011,
# noun_portion
1.7846512186329173,
# num_tokens
0.4449857820783286,
# med_token_prob
-0.35487604034835174,
# exact_alternatives
-0.7357808282729202,
# PPL_feature_code
-3.313119817081297,
# ADM_feature_code
-3.8508572602598976,
# PCL_feature_code
0.002625732530454115,
# other_feature_code
-2.5725460514389598,
# first_order
0.8582848733971942,
# combined_span
0.7845557182369852,
# close_locations
0.0,
# very_close_locations
0.0,
# base_score
0.0,
# base_score_margin
0.0,
# contained_locations
0.0,
# containing_locations
0.0,
]]),
'intercept_': array([-14.6902496]),
'n_iter_': array([44], dtype=int32),
}
contextual_classifier =\
{
'penalty': 'l1',
'dual': False,
'tol': 0.0001,
'C': 0.1,
'fit_intercept': True,
'intercept_scaling': 1,
'class_weight': None,
'random_state': None,
'solver': 'liblinear',
'max_iter': 100,
'multi_class': 'auto',
'verbose': 0,
'warm_start': False,
'n_jobs': None,
'l1_ratio': None,
'classes_': array([False, True]),
'coef_': array([[
# log_population
0.3787203315925731,
# name_count
0.47246832816657763,
# names_used
1.0765607603242244,
# exact_name_match
1.4705218728593559,
# multiple_spans
1.3801379355279673,
# span_length
0.21060539648691756,
# all_acronyms
-2.5491642087123516,
# cannonical_name_used
2.877038521477874,
# loc_NE_portion
1.6424801016350434,
# other_NE_portion
-0.38562595379247006,
# noun_portion
2.002630501746275,
# num_tokens
0.4152636087418877,
# med_token_prob
-0.32906537630371446,
# exact_alternatives
-0.8984979885089859,
# PPL_feature_code
-4.534782739767053,
# ADM_feature_code
-5.510120071836727,
# PCL_feature_code
0.003626341435930206,
# other_feature_code
-3.2323872260723783,
# first_order
1.451007214897749,
# combined_span
1.480736480667228,
# close_locations
0.1719448315096514,
# very_close_locations
-0.07886456859233734,
# base_score
-5.393131594700634,
# base_score_margin
3.119070808829057,
# contained_locations
0.1825358655341298,
# containing_locations
0.962090412821035,
]]),
'intercept_': array([-14.57426316]),
'n_iter_': array([52], dtype=int32),
}
# Logistic regression code from scipy
def predict_proba(X, classifier):
"""Probability estimation for OvR logistic regression.
Positive class probabilities are computed as
1. / (1. + np.exp(-classifier.decision_function(X)));
multiclass is handled by normalizing that over all classes.
"""
prob = np.dot(X, classifier['coef_'].T) + classifier['intercept_']
prob = prob.ravel() if prob.shape[1] == 1 else prob
prob *= -1
np.exp(prob, prob)
prob += 1
np.reciprocal(prob, prob)
if prob.ndim == 1:
return np.vstack([1 - prob, prob]).T
else:
# OvR normalization, like LibLinear's predict_probability
prob /= prob.sum(axis=1).reshape((prob.shape[0], -1))
return prob
def predict_proba_base(X):
return predict_proba(X, base_classifier)
def predict_proba_contextual(X):
return predict_proba(X, contextual_classifier)
