import numpy as np
from scipy.stats import randint as sp_randint, expon, truncnorm, uniform
[docs]class Config:
"""
Class that configures the execution process.
Attributes:
n_folds (int): The number of folds in the experiment
source_crs (int): The EPSG crs code used in input files
target_crs (int): The EPSG crs code to transform the data
clusters_pct (float): Percentage of data points, indicating how many \
clusters to create in order to query Overpass API for streets
osm_buffer (float): A buffer distance (in meters) to consider around \
each bounding box when querying Overpass API
osm_timeout (int): Timeout (in seconds) after five requests to \
Overpass API
max_overpass_tries (int): Maximum number of failed tries to extract the road network when querying the
Overpass API before quiting.
distance_thr (float): Distances in features greater than this value \
will be converted to this threshold
baseline_service (str): The name of the service to consider when \
measuring baseline scores
experiments_path (str): Path to folder that stores the experiments
services (list): The services (geocoders) used in the setup
supported_features (list): List of the supported features to choose \
from
included_features (list): List of the features to be included in the \
experiment
normalized_features (list): List of features to be normalized
supported_classifiers (list): List of the supported classifiers to \
choose from
included_classifiers (list): List of the classifiers to be included \
in the experiment
NB_hparams (dict): Parameters search space for Naive Bayes classifier
NN_hparams (dict): Parameters search space for Nearest Neighbors \
classifier
LR_hparams (dict): Parameters search space for Logistic Regression \
classifier
SVM_hparams (list): Parameters search space for SVM classifier
MLP_hparams (dict): Parameters search space for MLP classifier
DT_hparams (dict): Parameters search space for Decision Tree classifier
RF_hparams (dict): Parameters search space for Random Forest classifier
ET_hparams (dict): Parameters search space for Extra Trees classifier
"""
n_folds = 5
n_jobs = 4 #: int: Number of parallel jobs to be initiated. -1 means to utilize all available processors.
# accepted values: randomized, grid, hyperband - not yet implemented!!!
hyperparams_search_method = 'grid'
"""str: Search Method to use for finding best hyperparameters. (*randomized* | *grid*).
"""
#: int: Number of iterations that RandomizedSearchCV should execute. It applies only when
#: :attr:`hyperparams_search_method` equals to 'randomized'.
max_iter = 30
verbose = True
source_crs = 4326
target_crs = 3857
clusters_pct = 0.015
osm_buffer = 0.001
osm_timeout = 50
max_overpass_tries = 5
distance_thr = 5000.0
square_thr = 500000.0
baseline_service = 'original'
#: int: Seed to use by random number generators.
seed_no = 13
base_dir = '/media/disk/LGM-Geocoding'
services = [
'original',
'arcgis',
'nominatim',
]
supported_features = [
'normalized_coords',
'pairwise_coords_distances',
'pairwise_points_distances',
'centroid_coords_distances',
'centroid_points_distances',
'mean_centroids_coords_distances',
'mean_centroids_points_distances',
'nearest_street_distance_per_service',
'nearest_street_distance_by_centroid',
'zip_codes',
'common_nearest_street_distance',
'intersects_on_common_nearest_street',
'points_area',
'polar_coords',
]
included_features = [
# 'normalized_coords',
'pairwise_coords_distances',
'pairwise_points_distances',
'centroid_coords_distances',
'centroid_points_distances',
'mean_centroids_coords_distances',
'mean_centroids_points_distances',
'nearest_street_distance_per_service',
'nearest_street_distance_by_centroid',
# 'zip_codes',
'common_nearest_street_distance',
'intersects_on_common_nearest_street',
'points_area',
'polar_coords',
]
normalized_features = [
# 'normalized_coords',
'pairwise_coords_distances',
'pairwise_points_distances',
'centroid_coords_distances',
'centroid_points_distances',
'mean_centroids_coords_distances',
'mean_centroids_points_distances',
'nearest_street_distance_per_service',
'common_nearest_street_distance',
'points_area',
'polar_coords',
]
supported_classifiers = [
'Baseline',
'NaiveBayes',
'NearestNeighbors',
'LogisticRegression',
'SVM',
'MLP',
'DecisionTree',
'RandomForest',
'ExtraTrees',
'XGBoost'
]
included_classifiers = [
'Baseline',
'NaiveBayes',
'NearestNeighbors',
'LogisticRegression',
'SVM',
'MLP',
'DecisionTree',
'RandomForest',
'ExtraTrees',
'XGBoost'
]
NB_hparams = {}
NN_hparams = {
'n_neighbors': [2, 3, 5, 10]
}
LR_hparams = {
'max_iter': [100, 500],
'C': [0.001, 0.1, 1, 10, 1000]
}
SVM_hparams = [
{'kernel': ['rbf', 'sigmoid'], 'gamma': [1e-2, 1e-3, 1e-4, 1e-5],
'C': [0.001, 0.01, 0.1, 1, 10, 25, 50, 100, 1000], 'probability': [True]},
{'kernel': ['poly'], 'degree': [1, 2, 3], 'gamma': ['scale', 'auto'],
'C': [0.1, 1, 10, 25, 50, 100, 1000], 'max_iter': [10000], 'probability': [True]},
]
MLP_hparams = {
'hidden_layer_sizes': [(100, ), (50, 50, )],
# 'learning_rate_init': [0.0001, 0.01, 0.1],
'max_iter': [500, 1000],
'solver': ['sgd', 'adam'],
}
DT_hparams = {
'max_depth': [1, 4, 16, 32, 64],
'min_samples_split': [2, 5, 10, 20, 50, 100, 200],
'min_samples_leaf': [1, 2, 5, 10],
'max_features': list(np.arange(2, 11, 2)) + ["sqrt", "log2", None]
}
RF_hparams = {
'max_depth': [5, 10, 50, 100, 250, 300],
'n_estimators': [100, 250, 500, 1000],
# 'min_samples_leaf': [1, 5, 10],
'min_samples_split': [2, 10],
}
ET_hparams = {
'max_depth': [5, 10, 50, 100, 250, 300],
'n_estimators': [100, 250, 500, 1000],
# 'min_samples_leaf': [1, 5, 10],
'min_samples_split': [2, 10, 50],
}
XGB_hparams = {
"n_estimators": [500, 1000, 3000],
'max_depth': [5, 10, 50, 100, 250, 300],
# # hyperparameters to avoid overfitting
# 'eta': list(np.linspace(0.01, 0.3, 10)), # 'learning_rate'
# 'gamma': [0, 1, 5],
# 'subsample': [0.8, 0.9, 1],
# # Values from 0.3 to 0.8 if you have many columns (especially if you did one-hot encoding),
# # or 0.8 to 1 if you only have a few columns
# 'colsample_bytree': list(np.linspace(0.8, 1, 3)),
# 'min_child_weight': [1, 5, 10],
}
# These parameters constitute the search space for RandomizedSearchCV in our experiments.
NB_hparams_dist = {}
NN_hparams_dist = {
'n_neighbors': sp_randint(1, 20)
}
LR_hparams_dist = {
'max_iter': sp_randint(100, 1000),
'C': expon(scale=200)
}
SVM_hparams_dist = {
'C': expon(loc=0.01, scale=20),
# "C": uniform(2, 10),
"gamma": uniform(1e-5, 1e-2),
'kernel': ['rbf', 'poly', 'sigmoid'],
'degree': sp_randint(1, 3),
'class_weight': ['balanced', None],
'tol': [1e-3, 1e-4],
'max_iter': [100000],
'probability': [True],
# 'dual': [True, False]
}
DT_hparams_dist = {
'max_depth': sp_randint(10, 200),
'min_samples_split': sp_randint(2, 51),
'min_samples_leaf': sp_randint(1, 15),
# 'max_features': sp_randint(1, 11),
}
RF_hparams_dist = {
'bootstrap': [True, False],
# 'max_depth': [10, 20, 30, 40, 50, 60, 100, None],
'max_depth': sp_randint(10, 300),
"n_estimators": sp_randint(250, 2000),
# 'criterion': ['gini', 'entropy'],
# 'max_features': ['sqrt', 'log2'], # sp_randint(1, 11)
'min_samples_leaf': sp_randint(1, 10),
'min_samples_split': sp_randint(2, 50),
}
XGB_hparams_dist = {
"n_estimators": sp_randint(500, 4000),
'max_depth': sp_randint(3, 300),
# 'eta': expon(loc=0.01, scale=0.1), # 'learning_rate'
# hyperparameters to avoid overfitting
'gamma': sp_randint(0, 5),
'subsample': truncnorm(0.8, 1),
'colsample_bytree': truncnorm(0.8, 1),
'min_child_weight': sp_randint(1, 10),
}
MLP_hparams_dist = {
'learning_rate_init': expon(loc=0.0001, scale=0.1),
'max_iter': sp_randint(500, 2000),
'solver': ['sgd', 'adam']
}
