# Author: vkaff
# E-mail: vkaffes@imis.athena-innovation.gr
import numpy as np
from scipy.stats import randint as sp_randint, expon, truncnorm
#: str: Relative path to the datasets.
dataset = 'data/polypairs_dataset.csv'
#: int: Seed used by each of the random number generators.
seed_no = 2020
#: float: Proportion of the dataset to include in the test split. Accepted values should be between 0.0 and 1.0.
test_split_thres = 0.2
[docs]class MLConf:
"""
This class initializes parameters that correspond to the machine learning part of the framework.
These variables define the parameter grid for GridSearchCV:
:cvar SVM_hyperparameters: Defines the search space for SVM.
:vartype SVM_hyperparameters: :obj:`list`
:cvar MLP_hyperparameters: Defines the search space for MLP.
:vartype MLP_hyperparameters: :obj:`dict`
:cvar DecisionTree_hyperparameters: Defines the search space for Decision Trees.
:vartype DecisionTree_hyperparameters: :obj:`dict`
:cvar RandomForest_hyperparameters: Defines the search space for Random Forests and Extra-Trees.
:vartype RandomForest_hyperparameters: :obj:`dict`
:cvar XGBoost_hyperparameters: Defines the search space for XGBoost.
:vartype XGBoost_hyperparameters: :obj:`dict`
These variables define the parameter grid for RandomizedSearchCV where continuous distributions are used for
continuous parameters (whenever this is feasible):
:cvar SVM_hyperparameters_dist: Defines the search space for SVM.
:vartype SVM_hyperparameters_dist: :obj:`dict`
:cvar MLP_hyperparameters_dist: Defines the search space for MLP.
:vartype MLP_hyperparameters_dist: :obj:`dict`
:cvar DecisionTree_hyperparameters_dist: Defines the search space for Decision Trees.
:vartype DecisionTree_hyperparameters_dist: :obj:`dict`
:cvar RandomForest_hyperparameters_dist: Defines the search space for Random Forests and Extra-Trees.
:vartype RandomForest_hyperparameters_dist: :obj:`dict`
:cvar XGBoost_hyperparameters_dist: Defines the search space for XGBoost.
:vartype XGBoost_hyperparameters_dist: :obj:`dict`
"""
kfold_parameter = 5 #: int: The number of outer folds that splits the dataset for the k-fold cross-validation.
n_jobs = -1 #: int: Number of parallel jobs to be initiated. -1 means to utilize all available processors.
#: bool: Whether to build additional features or not, i.e., convex hull of polygons and dist of centroids.
extra_features = True
# accepted values: randomized, grid, hyperband - not yet implemented!!!
hyperparams_search_method = 'randomized'
"""str: Search Method to use for finding best hyperparameters. (*randomized* | *grid*).
See Also
--------
:func:`~polygon_classification.param_tuning.ParamTuning.fineTuneClassifiers`. Details on available inputs.
"""
#: int: Number of iterations that RandomizedSearchCV should execute. It applies only when
#: :attr:`hyperparams_search_method` equals to 'randomized'.
max_iter = 3
score = 'accuracy'
"""str: The metric to optimize on hyper-parameter tuning. Possible valid values presented on `Scikit predefined values`_.
.. _Scikit predefined values:
https://scikit-learn.org/stable/modules/model_evaluation.html#common-cases-predefined-values
"""
classifiers = [
# 'SVM',
# 'DecisionTree',
'RandomForest',
# 'ExtraTrees',
# 'XGBoost',
# 'MLP'
]
"""list of str: Define the classifiers to apply on code execution. Accepted values are:
- SVM
- DecisionTree
- RandomForest
- ExtraTrees
- XGBoost
- MLP.
"""
clf_custom_params = {
'SVM': {
# with scaler
# basic/extra
'C': 200, 'class_weight': 'balanced', 'gamma': 10, 'max_iter': 10000,
# 'kernel': 'rbf', 'tol': 0.001
},
'DecisionTree': {
# 'min_samples_leaf': 0.10218472045491575, 'min_samples_split': 0.46848801022523695, 'max_features': 10,
# 'class_weight': {1: 1, 4: 9}, 'max_depth': 70,
# with scaler
# 'max_features': 10, 'min_samples_leaf': 0.13084191454406335, 'class_weight': {1: 1, 4: 2}, 'max_depth': 79,
# 'min_samples_split': 0.7040970996893269,
# basic
# 'min_samples_leaf': 0.13896623393837215, 'class_weight': {1: 1, 4: 7}, 'max_depth': 42, 'max_features': 5,
# 'min_samples_split': 0.21705549723971926,
# extra
'class_weight': {1: 1, 4: 2}, 'max_depth': 80, 'max_features': 10, 'min_samples_leaf': 2,
'min_samples_split': 10,
# 'splitter': 'best',
},
'RandomForest': {
# with scaler
# basic
# 'max_depth': 62, 'min_samples_split': 2, 'n_estimators': 553, 'max_features': 'sqrt', 'bootstrap': False,
# 'criterion': 'entropy', 'min_samples_leaf': 2, 'class_weight': {1: 1, 4: 7},
# extra
'class_weight': {1: 2, 4: 1}, 'criterion': 'entropy', 'max_depth': 100, 'n_estimators': 1000,
# 'min_samples_split': 2,
# 'random_state': seed_no, 'n_jobs': n_jobs, # 'oob_score': True,
},
'ExtraTrees': {
# with scaler
# basic
# 'max_depth': 71, 'bootstrap': False, 'criterion': 'gini', 'class_weight': {1: 1, 4: 1},
# 'min_samples_leaf': 1, 'max_features': 'sqrt', 'min_samples_split': 7, 'n_estimators': 776,
# extra
'class_weight': {1: 2, 4: 1}, 'max_depth': 100,
# 'random_state': seed_no, 'n_jobs': n_jobs
},
'XGBoost': {
# with scaler
# basic
# 'n_estimators': 2549, 'min_child_weight': 1, 'max_depth': 62, 'scale_pos_weight': 1,
# 'colsample_bytree': 0.598605740971479, 'gamma': 1, 'eta': 0.17994840726392214,
# 'subsample': 0.7250606565532803,
# extra
'max_depth': 72, 'n_estimators': 21, 'scale_pos_weight': 3,
# 'random_state': seed_no, 'nthread': n_jobs, 'objective': "binary:logistic",
},
'MLP': {
# with scaler
# basic
# 'activation': 'logistic', 'solver': 'lbfgs', 'max_iter': 1000, 'tol': 0.0001,
# 'learning_rate_init': 0.16533315728128767,
# extra
# 'activation': 'relu', 'hidden_layer_sizes': (100,),
'learning_rate_init': 0.05, 'max_iter': 5000, 'solver': 'lbfgs', 'tol': 0.003,
}
}
# These parameters constitute the search space for GridSearchCV in our experiments.
SVM_hyperparameters = [
{
'kernel': ['rbf', 'sigmoid'],
'gamma': [1e-2, 0.1, 1, 5, 10, 30, 'scale'],
'C': [0.01, 0.1, 1, 10, 100, 200, 300],
'tol': [1e-3, 1e-2],
# 'probability': [True, False],
'max_iter': [5000],
'class_weight': [None, 'balanced', {1: 2, 4: 1}, {1: 3, 4: 1}],
},
{
'kernel': ['poly'],
'gamma': ['auto', 'scale', 1, 10, 30],
'C': [0.01, 0.1, 1, 10, 100, 200, 300],
'degree': [1, 2, 3], # degree=1 is the same as using a 'linear' kernel
'tol': [1e-3, 1e-2],
# 'probability': [True, False],
'max_iter': [5000],
'class_weight': [None, 'balanced', {1: 2, 4: 1}, {1: 3, 4: 1}],
},
# {'kernel': ['linear'], 'C': [0.001, 0.01, 0.1, 1, 10, 100, 1000], 'max_iter': [3000]}
]
DecisionTree_hyperparameters = {
'max_depth': [2, 3, 5, 10, 30, 50, 60, 80, 100],
'min_samples_split': [2, 4, 6, 10, 15, 25, 50],
'min_samples_leaf': [1, 2, 4, 10],
# 'min_samples_split': list(np.linspace(0.1, 1, 10)),
# 'min_samples_leaf': list(np.linspace(0.1, 0.5, 5)),
'max_features': list(np.arange(2, 11, 2)) + ["sqrt", "log2"],
'splitter': ['best', 'random'],
'class_weight': [None, 'balanced', {1: 2, 4: 1}, {1: 3, 4: 1}],
}
RandomForest_hyperparameters = {
# 'bootstrap': [True, False],
'max_depth': [5, 10, 20, 50, 70, 100],
'criterion': ['gini', 'entropy'],
# 'max_features': ['log2', 'sqrt'], # auto is equal to sqrt
# 'min_samples_leaf': [1, 2, 4],
'min_samples_split': [2, 3, 5],
"n_estimators": [100, 120, 200, 250],
'class_weight': [None, 'balanced', {1: 2, 4: 1}, {1: 3, 4: 1}],
}
XGBoost_hyperparameters = {
"n_estimators": [20, 30, 100, 300],
# 'eta': list(np.linspace(0.01, 0.2, 10)), # 'learning_rate'
## avoid overfitting
# Control the model complexity
'max_depth': [10, 30, 50, 70, 100],
'gamma': [0, 1, 5],
# 'reg_alpha': [1, 10],
# Add randomness to make training robust to noise
'subsample': [0.8, 0.9, 1],
# 'colsample_bytree': list(np.linspace(0.8, 1, 3)),
# for class imbalances, setting the parameters
# *) 'max_delta_step',
# *) 'min_child_weight' and
# *) 'scale_pos_weight'
# could help
'scale_pos_weight': [1, 2, 3],
# 'min_child_weight': [1, 5, 10],
}
MLP_hyperparameters = {
'hidden_layer_sizes': [(100,), (50, 50,)],
'learning_rate_init': [0.0001, 0.005, 0.01, 0.05, 0.1],
'max_iter': [3000],
'solver': ['lbfgs', 'sgd', 'adam'],
'activation': ['identity', 'logistic', 'tanh', 'relu'],
'tol': [1e-3, 1e-4],
}
# These parameters constitute the search space for RandomizedSearchCV in our experiments.
SVM_hyperparameters_dist = {
'C': expon(scale=100),
'gamma': expon(scale=.1),
'kernel': ['rbf', 'poly', 'sigmoid'],
'class_weight': ['balanced', None] + [{1: w, 4: 1} for w in range(1, 5)],
'degree': [1, 2, 3],
'tol': [1e-3, 1e-2],
'max_iter': [10000]
}
DecisionTree_hyperparameters_dist = {
'max_depth': sp_randint(10, 200),
'min_samples_split': sp_randint(2, 200),
'min_samples_leaf': sp_randint(1, 10),
'max_features': sp_randint(1, 11),
'class_weight': [None, 'balanced'] + [{1: w, 4: 1} for w in range(1, 5)],
}
RandomForest_hyperparameters_dist = {
# 'bootstrap': [True, False],
'max_depth': sp_randint(3, 200),
'criterion': ['gini', 'entropy'],
'max_features': ['sqrt', 'log2'], # sp_randint(1, 11)
'min_samples_leaf': sp_randint(1, 10),
'min_samples_split': sp_randint(2, 21),
"n_estimators": sp_randint(250, 1000),
'class_weight': ['balanced', None] + [{1: w, 4: 1} for w in range(1, 5)],
}
XGBoost_hyperparameters_dist = {
"n_estimators": sp_randint(20, 200),
# 'eta': expon(loc=0.01, scale=0.1), # 'learning_rate'
# hyperparameters to avoid overfitting
'max_depth': sp_randint(10, 200),
'gamma': sp_randint(0, 5),
'subsample': truncnorm(0.4, 0.7),
# 'colsample_bytree': truncnorm(0.8, 1),
# 'min_child_weight': sp_randint(1, 10),
'scale_pos_weight': sp_randint(1, 5),
"reg_alpha": truncnorm(0, 2),
'reg_lambda': sp_randint(1, 20),
}
MLP_hyperparameters_dist = {
'hidden_layer_sizes': [(100,), (50, 50,)],
'learning_rate_init': expon(loc=0.0001, scale=0.1),
'max_iter': [3000],
'solver': ['lbfgs', 'sgd', 'adam'],
'activation': ['identity', 'logistic', 'tanh', 'relu'],
'tol': [1e-3, 1e-4],
}
