def run_and_get_classifier(X_test, X_train, y_test, y_train):
# Declare the fitness function we want to use
def fit(y_true, y_pred):
# y_pred are floats in [0, n_classes-1]. To use accuracy metric we need
# to binarize the output using round_to_cls()
y_pred_bin = round_to_cls(y_pred, n_classes=2)
return accuracy_score(y_true, y_pred_bin)
# Initialize our classifier
clf = TrefleClassifier(
n_rules=4,
n_classes_per_cons=[2], # there is only 1 consequent with 2 classes
n_labels_per_mf=3, # use 3 labels LOW, MEDIUM, HIGH
default_cons=[0], # default rule yield the class 0
n_max_vars_per_rule=3, # WBCD dataset has 30 variables, here we force
# to use a maximum of 3 variables per rule
# to have a better interpretability
# In total we can have up to 3*4=12 different variables
# for a fuzzy system
n_generations=5,
fitness_function=fit,
verbose=True,
)
# Train our classifier
clf.fit(X_train, y_train)
# Make predictions
y_pred = clf.predict_classes(X_test)
clf.print_best_fuzzy_system()
print_score(y_pred, y_test)
return clf
def _create_fis(self):
def fit(y_true, y_pred):
rmse = -mean_squared_error(y_true, y_pred)
return rmse
# Initialize our classifier
clf = TrefleClassifier(
n_rules=3,
n_classes_per_cons=[3],
default_cons=[1],
n_max_vars_per_rule=4,
n_generations=5,
pop_size=100,
n_labels_per_mf=3,
dc_weight=2,
fitness_function=fit,
)
clf.fit(self._X_train, self._y_train)
return clf
def main():
np.random.seed(0)
random.seed(0)
# Load dataset
data = load_boston()
# Organize our data
X = data["data"]
print(X.shape)
y = data["target"]
y = np.reshape(y, (-1, 1)) # output needs to be at least 1 column wide
# Split our data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
# Declare the fitness function we want to use
def fit(y_true, y_pred):
# Here no need to threshold y_pred because we are using a regression
# metric
return -mean_squared_error(y_true, y_pred)
# Initialize our classifier
clf = TrefleClassifier(
n_rules=5,
n_classes_per_cons=[0], # In regression, there is no class (i.e. 0)
n_labels_per_cons=Label4, # use 4 labels LOW, MEDIUM, HIGH, VERY HIGH
# # for the consequent
# # Recall: even for continuous variables
# # we use a label e.g.
# # "[...] THEN temperature is LOW"
n_labels_per_mf=2, # use 2 labels LOW, HIGH (for the antecedents)
default_cons=[Label4.VERY_HIGH()], # default rule yield the 4th (and last) label
n_max_vars_per_rule=2,
n_generations=30,
fitness_function=fit,
verbose=True,
)
# Train our classifier
clf.fit(X_train, y_train)
# Make predictions
y_pred = clf.predict_classes(X_test)
# Alternatively, you can use predict() which return non-thresholded y_pred
# but you could need to add a threshold yourself. For example:
# y_pred_raw = clf.predict(X_test)
# y_pred = round_to_cls(y_pred_raw, n_classes=2)
clf.print_best_fuzzy_system()
# Evaluate accuracy
score = mean_squared_error(y_test, y_pred)
print("Score on test set: {:.3f}".format(score))
def main():
np.random.seed(0)
random.seed(0)
# Load dataset
data = load_breast_cancer()
# Organize our data
X = data["data"]
print(X.shape)
y = data["target"]
y = np.reshape(y, (-1, 1)) # output needs to be at least 1 column wide
# Split our data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
# Declare the fitness function we want to use
def fit(y_true, y_pred):
# y_pred are floats in [0, n_classes-1]. To use accuracy metric we need
# to binarize the output using round_to_cls()
y_pred_bin = round_to_cls(y_pred, n_classes=2)
return accuracy_score(y_true, y_pred_bin)
# Initialize our classifier
clf = TrefleClassifier(
n_rules=4,
n_classes_per_cons=[2], # there is only 1 consequent with 2 classes
n_labels_per_mf=3, # use 3 labels LOW, MEDIUM, HIGH
default_cons=[0], # default rule yield the class 0
n_max_vars_per_rule=3, # WBCD dataset has 30 variables, here we force
# to use a maximum of 3 variables per rule
# to have a better interpretability
# In total we can have up to 3*4=12 different variables
# for a fuzzy system
n_generations=20,
fitness_function=fit,
verbose=True,
)
# Train our classifier
clf.fit(X_train, y_train)
# Make predictions
y_pred = clf.predict_classes(X_test)
# Alternatively, you can use predict() which return non-thresholded y_pred
# but you could need to add a threshold yourself. For example:
# y_pred_raw = clf.predict(X_test)
# y_pred = round_to_cls(y_pred_raw, n_classes=2)
clf.print_best_fuzzy_system()
# Evaluate accuracy
score = accuracy_score(y_test, y_pred)
print("Score on test set: {:.3f}".format(score))
def main():
np.random.seed(0)
random.seed(0)
# Load dataset
data = load_iris()
# Organize our data
X = data["data"]
y = data["target"] # y.shape is (150,)
y = create_one_hot_from_array(y) # y.shape is now (150,3)
# Split our data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
# Declare the fitness function we want to use
def fit(y_true, y_pred):
# y_pred are floats in [0, n_classes-1]. To use accuracy metric we need
# to binarize the output using round_to_cls()
# Warning /!\ here since it has been one-hot-encoded we need to set
# n_classes=2 instead n_classes=N_CLASSES because each consequent
# is a binary class
y_pred_bin = round_to_cls(y_pred, n_classes=2)
return accuracy_score(y_true, y_pred_bin)
# Initialize our classifier
clf = TrefleClassifier(
n_rules=3, # here we need to increase the number of rule to 3
# # because we need at least 1 rule per class in the case
# # of a one-hot-encoded problem
n_classes_per_cons=[2, 2, 2], # there are 3 consequents with 2 classes
# # each.
n_labels_per_mf=4, # use 4 labels LOW, MEDIUM, HIGH, VERY HIGH
default_cons=[0, 0, 1], # default rule yield the class 2
n_max_vars_per_rule=4, # let's use the 4 iris variables (PL, PW, SL, SW)
n_generations=30,
fitness_function=fit,
verbose=True,
)
# Train our classifier
clf.fit(X_train, y_train)
# Make predictions
# y_pred = clf.predict_classes(X_test)
y_pred_raw = clf.predict(X_test)
y_pred = round_to_cls(y_pred_raw, n_classes=2)
clf.print_best_fuzzy_system()
# Evaluate accuracy
# Important /!\ the fitness can be different than the scoring function
score = accuracy_score(y_test, y_pred)
print("Score on test set: {:.3f}".format(score))
def main():
np.random.seed(0)
random.seed(0)
# Load dataset
data = load_iris()
N_CLASSES = 3
# Organize our data
X = data["data"]
y = data["target"]
y = np.reshape(y, (-1, 1)) # output needs to be at least 1 column wide
# Split our data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
# Declare the fitness function we want to use
def fit(y_true, y_pred):
# y_pred are floats in [0, n_classes-1]. To use accuracy metric we need
# to binarize the output using round_to_cls()
y_pred_bin = round_to_cls(y_pred, N_CLASSES)
return accuracy_score(y_true, y_pred_bin)
# Initialize our classifier
clf = TrefleClassifier(
n_rules=2,
n_classes_per_cons=[N_CLASSES], # there is only 1
# # consequent with 3 classes
n_labels_per_mf=4, # use 4 labels LOW, MEDIUM, HIGH, VERY HIGH
default_cons=[1], # default rule yield the class 1
n_max_vars_per_rule=4, # let's use the 4 iris variables (PL, PW, SL, SW)
n_generations=30,
fitness_function=fit,
verbose=True,
)
# Train our classifier
clf.fit(X_train, y_train)
# Make predictions
y_pred = clf.predict_classes(X_test)
clf.print_best_fuzzy_system()
# Evaluate f1 score.
# Important /!\ the fitness can be different than the scoring function
score = f1_score(y_test, y_pred, average="weighted")
print("Score on test set: {:.3f}".format(score))
def run():
import numpy as np
import random
np.random.seed(6)
random.seed(6)
# Load dataset
data = load_breast_cancer()
# data = load_iris()
# Organize our data
y_names = data["target_names"]
print("target names", y_names)
y = data["target"]
y = y.reshape(-1, 1)
X_names = data["feature_names"]
print("features names", X_names)
X = data["data"]
# X, y = make_classification(
# n_samples=1000, n_features=10, n_informative=5, n_classes=2
# )
# y = y.reshape(-1, 1)
# multi_class_y_col = np.random.randint(0, 4, size=y.shape)
# regr_y_col = np.random.random(size=y.shape) * 100 + 20
# y = np.hstack((y, multi_class_y_col, regr_y_col))
# print(y)
# Split our data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
def fit(y_true, y_pred):
y_pred_thresholded = round_to_cls(y_pred, n_classes=2)
fitness_val = accuracy_score(y_true, y_pred_thresholded)
return fitness_val
# Initialize our classifier
clf = TrefleClassifier(
n_rules=3,
n_classes_per_cons=[2],
default_cons=[1],
n_max_vars_per_rule=3,
n_generations=10,
pop_size=100,
n_labels_per_mf=3,
verbose=True,
dc_weight=1,
# p_positions_per_lv=16,
n_lv_per_ind_sp1=40,
fitness_function=fit,
)
# Train our classifier
model = clf.fit(X_train, y_train)
# Make predictions
y_pred = clf.predict(X_test)
clf.print_best_fuzzy_system()
tff_str = clf.get_best_fuzzy_system_as_tff()
print(tff_str)
yolo = TrefleFIS.from_tff(tff_str)
yolo.describe()
# fis = clf.get_best_fuzzy_system()
# print("best fis is ", end="")
# print(fis)
# FISViewer(fis).show()
# Evaluate accuracy
print("Simple run score: ")
y_pred_thresholded = round_to_cls(y_pred, n_classes=2)
print("acc", accuracy_score(y_test, y_pred_thresholded))
print(classification_report(y_test, y_pred_thresholded))
def main():
"""
Executing this method is time consuming
"""
np.random.seed(0)
random.seed(0)
# Load dataset
data = load_breast_cancer()
# Organize our data
X = data["data"]
print(X.shape)
y = data["target"]
y = np.reshape(y, (-1, 1)) # output needs to be at least 1 column wide
# Split our data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
# Declare the fitness function we want to use
def evaluate(y_true, y_pred):
# y_pred are floats in [0, n_classes-1]. To use accuracy metric we need
# to binarize the output using round_to_cls()
y_pred_bin = round_to_cls(y_pred, n_classes=2)
return accuracy_score(y_true, y_pred_bin)
# Initialize our classifier
# note that some arguments are mandatory such as n_rules. If you want to
# grid search that parameter you must set it to None before tuning it.
estimator = TrefleClassifier(
n_rules=None, # mandatory argument, set to None and tune it below
n_classes_per_cons=[2], # there is only 1 consequent with 2 classes
n_labels_per_mf=3, # use 3 labels LOW, MEDIUM, HIGH
default_cons=None,
n_max_vars_per_rule=3, # WBCD dataset has 30 variables, here we force
# to use a maximum of 3 variables per rule
# to have a better interpretability
# In total we can have up to 3*4=12 different variables
# for a fuzzy system
n_generations=20,
fitness_function=evaluate,
verbose=True,
)
def get_fitness_functions():
def get_recall_and_precision_score(y_true, y_pred):
y_pred_bin = round_to_cls(y_pred, n_classes=2)
recall = recall_score(y_true, y_pred_bin)
precision = precision_score(y_true, y_pred_bin)
return recall, precision
def ff1(y_true, y_pred):
recall, precision = get_recall_and_precision_score(y_true, y_pred)
return (1.0 * recall + 2.0 * precision) / 3.0
def ff2(y_true, y_pred):
recall, precision = get_recall_and_precision_score(y_true, y_pred)
return (2.0 * recall + 1.0 * precision) / 3.0
def ff3(y_true, y_pred):
recall, precision = get_recall_and_precision_score(y_true, y_pred)
return (1.0 * recall + 3.0 * precision) / 4.0
return ff1, ff2, ff3
tuned_parameters = [
{"n_rules": [2, 5], "default_cons": [[0], [1]]},
{
"n_rules": [4, 5],
"default_cons": [[0], [1]],
"n_max_vars_per_rule": [3, 5, 6],
},
{
"n_rules": [4],
"default_cons": [[0]],
"fitness_function": get_fitness_functions(),
},
]
# Note that the scoring reuses the evaluate function but fitness_function
# (i.e. the function that compares models for a given configuration/run)
# can be different than scoring function (i.e. the function that compares
# best individuals between different configurations/runs)
clf = GridSearchCV(estimator, tuned_parameters, cv=3, scoring=make_scorer(evaluate))
# Train our classifier
clf.fit(X_train, y_train)
print("Best params: ")
print(clf.best_params_)
best_estimator = clf.best_estimator_
y_pred_test = best_estimator.predict_classes(X_test)
print(accuracy_score(y_true=y_test, y_pred=y_pred_test))
best_estimator.print_best_fuzzy_system()