84 lines
3.2 KiB
Python
84 lines
3.2 KiB
Python
|
|
# Ce script montre un exemple pratique de l'overfit
|
|
# Le premier ANN entraîné a une structure complexe avec beaucoup de neurones
|
|
# Du coup il arrive à obtenir une très bonne précision sur le jeu d'entraînement (environ 96%)
|
|
# Mais il n'arrive pas à généraliser sur le jeu de test (précision à 81%)
|
|
# C'est un cas classique d'overfitting (surentraînement)
|
|
#
|
|
# Le deuxième ANN utilise Dropout pour réduire ce problème.
|
|
# La précision d'entraînement est plus faible (90%) et
|
|
# la précision de test a augmenté (à 85%)
|
|
# On a toujours un problème de surentraînement, mais bien mondre.
|
|
|
|
# Import libraries
|
|
import pandas as pd
|
|
|
|
# Import data
|
|
dataset = pd.read_csv('data/Churn_Modelling.csv')
|
|
X = dataset.iloc[:, 3:13]
|
|
y = dataset.iloc[:, 13]
|
|
|
|
# Encode categorical data and scale continuous data
|
|
from sklearn.preprocessing import OneHotEncoder, StandardScaler
|
|
from sklearn.compose import make_column_transformer
|
|
preprocess = make_column_transformer(
|
|
(OneHotEncoder(), ['Geography', 'Gender']),
|
|
(StandardScaler(), ['CreditScore', 'Age', 'Tenure', 'Balance',
|
|
'NumOfProducts', 'HasCrCard', 'IsActiveMember',
|
|
'EstimatedSalary']))
|
|
X = preprocess.fit_transform(X)
|
|
|
|
# Split in train/test
|
|
y = y.values
|
|
from sklearn.model_selection import train_test_split
|
|
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.2, random_state = 0)
|
|
|
|
# Part 2 - Now let's make the ANN!
|
|
|
|
# Importing the Keras libraries and packages
|
|
from keras.models import Sequential
|
|
from keras.layers import Dense
|
|
from keras.layers import Dropout
|
|
|
|
# ANN - Overfitting
|
|
classifier = Sequential()
|
|
classifier.add(Dense(units = 128, kernel_initializer = 'uniform', activation = 'relu', input_dim = 13))
|
|
classifier.add(Dense(units = 64, kernel_initializer = 'uniform', activation = 'relu'))
|
|
classifier.add(Dense(units = 32, kernel_initializer = 'uniform', activation = 'relu'))
|
|
classifier.add(Dense(units = 16, kernel_initializer = 'uniform', activation = 'relu'))
|
|
classifier.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))
|
|
classifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
|
|
classifier.fit(X_train, y_train, batch_size = 10, epochs = 100)
|
|
score_train = classifier.evaluate(X_train, y_train) # 96.6%
|
|
score_test = classifier.evaluate(X_test, y_test) # 80.7%
|
|
|
|
# ANN - Dropout
|
|
|
|
classifier = Sequential()
|
|
classifier.add(Dense(units = 128, kernel_initializer = 'uniform', activation = 'relu', input_dim = 13))
|
|
classifier.add(Dropout(0.2))
|
|
classifier.add(Dense(units = 64, kernel_initializer = 'uniform', activation = 'relu'))
|
|
classifier.add(Dropout(0.2))
|
|
classifier.add(Dense(units = 32, kernel_initializer = 'uniform', activation = 'relu'))
|
|
classifier.add(Dropout(0.2))
|
|
classifier.add(Dense(units = 16, kernel_initializer = 'uniform', activation = 'relu'))
|
|
classifier.add(Dropout(0.2))
|
|
classifier.add(Dense(units = 1, kernel_initializer = 'uniform', activation = 'sigmoid'))
|
|
classifier.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
|
|
classifier.fit(X_train, y_train, batch_size = 10, epochs = 100)
|
|
score_train = classifier.evaluate(X_train, y_train) # 90.6%
|
|
score_test = classifier.evaluate(X_test, y_test) # 85.7%
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|