# Simple NN using TensorFlow
# April 2019
# Eduardo Moura Cirilo Rocha
# compate optimizers and learning rates
import numpy as np
import pandas as pd # load csv data
from sklearn.model_selection import train_test_split # split into training and test set
import dataset as ds # load dataset
import matplotlib.pyplot as plt # plotting
import NN_Sigmoid as nn
# load dataset
noShow = ds.import_data_df([ds._FILE_PATHS['merged']])
noShow_X, noShow_y = noShow.iloc[:, :-1].values, noShow.iloc[:, -1].values
noShow_y = np.array([[i] for i in noShow_y])
# separate training and test set randomly keeping classes ratio
trainX, testX, trainY, testY = train_test_split(noShow_X,
noShow_y,
test_size=0.2,
random_state=42,
stratify=noShow_y)
# separate training in validation and training set
trainX, valX, trainY, valY = train_test_split(trainX,
trainY,
test_size=0.15,
random_state=42,
stratify=trainY)
def k_fold_cross_validation(k, hiddenLayers, numEpochs):
# load dataset
noShow = ds.import_data_df([ds._FILE_PATHS['merged']])
noShow_X, noShow_y = noShow.iloc[:, :-1].values, noShow.iloc[:, -1].values
noShow_y = np.array([[i] for i in noShow_y])
# Stratified k-fold
skf = StratifiedKFold(n_splits=k, shuffle=True)
skf.get_n_splits(noShow_X, noShow_y)
#print(skf)
# store results
losses = []
accuracies = []
f1s = []
fold = 1
for train_index, test_index in skf.split(noShow_X, noShow_y):
#print("Fold: "+str(fold))
fold += 1
trainX, testX = noShow_X[train_index], noShow_X[test_index]
trainY, testY = noShow_y[train_index], noShow_y[test_index]
# separate training in validation and training set
trainX, valX, trainY, valY = train_test_split(trainX,
trainY,
test_size=0.15,
random_state=42,
stratify=trainY)
# number of features
numFeatures = trainX.shape[1]
# number of classes
numLabels = trainY.shape[1]
# init
NN = nn.NN_Sigmoid(hiddenLayers,
numFeatures,
numLabels,
learning_rate=0.05,
cross_entropy_weight=4,
optimizer="Adam")
# train
NN.train(numEpochs,
trainX,
trainY,
valX=valX,
valY=valY,
val_epochs=25,
val_patience=5)
# test
_, loss, acc, f1 = NN.predict(testX, testY)
# close tf session
NN.close_session()
losses.append(loss)
accuracies.append(acc)
f1s.append(f1)
return losses, accuracies, f1s