def a5():
directories.A5()
# check tensorflow's GPU support
physical_devices = tf.config.experimental.list_physical_devices('GPU')
print("Num GPUs Available",
len(tf.config.experimental.list_physical_devices('GPU')))
tf.config.experimental.set_memory_growth(physical_devices[0], True)
features = 784
classes = 10
fltr = [3, 5]
h1 = [128, 384, 749]
# φόρτωση mnist από το keras
(x_train, y_train), (x_test, y_test) = mnist.load_data()
# κάνουμε το mnist reshape
x_train = x_train.reshape(x_train.shape[0], 28, 28, 1)
x_test = x_test.reshape(x_test.shape[0], 28, 28, 1)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
# ορισμός των labels
y_train = to_categorical(y_train, classes)
y_test = to_categorical(y_test, classes)
for fl in fltr:
for nodes in h1:
print(
f'Μοντέλο CNN με ένα επίπεδο 32 φίλτρων {fl}x{fl}, 2x2 MaxPooling και MLP {nodes}:10'
)
# model
model = Sequential()
# first layer: Convolution 2D, 32 filters of size 5x5
model.add(
Conv2D(32, (fl, fl),
input_shape=(28, 28, 1),
activation='relu',
kernel_initializer='he_uniform'))
# second layer: MaxPooling 2D, returns max value of image portion (2x2)
model.add(MaxPooling2D(pool_size=(2, 2)))
# third layer: Flatten results of previous layers to feed into the MLP
model.add(Flatten())
# fourth and output layer: our standard MLP
model.add(
Dense(nodes,
kernel_initializer='he_uniform',
activation='relu'))
model.add(
Dense(10,
kernel_initializer='he_uniform',
activation='softmax'))
# compile the model
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
fname = './logs/A5/32({},{})_0.2_{}-10.txt'.format(fl, fl, nodes)
directories.filecheck(fname)
acc = []
vloss = []
tloss = []
acc_sum = 0
loss_sum = 0
f = open(fname, 'w')
sys.stdout = f
fold = 1
kfold = KFold(5, shuffle=True, random_state=1)
for train, test in kfold.split(x_train):
xi_train, xi_test = x_train[train], x_train[test]
yi_train, yi_test = y_train[train], y_train[test]
print(f' fold # {fold}, TRAIN: {train}, TEST: {test}')
history = model.fit(xi_train,
yi_train,
epochs=10,
batch_size=200,
verbose=1,
validation_data=(xi_test, yi_test))
acc.append(history.history['val_accuracy'])
vloss.append(history.history['val_loss'])
tloss.append(history.history['loss'])
# Test the model after training
test_results = model.evaluate(x_test, y_test, verbose=1)
print(
f'Αποτελέσματα στο fold # {fold} - Loss: {test_results[0]} - Accuracy: {test_results[1]}'
)
fold = fold + 1
# save 5-fold cv results
loss_sum += test_results[0]
acc_sum += test_results[1]
# plots
# accuracy
plot_acc = plt.figure(1)
title1 = 'Validation Accuracy, CNN 32({},{}), max pooling, {}-10'.format(
fl, fl, nodes)
plt.title(title1, loc='center', pad=None)
plt.plot(np.mean(acc, axis=0))
plt.ylabel('acc')
plt.xlabel('epoch')
# loss
plot_loss = plt.figure(2)
title2 = 'Loss, CNN 32({},{}), max pooling, {}-10'.format(
fl, fl, nodes)
plt.title(title2, loc='center', pad=None)
plt.plot(np.mean(vloss, axis=0))
# train loss
plt.plot(np.mean(tloss, axis=0))
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['validation', 'train'], loc='upper left')
directories.filecheck('./plots/A5/{}.png'.format(title1))
directories.filecheck('./plots/A5/{}.png'.format(title2))
plot_acc.savefig('./plots/A5/{}.png'.format(title1), format='png')
plot_loss.savefig('./plots/A5/{}.png'.format(title2), format='png')
print(
f'Συνολικά αποτελέσματα- Loss {loss_sum/ 5} - Accuracy {acc_sum / 5}'
)
# επιστροφή stdout στην κονσόλα
f.close()
sys.stdout = sys.__stdout__
# απελευθερωση μνημης
print(f'Clearing session....')
tf.keras.backend.clear_session()
plt.close(1)
plt.close(2)
acc.clear()
vloss.clear()
tloss.clear()
x_train = x_train.reshape(x_train.shape[0], 784)
x_test = x_test.reshape(x_test.shape[0], 784)
# x_train,test are now matrices of matrices
# MinMax scaling
scaler = preprocessing.MinMaxScaler()
x_train = scaler.fit_transform(x_train)
x_test = scaler.fit_transform(x_test)
# ορισμός των labels
y_train = to_categorical(y_train, 10)
y_test = to_categorical(y_test, 10)
user_in = input('Run B4-A? y/n ')
if user_in == 'y':
fname = 'logs/B4/evaluation for best mlp.txt'
directories.filecheck(fname)
f = open(fname, 'w')
sys.stdout = f
butler.evaluate_best_mlp(num_indiv, crossrate, mutrate, x_test, y_test)
# επιστροφή stdout στην κονσόλα
f.close()
sys.stdout = sys.__stdout__
user_in = input('Run B4-B? y/n ')
if user_in == 'y':
x_train_selected, x_test_selected = butler.get_selected(
x_train, x_test, num_indiv, crossrate, mutrate)
# model
# Δημιουργία μοντέλου με χρήση του keras API
model = Sequential()
# Πρώτο κρυφό επίπεδο
def extra_layer(ep):
# αρχικοποίηση directories αποθήκευσης
directories.extra_layer()
# GPU support
physical_devices = tensorflow.config.experimental.list_physical_devices(
'GPU')
print("CUDA - Αριθμός διαθέσιμων GPUs:",
len(tensorflow.config.experimental.list_physical_devices('GPU')))
tensorflow.config.experimental.set_memory_growth(physical_devices[0], True)
# αρχικοποίηση μεταβλητών
features = 784
classes = 10
h1 = 794
h2 = [10, 50, 150, 250]
# φόρτωση mnist από το keras
(x_train, y_train), (x_test, y_test) = mnist.load_data()
# κάνουμε το mnist reshape
x_train = x_train.reshape(x_train.shape[0], features)
x_test = x_test.reshape(x_test.shape[0], features)
# MinMax scaling
scaler = preprocessing.MinMaxScaler()
x_train = scaler.fit_transform(x_train)
x_test = scaler.fit_transform(x_test)
# ορισμός των labels
y_train = to_categorical(y_train, classes)
y_test = to_categorical(y_test, classes)
# ορισμός input shape για το μοντέλο MLP βάσει των χαρακτηριστικών
input_shape = (features, )
# Έλεγχος για όλα τα H2
for h_2 in h2:
loss_sum = 0
acc_sum = 0
f_ce = "./logs/A2/Extra_Layer/results_CE_%s-%s.txt" % (h1, h_2)
f_mse = "./logs/A2/Extra_Layer/results_MSE_%s-%s.txt" % (h1, h_2)
directories.filecheck(f_ce)
directories.filecheck(f_mse)
# Δημιουργία μοντέλων με χρήση του keras API
model_ce = Sequential()
model_mse = Sequential()
# πρώτο κρυφό επίπεδο
model_ce.add(Dense(h1, input_shape=input_shape, activation='relu'))
model_ce.add(Dense(h_2, activation='relu'))
# δεύτερο κρυφό επίπεδο
model_mse.add(Dense(h1, input_shape=input_shape, activation='relu'))
model_mse.add(Dense(h_2, activation='relu'))
# επίπεδο εξόδου
model_ce.add(Dense(classes, activation='softmax'))
model_mse.add(Dense(classes, activation='softmax'))
# compile
# crossentropy
model_ce.compile(loss='categorical_crossentropy',
optimizer='SGD',
metrics=['accuracy'])
# mse
model_mse.compile(loss='mean_squared_error',
optimizer='SGD',
metrics=['accuracy'])
# αρχείο εξόδου
f = open(f_ce, 'w')
print(
'Μοντέλο Cross Entropy Loss για {} κόμβους στο πρώτο κρυφό επίπεδο και {} στο δεύτερο'
.format(h1, h_2))
sys.stdout = f
################################################################################################################
###################################### CROSS ENTROPY 5-FOLD CV #################################################
fold = 1
kfold = KFold(5, shuffle=True, random_state=1)
aval = []
lval = []
ltrain = []
for train, test in kfold.split(x_train):
# διαχωρισμός train-test indexes
xi_train, xi_test = x_train[train], x_train[test]
yi_train, yi_test = y_train[train], y_train[test]
print(f' fold # {fold}, TRAIN: {train}, TEST: {test}')
# fit μοντέλου
ce_history = model_ce.fit(xi_train,
yi_train,
epochs=ep,
batch_size=200,
verbose=1,
validation_data=(xi_test, yi_test))
# στατιστικά
aval.append(ce_history.history['val_accuracy'])
lval.append(ce_history.history['val_loss'])
ltrain.append(ce_history.history['loss'])
# μετρήσεις μοντέλου
ce_results = model_ce.evaluate(x_test, y_test, verbose=1)
print(
f'Αποτελέσματα στο fold # {fold} - Loss: {ce_results[0]} - Accuracy: {ce_results[1]}'
)
fold = fold + 1
# αποθήκευση για προβολή των αποτελεσμάτων 5-fold CV
loss_sum += ce_results[0]
acc_sum += ce_results[1]
# plots
# accuracy
plot_acc = plt.figure(1)
title1 = 'Validation Accuracy Crossentropy Model {}-{}-10'.format(
h1, h_2)
plt.title(title1, loc='center', pad=None)
plt.plot(np.mean(aval, axis=0))
plt.ylabel('acc')
plt.xlabel('epoch')
# loss
plot_loss = plt.figure(2)
title2 = 'Loss Crossentropy Model {}-{}-10'.format(h1, h_2)
plt.title(title2, loc='center', pad=None)
# validation loss
plt.plot(np.mean(lval, axis=0))
# train loss
plt.plot(np.mean(ltrain, axis=0))
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['validation', 'train'], loc='upper left')
# Save locally
directories.filecheck('./plots/A2/Extra_Layer/{}.png'.format(title1))
directories.filecheck('./plots/A2/Extra_Layer/{}.png'.format(title2))
plot_loss.savefig("./plots/A2/Extra_Layer/{}.png".format(title2),
format='png')
plot_acc.savefig("./plots/A2/Extra_Layer/{}.png".format(title1),
format='png')
# εκτυπωση αποτελεσμάτων
print(
f'Συνολικά Αποτελέσματα (Cross Entropy Model)- Loss {loss_sum / 5} - Accuracy {acc_sum / 5}'
)
# επιστροφή stdout στην κονσόλα
f.close()
sys.stdout = sys.__stdout__
# απελευθερωση μνημης
print(f'Clearing session....')
tensorflow.keras.backend.clear_session()
plt.close(1)
plt.close(2)
# αρχικοποίηση καινούριων μεταβλητών
loss_sum = 0
acc_sum = 0
fold = 1
aval.clear()
lval.clear()
ltrain.clear()
# νεο αρχείο εξόδου
f = open(f_mse, 'w')
print(
'Μοντέλο Mean Squared Error Loss για {} κόμβους στο πρώτο κρυφό επίπεδο και {} στο δεύτερο'
.format(h1, h_2))
sys.stdout = f
#######################################################################################################################
#################################### MEAN SQUARED ERROR 5-FOLD CV #####################################################
kfold = KFold(5, shuffle=True, random_state=1)
for train, test in kfold.split(x_train):
# διαχωρισμός train-test indexes
xi_train, xi_test = x_train[train], x_train[test]
yi_train, yi_test = y_train[train], y_train[test]
print(f' fold # {fold}, TRAIN: {train}, TEST: {test}')
# fit μοντέλου
mse_history = model_mse.fit(xi_train,
yi_train,
epochs=ep,
batch_size=200,
verbose=1,
validation_data=(xi_test, yi_test))
# αποθήκευση validation metrics για τα plots
aval.append(mse_history.history['val_accuracy'])
lval.append(mse_history.history['val_loss'])
ltrain.append(mse_history.history['loss'])
# μετρήσεις μοντέλου
mse_results = model_mse.evaluate(x_test, y_test, verbose=1)
print(
f'Αποτελέσματα στο fold # {fold} - Loss: {mse_results[0]} - Accuracy: {mse_results[1]}'
)
fold = fold + 1
# αποθήκευση για προβολή των αποτελεσμάτων 5-fold CV
loss_sum += mse_results[0]
acc_sum += mse_results[1]
# plots
# accuracy
plot_acc = plt.figure(1)
title1 = 'Validation Accuracy MSE Model {}-{}-10'.format(h1, h_2)
plt.title(title1, loc='center', pad=None)
plt.plot(np.mean(aval, axis=0))
plt.ylabel('acc')
plt.xlabel('epoch')
# loss
plot_loss = plt.figure(2)
title2 = 'Loss MSE Model {}-{}-10'.format(h1, h_2)
plt.title(title2, loc='center', pad=None)
# validation loss
plt.plot(np.mean(lval, axis=0))
# train loss
plt.plot(np.mean(ltrain, axis=0))
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['validation', 'train'], loc='upper left')
# Save locally
directories.filecheck('./plots/A2/Extra_Layer/{}.png'.format(title1))
directories.filecheck('./plots/A2/Extra_Layer/{}.png'.format(title2))
plot_loss.savefig("./plots/A2/Extra_Layer/{}.png".format(title2),
format='png')
plot_acc.savefig("./plots/A2/Extra_Layer/{}.png".format(title1),
format='png')
# εκτύπωση αποτελεσμάτων
print(
f'Results sum (MSE) - Loss {loss_sum / 5} - Accuracy {acc_sum / 5}'
)
f.close()
sys.stdout = sys.__stdout__
# καθαρισμός μνήμης
print(f'Clearing session....')
tensorflow.keras.backend.clear_session()
plt.close(1)
plt.close(2)
def a3(ep):
# αρχικοποίηση directories αποθήκευσης
directories.A3()
# GPU support
physical_devices = tensorflow.config.experimental.list_physical_devices(
'GPU')
print("CUDA - Αριθμός διαθέσιμων GPUs:",
len(tensorflow.config.experimental.list_physical_devices('GPU')))
tensorflow.config.experimental.set_memory_growth(physical_devices[0], True)
# αρχικοποίηση μεταβλητών
features = 784
classes = 10
loss_f = ['categorical_crossentropy', 'mean_squared_error']
h1 = 794
h2 = 50
learning_rates = [0.001, 0.001, 0.05, 0.1]
# φόρτωση mnist από το keras
(x_train, y_train), (x_test, y_test) = mnist.load_data()
# κάνουμε το mnist reshape
x_train = x_train.reshape(x_train.shape[0], features)
x_test = x_test.reshape(x_test.shape[0], features)
# MinMax scaling
scaler = preprocessing.MinMaxScaler()
x_train = scaler.fit_transform(x_train)
x_test = scaler.fit_transform(x_test)
# ορισμός των labels
y_train = to_categorical(y_train, classes)
y_test = to_categorical(y_test, classes)
# ορισμός input shape για το μοντέλο MLP βάσει των χαρακτηριστικών
input_shape = (features, )
for loss_fun in loss_f:
# Δημιουργία μοντέλου με χρήση του keras API
model = Sequential()
# Πρώτο κρυφό επίπεδο
model.add(Dense(h1, input_shape=input_shape, activation='relu'))
# Δεύτερο κρυφό επίπεδο
model.add(Dense(h2, activation='relu'))
# Επίπεδο εξόδου
model.add(Dense(classes, activation='softmax'))
i = 0
for lrate in learning_rates:
if i == 0:
m = 0.2
i = i + 1
else:
m = 0.6
i = i + 1
print(
f'Set SGD optimizer to learning rate={lrate} and momentum={m}')
opt = tensorflow.keras.optimizers.SGD(lr=lrate,
momentum=m,
decay=0.0,
nesterov=False)
fname = './logs/A3/results_{}{}_{}.txt'.format(loss_fun, lrate, m)
directories.filecheck(fname)
# compile
model.compile(loss=loss_fun, optimizer=opt, metrics=['accuracy'])
fold = 1
loss_sum = 0
acc_sum = 0
aval = []
lval = []
ltrain = []
f = open(fname, 'w')
sys.stdout = f
# 5-fold CV
kfold = KFold(5, shuffle=True, random_state=1)
for train, test in kfold.split(x_train):
# διαχωρισμός train-test indexes
xi_train, xi_test = x_train[train], x_train[test]
yi_train, yi_test = y_train[train], y_train[test]
print(f' fold # {fold}, TRAIN: {train}, TEST: {test}')
# fit μοντέλου
history = model.fit(
xi_train,
yi_train,
epochs=ep,
batch_size=200,
verbose=1,
validation_data=(xi_test, yi_test),
callbacks=[
tensorflow.keras.callbacks.EarlyStopping(
monitor='val_loss', patience=2)
])
# στατιστικά
aval.append(np.mean(history.history['val_accuracy']))
lval.append(np.mean(history.history['val_loss']))
ltrain.append(np.mean(history.history['loss']))
# μετρησεις μοντέλου
results = model.evaluate(x_test, y_test, verbose=1)
print(
f'Αποτελέσματα στο fold # {fold} - Loss: {results[0]} - Accuracy: {results[1]}'
)
fold += fold
# αποθήκευση για προβολή των αποτελεσμάτων 5-fold CV
loss_sum += results[0]
acc_sum += results[1]
# plots
# accuracy
plot_acc = plt.figure(1)
title1 = 'Validation Accuracy, {} model η={}, m={}'.format(
loss_fun, lrate, m)
plt.title(title1, loc='center', pad=None)
plt.plot(aval)
plt.ylabel('acc')
plt.xlabel('epoch')
# loss
plot_loss = plt.figure(2)
title2 = 'Loss, {} model η={}, m={}'.format(loss_fun, lrate, m)
plt.title(title2, loc='center', pad=None)
# validation loss
plt.plot(lval)
# train loss
plt.plot(ltrain)
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['validation', 'train'], loc='upper left')
# Save locally
directories.filecheck('./plots/A3/{}.png'.format(title1))
directories.filecheck('./plots/A3/{}.png'.format(title2))
plot_loss.savefig("./plots/A3/{}.png".format(title2), format='png')
plot_acc.savefig("./plots/A3/{}.png".format(title1), format='png')
# εκτύπωση αποτελεσμάτων
print(
f'Συνολικά Αποτελέσματα - Loss {loss_sum / 5} - Accuracy {acc_sum / 5}'
)
# επιστροφή stdout στην κονσόλα
f.close()
sys.stdout = sys.__stdout__
# απελευθερωση μνημης
print(f'Clearing session....')
tensorflow.keras.backend.clear_session()
plt.close(1)
plt.close(2)
aval.clear()
lval.clear()
ltrain.clear()
# μετρήσεις:
# fit = fitness scores για τη κάθε γενιά
# fittest = το καλύτερο σκορ στη γενιά
# best results gen = λίστα των καλύτερων σκορ κάθε γενιάς για κάθε τρέξιμο του αλγορίθμου (χρησιμοποιείται στο να ελέγχεται το ποσοστό βελτίωσης σε κάθε τρέξιμο)
# fitness_history = κάθε γραμμή πίνακα αποθηκεύει ανα γενιά την καλύτερη απόδοση του αλγορίθμου στο αντίστοιχο τρέξιμο
for num_indiv in no_of_individuals:
population = np.ones((num_indiv, 784))
for crossrate, mutrate in tweaks:
# άνοιγμα αρχείου αποθήκευσης
fname = "logs/B2/results_{}_{}_{}.txt".format(num_indiv, crossrate, mutrate)
directories.filecheck(fname)
f = open(fname, 'w')
sys.stdout = f
# αρχικοποιήσεις
solution = np.empty((iterations, 784))
solution_scores = np.empty((iterations, ))
fitness_history = np.zeros((10, num_gen))
gens_needed = []
average = 0
# GA
for iter in range(iterations):
best_results_gen = []
# ο αρχικός πληθυσμός παράγεται τυχαία
x_test = scaler.fit_transform(x_test) # ορισμός των labels y_train = to_categorical(y_train, classes) y_test = to_categorical(y_test, classes) # ορισμός input shape για το μοντέλο MLP βάσει των χαρακτηριστικών input_shape = (features, ) # αρχικοποίηση sum μετρήσεων loss_sum = 0 acc_sum = 0 # ορισμός directory αποθήκευσης stdout f_ce = "./logs/A2/Single_Layer/INPUT_results_CE_%s.txt" % h_1 f_mse = "./logs/A2/Single_Layer/INPUT_results_MSE_%s.txt" % h_1 directories.filecheck(f_ce) directories.filecheck(f_mse) # δημιουργία μοντέλων με χρήση του keras API model_ce = Sequential() model_mse = Sequential() # πρώτο κρυφό επίπεδο model_ce.add(Dense(h_1, input_shape=input_shape, activation='relu')) model_mse.add(Dense(h_1, input_shape=input_shape, activation='relu')) # επίπεδο εξόδου model_ce.add(Dense(classes, activation='softmax')) model_mse.add(Dense(classes, activation='softmax')) # compile # crossentropy model_ce.compile(loss='categorical_crossentropy',
