def filt(X_data, type_freq, lim_freq_1):
if type_freq == "low":
X_data_f = aug_freq_low(X_data, lim_freq_1, 3, "float")
if type_freq == "high":
X_data_f = aug_freq_high(X_data, lim_freq_1, 3, "float")
if type_freq == "none":
X_data_f = X_data
return (X_data_f)
print("Accuracy on train set: " + str(model.evaluate(X_train, Y_train, verbose=0)[1]))
print("Accuracy on test set: " + str(model.evaluate(X_test, Y_test, verbose=0)[1]))
model.save("models/SVHN_low_freq_" + str(lim_freq) + ".h5")
if (model_type == "high_freq"):
lim_freq = int(sys.argv[3])
print("limit frequency: " + str(lim_freq))
from data_augmentation import aug_freq_high
X_train = aug_freq_high(X_train, lim_freq, 3, type_r="float")
X_test = aug_freq_high(X_test, lim_freq, 3, type_r="float")
generator=ImageDataGenerator()
model = svhn_model()
filepath="svhn_weights_best_" + model_type + "_"+ str(lim_freq) +".hdf5"
checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')
def lr_schedule(epoch):
lr = 1e-2
if epoch > 15:
lr = 0.001
print('Learning rate: ', lr)
return lr
lr_scheduler = LearningRateScheduler(lr_schedule)
callbacks_list = [checkpoint, lr_scheduler]