def create_custom_noisy_set1(x_train_i, y_train_i):
ratio1 = {'mixed_transformations': 1}
ratio2 = {'random_image_eraser': 1}
ratio3 = {
"add_gaussian_noise": 0.2,
"add_sp_noise": 0.4,
"add_poisson_noise": .2,
'add_multiplicative_noise': .2
}
x_train_r1, y_train_r1 = get_random_shuffle_set(x_train_i, y_train_i,
ratio1)
x_train_r2, y_train_r2 = get_random_shuffle_set(x_train_i, y_train_i,
ratio2)
x_train_r3, y_train_r3 = get_random_shuffle_set(x_train_i, y_train_i,
ratio3)
final_x_train = np.concatenate((x_train_r1, x_train_r2, x_train_r3),
axis=0)
final_y_train = np.concatenate((y_train_r1, y_train_r2, y_train_r3),
axis=0)
x_train_shuffled, y_train_shuffled = get_random_data(final_x_train, final_y_train, 0, \
final_x_train.shape[0], max_samples=final_x_train.shape[0])
return x_train_shuffled, y_train_shuffled
def create_custom_noisy_set2(x_train_i, y_train_i):
ratio1 = {
'perform_swirl_transformation': .25,
'perform_random_affine_transform': .55
}
ratio2 = {'flip_rotate': .45, 'rotate_np': .55}
x_train_r1, y_train_r1 = get_random_shuffle_set(x_train_i, y_train_i,
ratio1)
x_train_r2, y_train_r2 = get_random_shuffle_set(x_train_i, y_train_i,
ratio2)
final_x_train = np.concatenate((x_train_r1, x_train_r2), axis=0)
final_y_train = np.concatenate((y_train_r1, y_train_r2), axis=0)
x_train_shuffled, y_train_shuffled = get_random_data(final_x_train, final_y_train, 0, \
final_x_train.shape[0], max_samples=final_x_train.shape[0])
return x_train_shuffled, y_train_shuffled
if __name__ == '__main__':
MAX_TRAINING = 40000
MAX_VALIDATION = 5000
MAX_TESTING = 3000
base = 'noisy_dataset/noisy/cifar_10_{}.npy'
x_train_max, y_train = np.load(base.format('x_train')), np.load(
base.format('y_train'))
x_test_max, y_test = np.load(base.format('x_test')), np.load(
base.format('y_test'))
y_train_max, y_test_max = to_categorical(
y_train, num_classes=10), to_categorical(y_test, num_classes=10)
train_cutoff = int(x_train_max.shape[0] - MAX_VALIDATION * 4)
x_train, y_train = get_random_data(x_train_max, y_train_max, 0,
train_cutoff, MAX_TRAINING)
x_val, y_val = get_random_data(x_train_max, y_train_max, train_cutoff + 1,
x_train_max.shape[0], MAX_VALIDATION)
x_test, y_test = get_random_data(x_test_max, y_test_max, 0,
x_test_max.shape[0], MAX_TESTING)
print("Number of Training Samples: X={}, Y={}".format(
x_train.shape[0], y_train.shape[0]))
print("Number of Validation Samples: X={}, Y={}".format(
x_val.shape[0], y_val.shape[0]))
print("Number of Test Samples: X={}, Y={}".format(x_test.shape[0],
y_test.shape[0]))
testing_optimizers = ['RAdam', 'Adam', 'NAdam', 'SGD']
testing_activations = ['tanh', 'relu', 'selu']
cifar = tf.keras.datasets.cifar10
(x_train, y_train), (x_test, y_test) = cifar.load_data()
x_train_sub, y_train_sub = fetch_selected_data_10(x_train, y_train)
x_test_sub, y_test_sub = fetch_selected_data_10(x_test, y_test)
cifar100 = tf.keras.datasets.cifar100
(x_train_o, y_train_o), (x_test_o, y_test_o) = cifar100.load_data()
x_train_o1, y_train_o1 = fetch_selected_data_100(x_train_o, y_train_o)
x_test_o1, y_test_o1 = fetch_selected_data_100(x_test_o, y_test_o)
x_train_np, x_test_sub = np.concatenate(
(x_train_sub, x_train_o1, x_test_o1),
axis=0) / 255.0, x_test_sub / 255.0
y_train_np = np.concatenate((y_train_sub, y_train_o1, y_test_o1), axis=0)
x_train_np_sub, y_train_np_sub = get_random_data(x_train_np, y_train_np, 0,
x_train_np.shape[0],
10000)
x_train_noisy1, y_train_noisy1 = create_custom_noisy_set1(
x_train_np_sub, y_train_np_sub)
x_train_noisy2, y_train_noisy2 = create_custom_noisy_set2(
x_train_np_sub, y_train_np_sub)
x_train_noisy = np.concatenate((x_train_noisy1, x_train_noisy2), axis=0)
y_train_noisy = np.concatenate((y_train_noisy1, y_train_noisy2), axis=0)
x_test_noisy, y_test_noisy = create_custom_noisy_set1(
x_test_sub, y_test_sub)
#x_test_less_noisy, y_test_less_noisy = get_random_shuffle_set(x_test_noisy, y_test_noisy, {'median_filtering':1}, True)
np.save('noisy_dataset/noisy/cifar_10_x_train.npy', x_train_noisy)
np.save('noisy_dataset/noisy/cifar_10_y_train.npy', y_train_noisy)
np.save('noisy_dataset/noisy/cifar_10_x_test.npy', x_test_noisy)