def augment_train_set(x_train, y_train, classes, N, dba_iters=5,
weights_method_name = 'as', distance_algorithm='DTW',
limit_N = True):
"""
This method takes a dataset and augments it using the method in icdm2017.
:param x_train: The original train set
:param y_train: The original labels set
:param N: The number of synthetic time series.
:param dba_iters: The number of dba iterations to converge.
:param weights_method_name: The method for assigning weights (see constants.py)
:param distance_algorithm: The name of the distance algorithm used (see constants.py)
"""
# get the weights function
#weights_fun = utils.constants.WEIGHTS_METHODS[weights_method_name]
weights_fun = get_weights_average_selected
# synthetic train set and labels
synthetic_x_train = []
synthetic_y_train = []
# loop through each class
k = 0
for c in classes:
k += 1
print('class {} of {}'.format(k, len(classes)))
# get the MTS for this class
c_x_train = x_train[np.where(y_train==c)]
if len(c_x_train) == 1 :
# skip if there is only one time series per set
continue
if limit_N == True:
# limit the nb_prototypes
nb_prototypes_per_class = min(N, len(c_x_train))
else:
# number of added prototypes will re-balance classes
nb_prototypes_per_class = N + (N - len(c_x_train))
# get the pairwise matrix
if weights_method_name == 'aa':
# then no need for dist_matrix
dist_pair_mat = None
else:
dist_pair_mat = calculate_dist_matrix(c_x_train, dist_fun, dist_fun_params)
t = 0
# loop through the number of synthtectic examples needed
for n in range(nb_prototypes_per_class):
t += 1
print('nb_prototypes_per_class {} of {}'.format(t, nb_prototypes_per_class))
# get the weights and the init for avg method
weights, init_avg = weights_fun(c_x_train, dist_pair_mat, distance_algorithm=distance_algorithm)
# get the synthetic data
synthetic_mts = dba(c_x_train, dba_iters, weights=weights)
# add the synthetic data to the synthetic train set
synthetic_x_train.append(synthetic_mts)
# add the corresponding label
synthetic_y_train.append(c)
# return the synthetic set
return np.array(synthetic_x_train), np.array(synthetic_y_train)
# get the pairwise matrix
if weights_method_name == 'aa':
# then no need for dist_matrix
dist_pair_mat = None
else:
dist_pair_mat = calculate_dist_matrix(c_x_train, dist_fun, dist_fun_params)
t = 0
# loop through the number of synthtectic examples needed
for n in range(nb_prototypes_per_class):
t += 1
print('nb_prototypes_per_class {} of {}'.format(t, nb_prototypes_per_class))
# get the weights and the init for avg method
weights, init_avg = weights_fun(c_x_train, dist_pair_mat, distance_algorithm=distance_algorithm)
# get the synthetic data
synthetic_mts = dba(c_x_train, dba_iters, weights=weights)
# add the synthetic data to the synthetic train set
synthetic_x_train.append(synthetic_mts)
# add the corresponding label
synthetic_y_train.append(c)
# return the synthetic set
return np.array(synthetic_x_train), np.array(synthetic_y_train)
if __name__ == '__main__':
tseries = utils.tseries
dist_pair_mat = calculate_dist_matrix(tseries, dist_fun, dist_fun_params)
weights, init_dba = get_weights_average_selected(tseries, dist_pair_mat)
print(weights)
print(init_dba)
def augment_train_set(x_train,
y_train,
N,
dba_iters=1,
weights_method_name='aa',
distance_algorithm='dtw',
limit_N=True):
"""
This method takes a dataset and augments it using the method in icdm2017.
:param x_train: The original train set
:param y_train: The original labels set
:param N: The number of synthetic time series.
:param dba_iters: The number of dba iterations to converge.
:param weights_method_name: The method for assigning weights (see constants.py)
:param distance_algorithm: The name of the distance algorithm used (see constants.py)
"""
# get the weights function
weights_fun = utils.constants.WEIGHTS_METHODS[weights_method_name]
# get the distance function
dist_fun = utils.constants.DISTANCE_ALGORITHMS[distance_algorithm]
# get the distance function params
dist_fun_params = utils.constants.DISTANCE_ALGORITHMS_PARAMS[
distance_algorithm]
# synthetic train set and labels
synthetic_x_train = []
synthetic_y_train = []
if limit_N == True:
# limit the nb_prototypes
nb_prototypes_per_class = min(N, len(x_train))
else:
# number of added prototypes will re-balance classes
nb_prototypes_per_class = N + (N - len(x_train))
# get the pairwise matrix
# loop through the number of synthtectic examples needed
for n in range(nb_prototypes_per_class):
# get the weights and the init for avg method
indices = np.random.randint(x_train.shape[0], size=20)
x_train_random = x_train[indices, :]
y_train_random = y_train[indices, :]
if weights_method_name == 'aa':
# then no need for dist_matrix
dist_pair_mat = None
else:
dist_pair_mat = calculate_dist_matrix(x_train_random, dist_fun,
dist_fun_params)
weights, init_avg = weights_fun(
x_train_random,
dist_pair_mat,
distance_algorithm=distance_algorithm)
# get the synthetic data
synthetic_mts_x = dba(x_train_random,
dba_iters,
verbose=False,
distance_algorithm=distance_algorithm,
weights=weights,
init_avg_method='manual',
init_avg_series=init_avg)
# add the synthetic data to the synthetic train set
synthetic_x_train.append(synthetic_mts_x)
new_y = 0
sum_weights = 0
weights_y = np.mean(weights, axis=1)
for s in range(y_train_random.shape[0]):
new_y += y_train_random[s] * weights_y[s]
sum_weights += weights_y[s]
# update the new weighted y
new_y = new_y / sum_weights
synthetic_y_train.append(new_y)
print('File', ind, 'new_data_number', n)
# return the synthetic set
return np.array(synthetic_x_train), np.array(synthetic_y_train)
# In[90]:
n = 100
syn_train_set = augment_train_set(X_p,
y,
n,
weights_method_name='as',
distance_algorithm='dtw')
# get the synthetic train and labels
syn_x_train, syn_y_train = syn_train_set
# concat the synthetic with the reduced random train and labels
aug_x_train = np.array(X_p.tolist() + syn_x_train.tolist())
aug_y_train = np.array(y.tolist() + syn_y_train.tolist())
# In[ ]:
aug_data = {'aug_x': aug_x_train, 'aug_y': aug_y_train}
aug_res_address = res_addr + 'session_' + ind + '/aug_power_data_session_' + ind + '.mat'
io.savemat(aug_res_address, aug_data)
original_data = {'x': X_p, 'y': y}
res_address = res_addr + 'session_' + ind + '/power_data_session_' + ind + '.mat'
io.savemat(res_address, original_data)
print(time.clock() - start_time)