def quick_load_bj_taxi_external(timeslots, description, length):
"""
load external data quickly, if time slots is the same
:param timeslots: the demanded time
:param description: global or local
:return:
"""
start_time = str(timeslots[0][0])
end_time = str(timeslots[-1][0])
path = r'/home/ryj/renyajie/exp/GLST_Net/inter_data/data'
filename = 'bj_taxi_external_{}_{}_{}_{}.h5'.format(
description, length, start_time, end_time)
f = h5py.File(os.path.join(path, filename), 'a')
# encode the time
encode_time = np.asarray([utils.encode_time(batch) for batch in timeslots])
# if the same, load directly
if 'timeslots' in f and (encode_time == f['timeslots'][()]).all():
print('cache load bj taxi {} external data from {} to {}'.format(
description, start_time, end_time))
print('-' * 30)
vacation = f['vacation'][()]
dayOfWeek = f['dayOfWeek'][()]
weather = f['weather'][()]
continuous_external = f['continuous_external'][()]
hour = f['hour'][()]
f.close()
return vacation, hour, dayOfWeek, weather, continuous_external
f.close()
# else calculate then cache
f = h5py.File(os.path.join(path, filename), 'w')
print('calculate bj taxi {} external data from {} to {}'.format(
description, start_time, end_time))
if description == 'global':
vacation, hour, dayOfWeek, weather, continuous_external = \
load_bj_taxi_external(timeslots, 'bj taxi global external data')
else:
vacation, hour, dayOfWeek, weather, continuous_external = \
load_bj_taxi_external(timeslots, 'bj taxi local external data')
print('cache bj taxi {} external data from {} to {}'.format(
description, start_time, end_time))
f['timeslots'] = encode_time
f['vacation'] = vacation
f['hour'] = hour
f['dayOfWeek'] = dayOfWeek
f['weather'] = weather
f['continuous_external'] = continuous_external
f.close()
return vacation, hour, dayOfWeek, weather, continuous_external
def quick_load_ny_bike_external(timeslots, description, length):
"""
load external data quickly, if time slots is the same
:param timeslots: the demanded time
:param description: global or local
:return:
"""
start_time = str(timeslots[0][0])
end_time = str(timeslots[-1][0])
path = r'/home/ryj/renyajie/exp/GLST_Net/inter_data/data'
filename = 'ny_bike_external_{}_{}_{}_{}.h5'.format(
description, length, start_time, end_time)
f = h5py.File(os.path.join(path, filename), 'a')
# encode the time
encode_time = np.asarray([utils.encode_time(batch) for batch in timeslots])
# if the same, load directly
if 'timeslots' in f and (encode_time == f['timeslots'][()]).all():
print('cache load ny bike {} external data from {} to {}'.format(
description, start_time, end_time))
print('-' * 30)
dayOfWeek = f['dayOfWeek'][()]
hour = f['hour'][()]
f.close()
return hour, dayOfWeek
f.close()
# else calculate then cache
f = h5py.File(os.path.join(path, filename), 'w')
print('calculate ny bike {} external data from {} to {}'.format(
description, start_time, end_time))
if description == 'global':
hour, dayOfWeek = load_ny_bike_external(
timeslots, 'ny bike global external data')
else:
hour, dayOfWeek = load_ny_bike_external(timeslots,
'ny bike local external data')
print('cache ny bike {} external data from {} to {}'.format(
description, start_time, end_time))
f['timeslots'] = encode_time
f['hour'] = hour
f['dayOfWeek'] = dayOfWeek
f.close()
return hour, dayOfWeek
def quick_get_data(dataset, predict_time, index_cut, neighbor_size, lstm_seq_len, flow_gate_len, att_lstm_num, att_lstm_seq_len):
start_time = str(predict_time[0])
end_time = str(predict_time[-1])
path = r'/home/ryj/renyajie/exp/GLST_Net/inter_data/data'
filename = '{}_baseline_stdn_neighbor_{}_{}_{}_{}.h5'.format(dataset, neighbor_size, lstm_seq_len, flow_gate_len, att_lstm_num, att_lstm_seq_len)
f = h5py.File(os.path.join(path, filename), 'a')
# encode the time
encode_time = np.asarray([utils.encode_time(batch) for batch in predict_time])
# if the same, load directly
if 'predict_time' in f and (encode_time == f['predict_time'][()]).all():
print('cache load stdn {} data from {} to {}, neighbor is {}'.format(dataset, start_time, end_time, neighbor_size))
print('-' * 30)
train_dict = {key + '_train' : f[key + '_train'][()] for key in data_name}
test_dict = {key + '_test' : f[key + '_test'][()] for key in data_name}
external_dim = f['external_dim'][()]
if(train_dict['lstm_input_train'].shape[1] == lstm_seq_len
and train_dict['flow_input_train'].shape[3] == neighbor_size * 2 + 1
and train_dict['att_nbhd_input_train'].shape[1] == att_lstm_num * att_lstm_seq_len
and train_dict['att_lstm_input_train'].shape[1] == att_lstm_num):
f.close()
return train_dict, test_dict, external_dim
f.close()
# else calculate then cache
f = h5py.File(os.path.join(path, filename), 'w')
print('calculate stdn {} data from {} to {}, neighbor is {}'.format(dataset, start_time, end_time, neighbor_size))
train_dict, test_dict, external_dim = \
get_data(dataset, predict_time, index_cut, neighbor_size, lstm_seq_len, flow_gate_len, att_lstm_num, att_lstm_seq_len)
print('cache stdn {} data from {} to {}, neighbor is {}'.format(dataset, start_time, end_time, neighbor_size))
f['predict_time'] = encode_time
for data in [test_dict, train_dict]:
for key, value in data.items():
f[key] = value
f['external_dim'] = external_dim
f.close()
return train_dict, test_dict, external_dim
def quick_get_data(dataset, predict_time, index_cut, neighbor_size, len_close,
len_period, len_trend):
"""
load st-res data quickly if there is the file
:param predict_time:
:param index_cut:
:param neighbor_size:
:param len_close:
:param len_period:
:param len_trend:
:return:
"""
start_time = str(predict_time[0])
end_time = str(predict_time[-1])
path = r'/home/ryj/renyajie/exp/GLST_Net/inter_data/data'
filename = '{}_baseline_stres_neighbor_{}_{}_{}_{}.h5'.format(
dataset, neighbor_size, len_close, len_period, len_trend)
f = h5py.File(os.path.join(path, filename), 'a')
# encode the time
encode_time = np.asarray(
[utils.encode_time(batch) for batch in predict_time])
# if the same, load directly
if 'predict_time' in f and (encode_time == f['predict_time'][()]).all():
print('cache load st-res {} data from {} to {}, neighbor is {}'.format(
dataset, start_time, end_time, neighbor_size))
print('-' * 30)
X_train = f['X_train'][()]
X_test = f['X_test'][()]
Y_train = f['Y_train'][()]
Y_test = f['Y_test'][()]
T_train = f['T_train'][()]
T_test = f['T_test'][()]
External_train = f['External_train'][()]
External_test = f['External_test'][()]
external_dim = f['external_dim'][()]
f.close()
return X_train, X_test, Y_train, Y_test, T_train, T_test, External_train, External_test, external_dim
f.close()
# else calculate then cache
f = h5py.File(os.path.join(path, filename), 'w')
print('calculate st-res {} data from {} to {}, neighbor is {}'.format(
dataset, start_time, end_time, neighbor_size))
X_train, X_test, Y_train, Y_test, T_train, T_test, External_train, External_test, external_dim = \
get_data(dataset, predict_time, index_cut, neighbor_size, len_close, len_period, len_trend)
encode_T_train = np.asarray(
[utils.encode_time(batch) for batch in T_train])
encode_T_test = np.asarray([utils.encode_time(batch) for batch in T_test])
print('cache st-res {} data from {} to {}, neighbor is {}'.format(
dataset, start_time, end_time, neighbor_size))
f['predict_time'] = encode_time
f['X_train'] = X_train
f['X_test'] = X_test
f['Y_train'] = Y_train
f['Y_test'] = Y_test
f['T_train'] = encode_T_train
f['T_test'] = encode_T_test
f['External_train'] = External_train
f['External_test'] = External_test
f['external_dim'] = external_dim
f.close()
return X_train, X_test, Y_train, Y_test, T_train, T_test, External_train, External_test, external_dim
def quick_get_data(dataset, predict_time, hours, days, weeks):
"""
load astgcn data quickly if there is the file
:param predict_time:
:param index_cut:
:param neighbor_size:
:param hours:
:param days:
:param weeks:
:param predict_len: 1 default
:return:
"""
start_time = str(predict_time[0])
end_time = str(predict_time[-1])
path = r'/home/ryj/renyajie/exp/GLST_Net/inter_data/data'
filename = '{}_baseline_astgcn_{}_{}_{}.h5'.format(dataset, hours, days,
weeks)
f = h5py.File(os.path.join(path, filename), 'a')
# encode the time
encode_time = np.asarray(
[utils.encode_time(batch) for batch in predict_time])
# if the same, load directly
if 'predict_time' in f and (encode_time == f['predict_time'][()]).all():
print('cache load astgcn {} data from {} to {}'.format(
dataset, start_time, end_time))
print('-' * 30)
X_train = f['X_train'][()]
X_test = f['X_test'][()]
Y_train = f['Y_train'][()]
Y_test = f['Y_test'][()]
T_train = f['T_train'][()]
T_test = f['T_test'][()]
f.close()
return X_train, X_test, Y_train, Y_test, T_train, T_test
f.close()
# else calculate then cache
f = h5py.File(os.path.join(path, filename), 'w')
print('calculate astgcn {} data from {} to {}'.format(
dataset, start_time, end_time))
X_train, X_test, Y_train, Y_test, T_train, T_test = \
get_data(dataset, predict_time, hours, days, weeks)
encode_T_train = np.asarray(
[utils.encode_time(batch) for batch in T_train])
encode_T_test = np.asarray([utils.encode_time(batch) for batch in T_test])
print('cache astgcn {} data from {} to {}'.format(dataset, start_time,
end_time))
f['predict_time'] = encode_time
f['X_train'] = X_train
f['X_test'] = X_test
f['Y_train'] = Y_train
f['Y_test'] = Y_test
f['T_train'] = encode_T_train
f['T_test'] = encode_T_test
f.close()
return X_train, X_test, Y_train, Y_test, T_train, T_test
def load_ny_bike(proportion_test,
len_global=7,
len_local=4,
neighbor_size=3,
region_num=5):
"""
load all the data
args:
len_global: the time length of global data
len_local: the time length of local data
neighbor_size: the local size, size = (val * 2 + 1) * (val * 2 + 1)
region_num: how many regions that a map contains
ret:
train set and test set, including:
1. global external and flow data
2. local external and flow data
3. ground truth
"""
date, data, mmn, index = load_ny_bike_flow()
# get global and local flow data, ground truth and the corresponding date
global_flow, stack_local_flow, ground_truth, current_local_flow, index_cut, \
predict_time, global_timeslots, local_timeslots = \
utils.get_flow_data(date, data, len_global, len_local, neighbor_size, region_num,
unit_len=24, width=16, height=8)
# get global and local external data
g_hour, g_dayOfWeek = quick_load_ny_bike_external(global_timeslots,
'global', len_global)
t_hour, t_dayOfWeek = quick_load_ny_bike_external(local_timeslots, 'local',
len_local)
# change encode to ascii for time
predict_time = np.asarray(utils.encode_time(predict_time))
global_timeslots = np.asarray(
[utils.encode_time(batch) for batch in global_timeslots])
local_timeslots = np.asarray(
[utils.encode_time(batch) for batch in local_timeslots])
# build train set and test set according to the param:len_test
data_dict = {
'global_flow': global_flow,
'stack_local_flow': stack_local_flow,
'ground_truth': ground_truth,
'current_local_flow': current_local_flow,
'index_cut': index_cut,
'predict_time': predict_time,
'global_timeslots': global_timeslots,
'local_timeslots': local_timeslots,
'g_hour': g_hour,
'g_dayOfWeek': g_dayOfWeek,
't_hour': t_hour,
't_dayOfWeek': t_dayOfWeek
}
data_dict = utils.duplicate_data(data_dict, region_num)
total_length = g_dayOfWeek.shape[0]
len_test = math.ceil(total_length * proportion_test)
len_train = total_length - len_test
print('train set length {:d}\ntest set length {:d}\n{}'.format(
len_train, len_test, '-' * 30))
train_set, test_set, data_name = utils.divide_train_and_test(
len_test, data_dict)
print('predict start: {}\npredict end: {}'.format(
data_dict["predict_time"][0].decode('utf-8'),
data_dict["predict_time"][-1].decode('utf-8')))
print('-' * 30)
return train_set, test_set, mmn, data_name
def load_rdw_ny_bike(proportion_test,
len_recent=4,
len_daily=4,
len_week=4,
neighbor_size=2,
region_num=5):
"""
load all the data
args:
len_global: the time length of global data
len_local: the time length of local data
neighbor_size: the local size, size = (val * 2 + 1) * (val * 2 + 1)
region_num: how many regions that a map contains
ret:
train set and test set, including:
1. global external and flow data
2. local external and flow data
3. ground truth
"""
date, data, mmn, index = load_ny_bike_flow()
# get global and local flow data, ground truth and the corresponding date
recent_local_flow, daily_local_flow, week_local_flow, ground_truth, current_local_flow \
, index_cut, predict_time, recent_time, daily_time, week_time, current_time = \
utils.get_flow_rdw_data(date, data, len_recent, len_daily, len_week,
neighbor_size, region_num, unit_len=24, width=16, height=8)
# get recent, daily, week, current external data
recent_hour, recent_dayOfWeek = quick_load_ny_bike_external(
recent_time, 'recent', len_recent)
daily_hour, daily_dayOfWeek = quick_load_ny_bike_external(
daily_time, 'daily', len_daily)
week_hour, week_dayOfWeek = quick_load_ny_bike_external(
week_time, 'week', len_week)
current_hour, current_dayOfWeek = quick_load_ny_bike_external(
current_time, 'current', 1)
# change encode to ascii for time
predict_time = np.asarray(utils.encode_time(predict_time))
recent_time = np.asarray(
[utils.encode_time(batch) for batch in recent_time])
daily_time = np.asarray([utils.encode_time(batch) for batch in daily_time])
week_time = np.asarray([utils.encode_time(batch) for batch in week_time])
current_time = np.asarray(
[utils.encode_time(batch) for batch in current_time])
# build train set and test set according to the param:len_test
data_dict = {
'recent_local_flow': recent_local_flow,
'daily_local_flow': daily_local_flow,
'week_local_flow': week_local_flow,
'current_local_flow': current_local_flow,
'ground_truth': ground_truth,
'index_cut': index_cut,
'predict_time': predict_time,
'recent_time': recent_time,
'daily_time': daily_time,
'week_time': week_time,
'current_time': current_time,
'recent_hour': recent_hour,
'recent_dayOfWeek': recent_dayOfWeek,
'daily_hour': daily_hour,
'daily_dayOfWeek': daily_dayOfWeek,
'week_hour': week_hour,
'week_dayOfWeek': week_dayOfWeek,
'current_hour': current_hour,
'current_dayOfWeek': current_dayOfWeek
}
data_dict = utils.duplicate_rdw_data(data_dict, region_num)
total_length = current_time.shape[0]
len_test = math.ceil(total_length * proportion_test)
len_train = total_length - len_test
print('train set length {:d}\ntest set length {:d}\n{}'.format(
len_train, len_test, '-' * 30))
train_set, test_set, data_name = utils.divide_train_and_test(
len_test, data_dict)
print('predict start: {}\npredict end: {}'.format(
data_dict["predict_time"][0].decode('utf-8'),
data_dict["predict_time"][-1].decode('utf-8')))
print('-' * 30)
return train_set, test_set, mmn, data_name
