def run_train_EARs(exps_dir):
n_var = N_VAR
cols = VARS
for n in range(0, n_var):
# load json configure file
config_file = exps_dir + 'saved_models/ear/ear_' + cols[
n] + '_configs.json'
configs, configs_dict = get_config_from_json(config_file)
option_lags = configs.option_lags
best_params = configs_dict # need grid-search
args = {}
args['look_back'] = best_params['lag']
# make model
model = EAR(args).make_model()
print(model.summary())
# training
x, y = cons_ur_data(data_file=exps_dir + 'dataset/training.csv',
col=cols[n],
look_back=best_params['lag'])
model.fit(
x,
y,
batch_size=32,
epochs=Max_Epoch,
callbacks=[EarlyStopping(monitor='loss', patience=8, mode='min')],
validation_split=0,
verbose=2)
# save model
saved_models_file = exps_dir + 'saved_models/ear/ear_' + cols[
n] + '_weights'
model.save_weights(saved_models_file)
print('Model saved ... ', saved_models_file)
return
def run_visual_SEPNets(exps_dir, pretrain, trainable):
print('----pretrain: ', pretrain, '\t----trainable: ', trainable)
n_var = N_VAR
cols = VARS
p_list = []
ear_weights_files = []
for n in range(n_var):
# load json configure file
config_file = exps_dir + 'saved_models/ear/ear_' + cols[
n] + '_configs.json'
configs, configs_dict = get_config_from_json(config_file)
p_list.append(configs.lag)
ear_weights_files.append(exps_dir + 'saved_models/ear/ear_' + cols[n] +
'_weights')
args = {}
args['look_back'] = Max_Window
args['n_var'] = n_var
args['p_list'] = p_list
args['se_weights_files'] = ear_weights_files
args['pretrain'] = pretrain
args['trainable'] = trainable
# make model
model = SEPNets(args).make_model()
print(model.summary())
print(model.get_weights()[-2:])
return
def run_train_SEPNets(exps_dir, pretrain, trainable):
print('----pretrain: ', pretrain, '\t----trainable: ', trainable)
n_var = N_VAR
cols = VARS
p_list = []
ear_weights_files = []
for n in range(n_var):
# load json configure file
config_file = exps_dir + 'saved_models/ear/ear_' + cols[
n] + '_configs.json'
configs, configs_dict = get_config_from_json(config_file)
p_list.append(configs.lag)
ear_weights_files.append(exps_dir + 'saved_models/ear/ear_' + cols[n] +
'_weights')
args = {}
args['look_back'] = Max_Window
args['n_var'] = n_var
args['p_list'] = p_list
args['se_weights_files'] = ear_weights_files
args['pretrain'] = pretrain
args['trainable'] = trainable
# make model
model = SEPNets(args).make_model()
model.summary()
# training
x, y = cons_mv_data(data_file=exps_dir + 'dataset/training.csv',
cols=cols[:N_VAR],
look_back=Max_Window)
model.fit(
x,
y,
batch_size=32,
epochs=Max_Epoch,
callbacks=[EarlyStopping(monitor='loss', patience=3, mode='min')],
validation_split=0,
verbose=2)
# save model
saved_models_file = exps_dir + 'saved_models/sepnets/model_weights_' + \
str(pretrain) + '_' + str(trainable) + '_' + str(N_VAR)
model.save_weights(saved_models_file)
print('Model saved ... ', saved_models_file)
return
def run_visual_ARs(exps_dir):
n_var = N_VAR
cols = VARS
p_list = []
ar_weights_files = []
for n in range(n_var):
# load json configure file
config_file = exps_dir + 'saved_models/ar/ar_' + cols[
n] + '_configs.json'
configs, _ = get_config_from_json(config_file)
p_list.append(configs.lag)
ar_weights_files.append(exps_dir + 'saved_models/ar/ar_' + cols[n] +
'_weights')
# make model
ar_models = [AR({'look_back': p}).make_model() for p in p_list]
# load weights
[
ar_models[idx].load_weights(ar_weights_files[idx])
for idx in range(n_var)
]
# print
print(ar_models[0].summary())
[print(ar_models[idx].get_weights()[-1]) for idx in range(n_var)]
return
from models.lstnets import LSTNet_multi_inputs
from data_loader.generator import cons_mv_data
from sklearn.metrics import mean_absolute_error
from utils.tools import mean_mae
from bunch import Bunch
from keras.callbacks import EarlyStopping
EXP_DIRS = ['../../exp_ElectricityLoad/',
'../../exp_210100063/',
'../../exp_201812/',
'../../exp_210100112/']
EXP_DIR = EXP_DIRS[3]
exp_config, _exp_config = get_config_from_json(EXP_DIR + 'exp_config.json')
N_VAR = exp_config.N_VAR
VARS = exp_config.VARS
Max_Window = exp_config.Max_Window
Max_Epoch = exp_config.Max_Epoch
Skip = exp_config.Skip
Period = exp_config.Period
MODE_LIST = ['train', 'test', 'visual']
MODE = MODE_LIST[1]
ARGS = {
"window": Max_Window,
'hidRNN': 16,
'hidCNN': 3,
def run_test_ARs(exps_dir):
n_var = N_VAR
cols = VARS
p_list = []
ar_weights_files = []
for n in range(n_var):
# load json configure file
config_file = exps_dir + 'saved_models/ar/ar_' + cols[
n] + '_configs.json'
configs, _ = get_config_from_json(config_file)
p_list.append(configs.lag)
ar_weights_files.append(exps_dir + 'saved_models/ar/ar_' + cols[n] +
'_weights')
# make model
ar_models = [AR({'look_back': p}).make_model() for p in p_list]
# load weights
[
ar_models[idx].load_weights(ar_weights_files[idx])
for idx in range(n_var)
]
# make forecast
test_x, test_y = cons_mv_data(data_file=exps_dir + 'dataset/testing.csv',
cols=cols,
look_back=Max_Window)
test_y_ar_pred = [
ar_models[idx].predict(test_x[:, -p_list[idx]:, idx])
for idx in range(n_var)
] # make auto-regression prediction with shape of (n_var, batch_size)
test_y_ar_pred = np.concatenate(test_y_ar_pred,
axis=-1) # (batch_size, n_var)
test_y_point_pred = test_x[:, -1, :]
# make forecast for training
train_x, train_y = cons_mv_data(data_file=exps_dir +
'dataset/training.csv',
cols=cols,
look_back=Max_Window)
train_y_ar_pred = [
ar_models[idx].predict(train_x[:, -p_list[idx]:, idx])
for idx in range(n_var)
]
train_y_ar_pred = np.concatenate(train_y_ar_pred, axis=-1)
train_y_point_pred = train_x[:, -1, :]
# mean mae
print('mean-overall-mae:\t', mean_mae(test_y).mean())
print('mean-mae:\n', mean_mae(test_y))
# pre-point mae
print('point-overall-mae:\t', mean_absolute_error(test_y,
test_y_point_pred))
print(
'point-mae:\n',
mean_absolute_error(test_y,
test_y_point_pred,
multioutput='raw_values'))
# ar model mae
print('ar_model-overall-mae:\t',
mean_absolute_error(test_y, test_y_ar_pred))
print(
'ar_model-mae:\n',
mean_absolute_error(test_y, test_y_ar_pred, multioutput='raw_values'))
np.savez_compressed(exps_dir + 'results/y_ar_pred',
train_y_pred=train_y_ar_pred,
test_y_pred=test_y_ar_pred)
return
def run_test_SEPNets(exps_dir, pretrain, trainable):
print('----pretrain: ', pretrain, '\t----trainable: ', trainable)
n_var = N_VAR
cols = VARS
p_list = []
ear_weights_files = []
for n in range(n_var):
# load json configure file
config_file = exps_dir + 'saved_models/ear/ear_' + cols[
n] + '_configs.json'
configs, _ = get_config_from_json(config_file)
p_list.append(configs.lag)
ear_weights_files.append(exps_dir + 'saved_models/ear/ear_' + cols[n] +
'_weights')
# make model
args = {}
args['look_back'] = Max_Window
args['n_var'] = n_var
args['p_list'] = p_list
args['se_weights_files'] = ear_weights_files
args['pretrain'] = pretrain
args['trainable'] = trainable
# make model
model = SEPNets(args).make_model()
print(model.summary())
# load weights
saved_models_file = exps_dir + 'saved_models/sepnets/model_weights_' + \
str(pretrain) + '_' + str(trainable) + '_' + str(N_VAR)
model.load_weights(saved_models_file)
# make testing forecast
test_x, test_y = cons_mv_data(data_file=exps_dir + 'dataset/testing.csv',
cols=cols[:N_VAR],
look_back=Max_Window)
test_y_sepnets_pred = model.predict(test_x)
test_y_point_pred = test_x[:, -1, :]
# make training forecast
train_x, train_y = cons_mv_data(data_file=exps_dir +
'dataset/training.csv',
cols=cols[:N_VAR],
look_back=Max_Window)
train_y_sepnets_pred = model.predict(train_x)
# mean mae
print('mean-overall-mae:\t', mean_mae(test_y).mean())
print('mean-mae:\n', mean_mae(test_y))
# pre-point mae
print('point-overall-mae:\t', mean_absolute_error(test_y,
test_y_point_pred))
print(
'point-mae:\n',
mean_absolute_error(test_y,
test_y_point_pred,
multioutput='raw_values'))
# sepnets model mae
print('sepnets_model-overall-mae:\t',
mean_absolute_error(test_y, test_y_sepnets_pred))
print(
'sepnets_model-mae:\n',
mean_absolute_error(test_y,
test_y_sepnets_pred,
multioutput='raw_values'))
saved_results = exps_dir+'results/y_sepnets_pred_'+ \
str(pretrain) + '_' + str(trainable)
np.savez_compressed(saved_results,
train_y_pred=train_y_sepnets_pred,
test_y_pred=test_y_sepnets_pred)
return