def train_sckit_dirregression(architecture, config, runconfig):
"""
Training process for architecture with direct regression of ahead time steps
:return:
"""
if type(config['data']['ahead']) == list:
iahead, sahead = config['data']['ahead']
else:
iahead, sahead = 1, config['data']['ahead']
lresults = []
for ahead in range(iahead, sahead + 1):
if runconfig.verbose:
print(
'************************************************************')
print(f'Steps Ahead = {ahead} ')
# Dataset
dataset = Dataset(config=config['data'], data_path=wind_data_path)
dataset.generate_dataset(ahead=ahead,
mode=architecture.data_mode,
remote=runconfig.remote)
train_x, train_y, val_x, val_y, test_x, test_y = dataset.get_data_matrices(
)
############################################
# Model
arch = architecture(config, runconfig)
arch.generate_model()
if runconfig.verbose:
arch.summary()
dataset.summary()
print()
############################################
# Training
arch.train(train_x, train_y, val_x, val_y)
############################################
# Results
if 'descale' not in config['training'] or config['training']['descale']:
lresults.append([ahead] + arch.evaluate(
val_x, val_y, test_x, test_y, scaler=dataset.scaler))
else:
lresults.append([ahead] +
arch.evaluate(val_x, val_y, test_x, test_y))
print(strftime('%Y-%m-%d %H:%M:%S'))
if config is not None:
updateprocess(config, ahead)
arch.log_result(lresults)
return lresults
def train_sckit_sequence2sequence(architecture, config, runconfig):
"""
Training process for architecture with direct regression of ahead time steps
:return:
"""
ahead = config['data']['ahead'] if (type(config['data']['ahead']) == list) else [1, config['data']['ahead']]
if type(config['data']['ahead']) == list:
iahead, sahead = config['data']['ahead']
else:
iahead, sahead = 1, config['data']['ahead']
# lresults = []
# Dataset
dataset = Dataset(config=config['data'], data_path=wind_data_path)
dataset.generate_dataset(ahead=ahead, mode=architecture.data_mode, remote=runconfig.remote)
train_x, train_y, val_x, val_y, test_x, test_y = dataset.get_data_matrices()
############################################
# Model
arch = architecture(config, runconfig)
arch.generate_model()
if runconfig.verbose:
arch.summary()
dataset.summary()
print()
############################################
# Training
arch.train(train_x, train_y, val_x, val_y)
############################################
saverrors = None
print('scikit')
print(config['training'])
print(config['training']['saverrors'])
if 'saverrors' in config['training'] and config['training']['saverrors']:
# saverrors = f'-{ahead[0]}-{ahead[1]}-R{iter}' ¿?por que no va?
saverrors = f'-{ahead[0]}-{ahead[1]}'
# Results
if 'descale' not in config['training'] or config['training']['descale']:
lresults = arch.evaluate(val_x, val_y, test_x, test_y, scaler=dataset.scaler,save_errors=saverrors)
else:
lresults = arch.evaluate(val_x, val_y, test_x, test_y, save_errors=saverrors)
print(strftime('%Y-%m-%d %H:%M:%S'))
if config is not None:
updateprocess(config, ahead)
arch.log_result(lresults)
return lresults
def train_persistence(architecture, config, runconfig):
"""
Training process for architecture with direct regression of ahead time steps
:return:
"""
if type(config['data']['ahead']) == list:
iahead, sahead = config['data']['ahead']
else:
iahead, sahead = 1, config['data']['ahead']
lresults = []
for ahead in range(iahead, sahead + 1):
if runconfig.verbose:
print(
"-----------------------------------------------------------------------------"
)
print(f"Steps Ahead = {ahead}")
# Dataset
dataset = Dataset(config=config['data'], data_path=wind_data_path)
dataset.generate_dataset(ahead=[ahead, ahead],
mode=architecture.data_mode,
remote=runconfig.remote)
train_x, train_y, val_x, val_y, test_x, test_y = dataset.get_data_matrices(
)
# Architecture
arch = architecture(config, runconfig)
if runconfig.verbose:
dataset.summary()
if 'descale' not in config['training'] or config['training']['descale']:
lresults.append([ahead] + arch.evaluate(
val_x, val_y, test_x, test_y, scaler=dataset.scaler))
else:
lresults.append([ahead] +
arch.evaluate(val_x, val_y, test_x, test_y))
print(strftime('%Y-%m-%d %H:%M:%S'))
# Update result in db
if config is not None:
updateprocess(config, ahead)
del dataset
arch.log_result(lresults)
return lresults
def train_sckit_sequence2sequence(architecture, config, runconfig):
"""
Training process for architecture with direct regression of ahead time steps
:return:
"""
ahead = config['data']['ahead'] if (type(config['data']['ahead']) == list) else [1, config['data']['ahead']]
if type(config['data']['ahead']) == list:
iahead, sahead = config['data']['ahead']
else:
iahead, sahead = 1, config['data']['ahead']
# lresults = []
# Dataset
dataset = Dataset(config=config['data'], data_path=wind_data_path)
dataset.generate_dataset(ahead=ahead, mode=architecture.data_mode, remote=runconfig.remote)
train_x, train_y, val_x, val_y, test_x, test_y = dataset.get_data_matrices()
############################################
# Model
arch = architecture(config, runconfig)
arch.generate_model()
if runconfig.verbose:
arch.summary()
dataset.summary()
print()
############################################
# Training
arch.train(train_x, train_y, val_x, val_y)
############################################
# Results
lresults = arch.evaluate(val_x, val_y, test_x, test_y)
print(strftime('%Y-%m-%d %H:%M:%S'))
if config is not None:
updateprocess(config, ahead)
arch.log_result(lresults)
return lresults
def train_dirregression(architecture, config, runconfig):
"""
Training process for architecture with direct regression of ahead time steps
:return:
"""
if type(config['data']['ahead']) == list:
iahead, sahead = config['data']['ahead']
else:
iahead, sahead = 1, config['data']['ahead']
lresults = []
if 'iter' in config['training']:
niter = config['training']['iter']
else:
niter = 1
for iter in range(niter):
for ahead in range(iahead, sahead + 1):
if runconfig.verbose:
print('-----------------------------------------------------------------------------')
print(f"Steps Ahead = {ahead}")
# Dataset
dataset = Dataset(config=config['data'], data_path=wind_data_path)
dataset.generate_dataset(ahead=[ahead, ahead], mode=architecture.data_mode, remote=runconfig.remote)
train_x, train_y, val_x, val_y, test_x, test_y = dataset.get_data_matrices()
############################################
# Model
config['idimensions'] = train_x.shape[1:]
arch = architecture(config, runconfig)
if runconfig.multi == 1:
arch.generate_model()
else:
with tf.device('/cpu:0'):
arch.generate_model()
if runconfig.verbose:
arch.summary()
arch.plot()
dataset.summary()
print()
############################################
# Training
arch.train(train_x, train_y, val_x, val_y)
############################################
# Results
lresults.append([ahead] + arch.evaluate(val_x, val_y, test_x, test_y))
print(strftime('%Y-%m-%d %H:%M:%S'))
# Update result in db
if config is not None and not runconfig.proxy:
from Wind.DataBaseConfigurations import updateprocess
updateprocess(config, ahead)
arch.save('-A%d-R%02d' % (ahead, iter))
del dataset
arch.log_result(lresults)
return lresults
def train_sjoint_sequence2sequence(architecture, config, runconfig):
"""
Training process for architecture with multiple blocks of horizons
Multihorizon SJOINT strategy
The training is done separately for blocks of horizons (if block size is 1 this is dirregression)
:return:
"""
if type(config['data']['ahead']) == list:
iahead, sahead = config['data']['ahead']
else:
iahead, sahead = 1, config['data']['ahead']
# Number of consecutive horizon elements to join in a prediction
slice = config['data']['slice']
# if (sahead - (iahead-1)) % slice != 0:
# raise NameError("SJOINT: slice has to be a divisor of the horizon length")
lresults = []
if 'iter' in config['training']:
niter = config['training']['iter']
else:
niter = 1
lmodels = []
steps = [[
i, j
] for i, j in zip(range(iahead, sahead +
1, slice), range(slice, sahead + slice + 1, slice))
]
steps[-1][1] = sahead
for iter in range(niter):
# Loads the dataset once and slices the y matrix for training and evaluation
dataset = Dataset(config=config['data'], data_path=wind_data_path)
dataset.generate_dataset(ahead=[iahead, sahead],
mode=architecture.data_mode,
remote=runconfig.remote)
train_x, train_y, val_x, val_y, test_x, test_y = dataset.get_data_matrices(
)
for recit, ahead in enumerate(steps):
if runconfig.verbose:
print(
'-----------------------------------------------------------------------------'
)
print(f"Steps Ahead = {ahead}")
############################################
# Model
config['idimensions'] = train_x.shape[1:]
config['odimensions'] = ahead[1] - ahead[0] + 1
nconfig = deepcopy(config)
nconfig['data']['ahead'] = ahead
arch = architecture(nconfig, runconfig)
if runconfig.multi == 1:
arch.generate_model()
else:
with tf.device('/cpu:0'):
arch.generate_model()
if runconfig.verbose:
arch.summary()
arch.plot()
dataset.summary()
print()
############################################
# Training with the current slice
arch.train(train_x, train_y[:, ahead[0] - 1:ahead[1]], val_x,
val_y[:, ahead[0] - 1:ahead[1]])
############################################
# Results
if 'descale' not in config['training'] or config['training'][
'descale']:
lresults.extend(
arch.evaluate(val_x,
val_y[:, ahead[0] - 1:ahead[1]],
test_x,
test_y[:, ahead[0] - 1:ahead[1]],
scaler=dataset.scaler))
else:
lresults.extend(
arch.evaluate(val_x, val_y[:, ahead[0] - 1:ahead[1]],
test_x, test_y[:, ahead[0] - 1:ahead[1]]))
print(strftime('%Y-%m-%d %H:%M:%S'))
# Update result in db
if config is not None and not runconfig.proxy:
from Wind.DataBaseConfigurations import updateprocess
updateprocess(config, ahead)
arch.save(f"-{ahead[0]}-{ahead[1]}-S{recit:02d}-R{iter:02d}")
arch.log_result(lresults)
return lresults