class TCMF(BaseModel):
"""
MF regularized TCN + TCN. This version is not for automated searching yet.
"""
def __init__(self):
"""
Initialize hyper parameters
:param check_optional_config:
:param future_seq_len:
"""
# models
self.model = None
self.model_init = False
def set_params(self, **config):
self.vbsize = config.get("vbsize", 128)
self.hbsize = config.get("hbsize", 256)
self.num_channels_X = config.get("num_channels_X",
[32, 32, 32, 32, 32, 1])
self.num_channels_Y = config.get("num_channels_Y",
[16, 16, 16, 16, 16, 1])
self.kernel_size = config.get("kernel_size", 7)
self.dropout = config.get("dropout", 0.1)
self.rank = config.get("rank", 64)
self.kernel_size_Y = config.get("kernel_size_Y", 7)
self.lr = config.get("learning_rate", 0.0005)
self.val_len = config.get("val_len", 24)
self.normalize = config.get("normalize", False)
self.start_date = config.get("start_date", "2020-4-1")
self.freq = config.get("freq", "1H")
self.covariates = config.get('covariates', None)
self.use_time = config.get("use_time", True)
self.dti = config.get("dti", None)
self.svd = config.get("svd", True)
self.period = config.get("period", 24)
self.alt_iters = config.get("alt_iters", 10)
self.y_iters = config.get("y_iters", 10)
self.init_epoch = config.get("init_FX_epoch", 100)
self.max_FX_epoch = config.get("max_FX_epoch", 300)
self.max_TCN_epoch = config.get("max_TCN_epoch", 300)
def _build(self, **config):
"""
build the models and initialize.
:param config: hyper parameters for building the model
:return:
"""
self.set_params(**config)
self.model = DeepGLO(vbsize=self.vbsize,
hbsize=self.hbsize,
num_channels_X=self.num_channels_X,
num_channels_Y=self.num_channels_Y,
kernel_size=self.kernel_size,
dropout=self.dropout,
rank=self.rank,
kernel_size_Y=self.kernel_size_Y,
lr=self.lr,
val_len=self.val_len,
normalize=self.normalize,
start_date=self.start_date,
freq=self.freq,
covariates=self.covariates,
use_time=self.use_time,
dti=self.dti,
svd=self.svd,
period=self.period,
forward_cov=False)
self.model_init = True
def fit_eval(self, x, y=None, verbose=0, num_workers=None, **config):
"""
Fit on the training data from scratch.
Since the rolling process is very customized in this model,
we enclose the rolling process inside this method.
:param x: training data, an array in shape (nd, Td),
nd is the number of series, Td is the time dimension
:param y: None. target is extracted from x directly
:param verbose:
:param num_workers: number of workers to use.
:return: the evaluation metric value
"""
if not self.model_init:
self._build(**config)
if num_workers is None:
num_workers = TCMF.get_default_num_workers()
val_loss = self.model.train_all_models(
x,
alt_iters=self.alt_iters,
y_iters=self.y_iters,
init_epochs=self.init_epoch,
max_FX_epoch=self.max_FX_epoch,
max_TCN_epoch=self.max_TCN_epoch,
num_workers=num_workers,
)
return val_loss
def fit_incremental(self, x):
"""
Incremental fitting given a pre-trained model.
:param x: incremental data
:param config: fitting parameters
:return:
"""
# TODO incrementally train models
pass
@staticmethod
def get_default_num_workers():
from zoo.ray import RayContext
try:
ray_ctx = RayContext.get(initialize=False)
num_workers = ray_ctx.num_ray_nodes
except:
num_workers = 1
return num_workers
def predict(self, x=None, horizon=24, mc=False, num_workers=None):
"""
Predict horizon time-points ahead the input x in fit_eval
:param x: We don't support input x currently.
:param horizon: horizon length to predict
:param mc:
:param num_workers: the number of workers to use. Note that there has to be an activate
RayContext if num_workers > 1.
:return:
"""
if x is not None:
raise ValueError("We don't support input x directly.")
if self.model is None:
raise Exception(
"Needs to call fit_eval or restore first before calling predict"
)
if num_workers is None:
num_workers = TCMF.get_default_num_workers()
if num_workers > 1:
import ray
from zoo.ray import RayContext
try:
RayContext.get(initialize=False)
except:
try:
# detect whether ray has been started.
ray.put(None)
except:
raise RuntimeError(
f"There must be an activate ray context while running with "
f"{num_workers} workers. You can either start and init a "
f"RayContext by init_orca_context(..., init_ray_on_spark="
f"True) or start Ray with ray.init()")
out = self.model.predict_horizon(
future=horizon,
bsize=90,
normalize=False,
num_workers=num_workers,
)
return out[:, -horizon::]
def evaluate(self, x=None, y=None, metrics=None, num_workers=None):
"""
Evaluate on the prediction results and y. We predict horizon time-points ahead the input x
in fit_eval before evaluation, where the horizon length equals the second dimension size of
y.
:param x: We don't support input x currently.
:param y: target. We interpret the second dimension of y as the horizon length for
evaluation.
:param metrics: a list of metrics in string format
:param num_workers: the number of workers to use in evaluate. It defaults to 1.
:return: a list of metric evaluation results
"""
if x is not None:
raise ValueError("We don't support input x directly.")
if y is None:
raise ValueError("Input invalid y of None")
if self.model is None:
raise Exception(
"Needs to call fit_eval or restore first before calling predict"
)
if len(y.shape) == 1:
y = np.expand_dims(y, axis=1)
horizon = 1
else:
horizon = y.shape[1]
result = self.predict(x=None, horizon=horizon, num_workers=num_workers)
if y.shape[1] == 1:
multioutput = 'uniform_average'
else:
multioutput = 'raw_values'
return [
Evaluator.evaluate(m, y, result, multioutput=multioutput)
for m in metrics
]
def save(self, model_file):
pickle.dump(self.model, open(model_file, "wb"))
def restore(self, model_file):
with open(model_file, 'rb') as f:
self.model = pickle.load(f)
self.model_init = True
def _get_optional_parameters(self):
return {}
def _get_required_parameters(self):
return {}
class TCMF(BaseModel):
"""
MF regularized TCN + TCN. This version is not for automated searching yet.
"""
def __init__(self):
"""
Initialize hyper parameters
:param check_optional_config:
:param future_seq_len:
"""
# models
self.model = None
self.model_init = False
def build(self, config):
"""
build the models and initialize.
:param config: hyper parameters for building the model
:return:
"""
self.model = DeepGLO(
vbsize=config.get("vbsize", 128),
hbsize=config.get("hbsize", 256),
num_channels_X=config.get("num_channels_X",
[32, 32, 32, 32, 32, 1]),
num_channels_Y=config.get("num_channels_Y",
[16, 16, 16, 16, 16, 1]),
kernel_size=config.get("kernel_size", 7),
dropout=config.get("dropout", 0.1),
rank=config.get("rank", 64),
kernel_size_Y=config.get("kernel_size_Y", 7),
lr=config.get("learning_rate", 0.0005),
normalize=config.get("normalize", False),
use_time=config.get("use_time", True),
svd=config.get("svd", True),
forward_cov=False)
self.model_init = True
def fit_eval(self, x, y=None, verbose=0, num_workers=None, **config):
"""
Fit on the training data from scratch.
Since the rolling process is very customized in this model,
we enclose the rolling process inside this method.
:param x: training data, an array in shape (nd, Td),
nd is the number of series, Td is the time dimension
:param y: None. target is extracted from x directly
:param verbose:
:param num_workers: number of workers to use.
:return: the evaluation metric value
"""
if not self.model_init:
self.build(config)
if num_workers is None:
num_workers = TCMF.get_default_num_workers()
covariates = config.get('covariates', None)
dti = config.get("dti", None)
self._check_covariates_dti(covariates=covariates,
dti=dti,
ts_len=x.shape[1])
val_loss = self.model.train_all_models(
x,
val_len=config.get("val_len", 24),
start_date=config.get("start_date", "2020-4-1"),
freq=config.get("freq", "1H"),
covariates=covariates,
dti=dti,
period=config.get("period", 24),
init_epochs=config.get("init_FX_epoch", 100),
alt_iters=config.get("alt_iters", 10),
y_iters=config.get("y_iters", 10),
max_FX_epoch=config.get("max_FX_epoch", 300),
max_TCN_epoch=config.get("max_TCN_epoch", 300),
num_workers=num_workers,
)
return val_loss
def fit_incremental(self, x, covariates_new=None, dti_new=None):
"""
Incremental fitting given a pre-trained model.
:param x: incremental data
:param covariates_new: covariates corresponding to the incremental x
:param dti_new: dti corresponding to the incremental x
:return:
"""
if x is None:
raise ValueError("Input invalid x of None")
if self.model is None:
raise Exception(
"Needs to call fit_eval or restore first before calling "
"fit_incremental")
self._check_covariates_dti(covariates=covariates_new,
dti=dti_new,
ts_len=x.shape[1],
method_name='fit_incremental')
self.model.inject_new(x,
covariates_new=covariates_new,
dti_new=dti_new)
@staticmethod
def get_default_num_workers():
from zoo.ray import RayContext
try:
ray_ctx = RayContext.get(initialize=False)
num_workers = ray_ctx.num_ray_nodes
except:
num_workers = 1
return num_workers
def predict(self,
x=None,
horizon=24,
mc=False,
future_covariates=None,
future_dti=None,
num_workers=None):
"""
Predict horizon time-points ahead the input x in fit_eval
:param x: We don't support input x currently.
:param horizon: horizon length to predict
:param mc:
:param future_covariates: covariates corresponding to future horizon steps data to predict.
:param future_dti: dti corresponding to future horizon steps data to predict.
:param num_workers: the number of workers to use. Note that there has to be an activate
RayContext if num_workers > 1.
:return:
"""
if x is not None:
raise ValueError("We don't support input x directly.")
if self.model is None:
raise Exception(
"Needs to call fit_eval or restore first before calling predict"
)
self._check_covariates_dti(covariates=future_covariates,
dti=future_dti,
ts_len=horizon,
method_name="predict")
if num_workers is None:
num_workers = TCMF.get_default_num_workers()
if num_workers > 1:
import ray
from zoo.ray import RayContext
try:
RayContext.get(initialize=False)
except:
try:
# detect whether ray has been started.
ray.put(None)
except:
raise RuntimeError(
f"There must be an activate ray context while running with "
f"{num_workers} workers. You can either start and init a "
f"RayContext by init_orca_context(..., init_ray_on_spark="
f"True) or start Ray with ray.init()")
out = self.model.predict_horizon(
future=horizon,
bsize=90,
num_workers=num_workers,
future_covariates=future_covariates,
future_dti=future_dti,
)
return out[:, -horizon::]
def evaluate(self,
x=None,
y=None,
metrics=None,
target_covariates=None,
target_dti=None,
num_workers=None):
"""
Evaluate on the prediction results and y. We predict horizon time-points ahead the input x
in fit_eval before evaluation, where the horizon length equals the second dimension size of
y.
:param x: We don't support input x currently.
:param y: target. We interpret the second dimension of y as the horizon length for
evaluation.
:param metrics: a list of metrics in string format
:param target_covariates: covariates corresponding to target_value.
2-D ndarray or None.
The shape of ndarray should be (r, horizon), where r is the number of covariates.
Global covariates for all time series. If None, only default time coveriates will be
used while use_time is True. If not, the time coveriates used is the stack of input
covariates and default time coveriates.
:param target_dti: dti corresponding to target_value.
DatetimeIndex or None.
If None, use default fixed frequency DatetimeIndex generated with the last date of x in
fit and freq.
:param num_workers: the number of workers to use in evaluate. It defaults to 1.
:return: a list of metric evaluation results
"""
if x is not None:
raise ValueError("We don't support input x directly.")
if y is None:
raise ValueError("Input invalid y of None")
if self.model is None:
raise Exception(
"Needs to call fit_eval or restore first before calling predict"
)
if len(y.shape) == 1:
y = np.expand_dims(y, axis=1)
horizon = 1
else:
horizon = y.shape[1]
result = self.predict(x=None,
horizon=horizon,
future_covariates=target_covariates,
future_dti=target_dti,
num_workers=num_workers)
if y.shape[1] == 1:
multioutput = 'uniform_average'
else:
multioutput = 'raw_values'
return [
Evaluator.evaluate(m, y, result, multioutput=multioutput)
for m in metrics
]
def save(self, model_file):
pickle.dump(self.model, open(model_file, "wb"))
def restore(self, model_file):
with open(model_file, 'rb') as f:
self.model = pickle.load(f)
self.model_init = True
def _get_optional_parameters(self):
return {}
def _get_required_parameters(self):
return {}
def _check_covariates_dti(self,
covariates=None,
dti=None,
ts_len=24,
method_name='fit'):
if covariates is not None and not isinstance(covariates, np.ndarray):
raise ValueError("Input covariates must be a ndarray. Got",
type(covariates))
if covariates is not None and not covariates.ndim == 2:
raise ValueError(
"You should input a 2-D ndarray of covariates. But Got dimension of",
covariates.ndim)
if covariates is not None and not covariates.shape[1] == ts_len:
raise ValueError(
f"The second dimension shape of covariates should be {ts_len}, "
f"but got {covariates.shape[1]} instead.")
if dti is not None and not isinstance(dti, pd.DatetimeIndex):
raise ValueError("Input dti must be a pandas DatetimeIndex. Got",
type(dti))
if dti is not None and len(dti) != ts_len:
raise ValueError(f"Input dti length should be equal to {ts_len}, "
f"but got {len(dti)} instead.")
if method_name != 'fit':
# covariates and dti should be consistent with that in fit
if self.model.covariates is None and covariates is not None:
raise ValueError(
f"Find valid covariates in {method_name} but invalid covariates "
f"in fit. Please keep them in consistence!")
if self.model.covariates is not None and covariates is None:
raise ValueError(
f"Find valid covariates in fit but invalid covariates in "
f"{method_name}. Please keep them in consistence!")
if self.model.covariates is not None \
and self.model.covariates.shape[0] != covariates.shape[0]:
raise ValueError(
f"The input covariates number in {method_name} should be the same "
f"as the input covariates number in fit. Got {covariates.shape[0]}"
f"and {self.model.covariates.shape[0]} respectively.")
if self.model.dti is None and dti is not None:
raise ValueError(
f"Find valid dti in {method_name} but invalid dti in fit. "
f"Please keep them in consistence!")
if self.model.dti is not None and dti is None:
raise ValueError(
f"Find valid dti in fit but invalid dti in {method_name}. "
f"Please keep them in consistence!")