def Dist_Objective(predt: np.ndarray, data: lgb.Dataset):
"""A customized objective function to train each distributional parameter using custom gradient and hessian.
"""
target = torch.tensor(data.get_label())
# When num_class!= 0, preds has shape (n_obs, n_dist_param).
# Each element in a row represents a raw prediction (leaf weight, hasn't gone through response function yet).
predt = predt.reshape(-1, Gaussian.n_dist_param(), order="F")
preds_location = Gaussian.param_dict()["location"](predt[:, 0])
preds_location = torch.tensor(preds_location, requires_grad=True)
preds_scale = Gaussian.param_dict()["scale"](predt[:, 1])
preds_scale = torch.tensor(preds_scale, requires_grad=True)
# Weights
if data.get_weight() == None:
# Use 1 as weight if no weights are specified
weights = np.ones_like(target, dtype=float)
else:
weights = data.get_weight()
# Initialize Gradient and Hessian Matrices
grad = np.zeros(shape=(len(target), Gaussian.n_dist_param()))
hess = np.zeros(shape=(len(target), Gaussian.n_dist_param()))
# Specify Metric for Auto Derivation
dGaussian = Normal(preds_location, preds_scale)
autograd_metric = -dGaussian.log_prob(target).nansum()
# Location
grad[:, 0] = stabilize_derivative(
auto_grad(metric=autograd_metric, parameter=preds_location, n=1) *
weights, Gaussian.stabilize)
hess[:, 0] = stabilize_derivative(
auto_grad(metric=autograd_metric, parameter=preds_location, n=2) *
weights, Gaussian.stabilize)
# Scale
grad[:, 1] = stabilize_derivative(
auto_grad(metric=autograd_metric, parameter=preds_scale, n=1) *
weights, Gaussian.stabilize)
hess[:, 1] = stabilize_derivative(
auto_grad(metric=autograd_metric, parameter=preds_scale, n=2) *
weights, Gaussian.stabilize)
# Reshaping
grad = grad.ravel(order="F")
hess = hess.ravel(order="F")
return grad, hess
def Dist_Objective(predt: np.ndarray, data: lgb.Dataset):
"""A customized objective function to train each distributional parameter using custom gradient and hessian.
"""
target = data.get_label()
# When num_class!= 0, preds has shape (n_obs, n_dist_param).
# Each element in a row represents a raw prediction (leaf weight, hasn't gone through response function yet).
predt = predt.reshape(-1, Gaussian.n_dist_param(), order="F")
preds_location = Gaussian.param_dict()["location"](predt[:, 0])
preds_scale = Gaussian.param_dict()["scale"](predt[:, 1])
# Weights
if data.get_weight() == None:
# Use 1 as weight if no weights are specified
weights = np.ones_like(target, dtype=float)
else:
weights = data.get_weight()
# Initialize Gradient and Hessian Matrices
grad = np.zeros(shape=(len(target), Gaussian.n_dist_param()))
hess = np.zeros(shape=(len(target), Gaussian.n_dist_param()))
# Location
grad[:, 0] = Gaussian.gradient_location(y=target,
location=preds_location,
scale=preds_scale,
weights=weights)
hess[:, 0] = Gaussian.hessian_location(scale=preds_scale,
weights=weights)
# Scale
grad[:, 1] = Gaussian.gradient_scale(y=target,
location=preds_location,
scale=preds_scale,
weights=weights)
hess[:, 1] = Gaussian.hessian_scale(scale=preds_scale,
weights=weights)
# Reshaping
grad = grad.ravel(order="F")
hess = hess.ravel(order="F")
return grad, hess
def _evaluate(self, scores: np.ndarray, clases: lgb.Dataset) -> Tuple[str, int, bool]:
labels = clases.get_label()
weights = clases.get_weight()
score_corte = self.prob_corte
nombre, valor = self._evaluar_funcion_ganancia(scores, labels, weights, score_corte)
return nombre, valor, True
def Dist_Objective(predt: np.ndarray, data: lgb.Dataset):
"""A customized objective function to train each distributional parameter using custom gradient and hessian.
"""
target = data.get_label()
# When num_class!= 0, preds has shape (n_obs, n_dist_param).
# Each element in a row represents a raw prediction (leaf weight, hasn't gone through response function yet).
preds_expectile = predt.reshape(-1,
Expectile.n_dist_param(),
order="F")
# Weights
if data.get_weight() == None:
# Use 1 as weight if no weights are specified
weights = np.ones_like(target, dtype=float)
else:
weights = data.get_weight()
# Initialize Gradient and Hessian Matrices
grad = np.zeros(shape=(len(target), len(Expectile.expectiles)))
hess = np.zeros(shape=(len(target), len(Expectile.expectiles)))
for i in range(len(Expectile.expectiles)):
grad[:, i] = Expectile.gradient_expectile(
y=target,
expectile=preds_expectile[:, i],
tau=Expectile.expectiles[i],
weights=weights)
hess[:,
i] = Expectile.hessian_expectile(y=target,
expectile=preds_expectile[:,
i],
tau=Expectile.expectiles[i],
weights=weights)
# Reshaping
grad = grad.ravel(order="F")
hess = hess.ravel(order="F")
return grad, hess
def __call__(self, pred: np.ndarray,
dtrain: lgb.Dataset) -> Tuple[str, float, bool]:
label = dtrain.get_label()
weights = dtrain.get_weight()
if label.shape[0] != pred.shape[0]:
pred = pred.reshape((label.shape[0], -1), order='F')
label = label.astype(np.int32)
pred = self.bw_func(pred)
# for weighted case
try:
val = self.metric_func(label, pred, sample_weight=weights)
except TypeError:
val = self.metric_func(label, pred)
# TODO: what if grouped case
return 'Opt metric', val, self.greater_is_better