def max_abs_normalization(x: np.ndarray) -> np.ndarray:
"""
max abs normalization
formula:
y = x / max(|x|)
"""
factor = x.abs().max()
if factor == 0:
canceled_to_avoid_div0("max_abs_normalization")
return x
else:
return x / factor
def numpy_normalised_quantile_loss(y: ndarray, y_pred: ndarray,
quantile: float) -> float:
"""Computes normalised quantile loss for numpy arrays.
Uses the q-Risk metric as defined in the "Training Procedure" section of the
main TFT paper.
Args:
y: Targets
y_pred: Predictions
quantile: Quantile to use for loss calculations (between 0 & 1)
Returns:
Float for normalised quantile loss.
"""
prediction_underflow: ndarray = y - y_pred
weighted_errors: ndarray = quantile * np.maximum(prediction_underflow, 0.) \
+ (1. - quantile) * np.maximum(-prediction_underflow, 0.)
quantile_loss: float = weighted_errors.mean()
normaliser: float = y.abs().mean()
return 2 * quantile_loss / normaliser