Merge pull request #125 from basf/develop

include quantile regression

README.md

@@ -200,6 +200,7 @@ MambularLSS allows you to model the full distribution of a response variable, no
 - **negativebinom**: For over-dispersed count data.
 - **inversegamma**: Often used as a prior in Bayesian inference.
 - **categorical**: For data with more than two categories.
+- **Quantile**: For quantile regression using the pinball loss.
 
 These distribution classes make MambularLSS versatile in modeling various data types and distributions.
 

mambular/__version__.py

@@ -1,4 +1,4 @@
 """Version information."""
 
 # The following line *must* be the last in the module, exactly as formatted:
-__version__ = "0.2.2"
+__version__ = "0.2.3"

mambular/models/sklearn_base_lss.py

@@ -30,6 +30,7 @@
     NormalDistribution,
     PoissonDistribution,
     StudentTDistribution,
+    Quantile,
 )
 from lightning.pytorch.callbacks import ModelSummary
 
@@ -210,11 +211,9 @@ def build_model(
             X, y, X_val, y_val, val_size=val_size, random_state=random_state
         )
 
-        num_classes = len(np.unique(y))
-
         self.task_model = TaskModel(
             model_class=self.base_model,
-            num_classes=num_classes,
+            num_classes=self.family.param_count,
             config=self.config,
             cat_feature_info=self.data_module.cat_feature_info,
             num_feature_info=self.data_module.num_feature_info,
@@ -347,6 +346,7 @@ def fit(
             "negativebinom": NegativeBinomialDistribution,
             "inversegamma": InverseGammaDistribution,
             "categorical": CategoricalDistribution,
+            "quantile": Quantile,
         }
 
         if distributional_kwargs is None:

mambular/utils/distributions.py

@@ -504,3 +504,52 @@ def compute_loss(self, predictions, y_true):
         # Compute the negative log-likelihood
         nll = -cat_dist.log_prob(y_true).mean()
         return nll
+
+
+class Quantile(BaseDistribution):
+    """
+    Quantile Regression Loss class.
+
+    This class computes the quantile loss (also known as pinball loss) for a set of quantiles.
+    It is used to handle quantile regression tasks where we aim to predict a given quantile of the target distribution.
+
+    Parameters
+    ----------
+    name : str, optional
+        The name of the distribution, by default "Quantile".
+    quantiles : list of float, optional
+        A list of quantiles to be used for computing the loss, by default [0.25, 0.5, 0.75].
+
+    Attributes
+    ----------
+    quantiles : list of float
+        List of quantiles for which the pinball loss is computed.
+
+    Methods
+    -------
+    compute_loss(predictions, y_true)
+        Computes the quantile regression loss between the predictions and true values.
+    """
+
+    def __init__(self, name="Quantile", quantiles=[0.25, 0.5, 0.75]):
+        param_names = [
+            f"q_{q}" for q in quantiles
+        ]  # Use string representations of quantiles
+        super().__init__(name, param_names)
+        self.quantiles = quantiles
+
+    def compute_loss(self, predictions, y_true):
+
+        assert not y_true.requires_grad  # Ensure y_true does not require gradients
+        assert predictions.size(0) == y_true.size(0)  # Ensure batch size matches
+
+        losses = []
+        for i, q in enumerate(self.quantiles):
+            errors = y_true - predictions[:, i]  # Calculate errors for each quantile
+            # Compute the pinball loss
+            quantile_loss = torch.max((q - 1) * errors, q * errors)
+            losses.append(quantile_loss)
+
+        # Sum losses across quantiles and compute mean
+        loss = torch.mean(torch.stack(losses, dim=1).sum(dim=1))
+        return loss