FIX Covariates: enforce length invariant, fix slice() indexing

benchmark_utils/base_solver.py

@@ -251,7 +251,9 @@ def forecast(
             for cutoff_idx, cutoff in enumerate(cutoffs):
                 hist = series[:cutoff]  # (T_cutoff, C)
                 inputs.append(torch.from_numpy(hist))
-                covariates.append(x.covariates.slice(cutoff, prediction_length))
+                covariates.append(
+                    x.covariates.slice(series_idx, cutoff, prediction_length)
+                )
                 layout.append((series_idx, cutoff_idx))
 
         if not inputs:

benchmark_utils/covariates.py

@@ -1,11 +1,11 @@
 """Covariates payload passed to forecasting adapters.
 
-A small dataclass so the contract is typed and IDE-discoverable. All
-three fields default to empty sequences, so datasets without covariates
-can just pass ``Covariates()``.
+Each field is either ``None`` (that covariate kind is absent for every
+series) or a sequence with one entry per series. Datasets without
+covariates just pass ``Covariates()`` (all fields ``None``).
 """
 
-from dataclasses import dataclass, field
+from dataclasses import dataclass
 from typing import Sequence
 
 import numpy as np
@@ -15,38 +15,58 @@
 class Covariates:
     """Per-series covariates aligned with the ``x`` sequence in ``predict``.
 
-    Each field is a sequence whose length equals ``len(x)``. Within a
-    series, the inner structure depends on the covariate kind — see the
-    forecasting predict() contract in :mod:`benchmark_utils.adapters.base`.
+    Each field is either ``None`` (absent for every series) or a sequence
+    whose length equals ``len(x)`` — one entry per series. The per-series
+    element shapes are listed below; see the forecasting predict() contract
+    in :mod:`benchmark_utils.adapters.base`.
 
     Parameters
     ----------
     static_covars
-        Shape is (channels,)
+        ``None``, or a length-``len(x)`` sequence of arrays of shape (channels,)
     hist_covars
-        Shape is (time, channels)
+        ``None``, or a length-``len(x)`` sequence of arrays of shape (time, channels)
     future_covars
-        Shape is (time, channels)
+        ``None``, or a length-``len(x)`` sequence of arrays of shape (time, channels)
     """
 
-    static_covars: Sequence[np.ndarray] = field(default_factory=list)
-    hist_covars: Sequence[np.ndarray] = field(default_factory=list)
-    future_covars: Sequence[np.ndarray] = field(default_factory=list)
+    static_covars: Sequence[np.ndarray] | None = None
+    hist_covars: Sequence[np.ndarray] | None = None
+    future_covars: Sequence[np.ndarray] | None = None
 
     def __post_init__(self):
-        if len(self.static_covars) != len(self.hist_covars) != len(self.future_covars):
+        # Every provided (non-None) field must cover the same set of series.
+        lengths = {
+            len(f)
+            for f in (self.static_covars, self.hist_covars, self.future_covars)
+            if f is not None
+        }
+        if len(lengths) > 1:
             raise ValueError(
-                "All covariate sequences must have the same length as x"
+                "All provided covariate sequences must have the same length"
             )
-    
+
     def __len__(self) -> int:
-        # or hist_covars or future_covars, they all have the same length
-        return len(self.static_covars)
+        """Number of series covered (0 if no covariates are present)."""
+        for f in (self.static_covars, self.hist_covars, self.future_covars):
+            if f is not None:
+                return len(f)
+        return 0
+
+    def slice(self, series_idx: int, cutoff: int, horizon: int) -> 'Covariates':
+        """Covariates for a single ``(series, cutoff)`` window.
 
-    def slice(self, cutoff: int, horizon: int) -> 'Covariates':
-        """Get covariates for a single series."""
+        Selects series ``series_idx`` and slices its time axis: history up to
+        ``cutoff`` and the future window ``[cutoff, cutoff + horizon)``.
+        """
         return Covariates(
-            static_covars=self.static_covars,
-            hist_covars=self.hist_covars[:cutoff],
-            future_covars=self.future_covars[cutoff:cutoff + horizon],
+            static_covars=None
+            if self.static_covars is None
+            else [self.static_covars[series_idx]],
+            hist_covars=None
+            if self.hist_covars is None
+            else [self.hist_covars[series_idx][:cutoff]],
+            future_covars=None
+            if self.future_covars is None
+            else [self.future_covars[series_idx][cutoff:cutoff + horizon]],
         )

benchmark_utils/inputs.py

@@ -26,8 +26,9 @@ class ForecastInput:
     cutoff_indexes : sequence of sequence of int
         Jagged — per-series timestep indexes at which a forecast starts.
     covariates : Covariates
-        Static / historical / future covariates aligned with ``x``.
-        Defaults to empty.
+        Static / historical / future covariates aligned with ``x``. Each
+        covariate field is either ``None`` (absent for every series) or has
+        one entry per series (length ``len(x)``). Defaults to all-``None``.
     """
 
     x: Sequence[np.ndarray]
@@ -37,6 +38,12 @@ class ForecastInput:
     def __post_init__(self):
         if len(self.x) != len(self.cutoff_indexes):
             raise ValueError("x and cutoff_indexes must have the same length")
-        # TODO len(self.covariates) == 0 for some
-        # if len(self.x) != len(self.covariates):
-        #     raise ValueError("x and covariates must have the same length")
+        # Covariates either cover all series (one entry each) or are absent.
+        # len(covariates) is the shared per-field length (0 when absent), so
+        # covariate-free inputs need no boilerplate.
+        n_cov = len(self.covariates)
+        if n_cov not in (0, len(self.x)):
+            raise ValueError(
+                f"covariates must cover every series: got length {n_cov}, "
+                f"expected 0 (absent) or len(x)={len(self.x)}"
+            )

datasets/enedis.py

@@ -158,7 +158,6 @@ def get_data(self):
         y_train = [target[first_cut : first_cut + pred_len]]
 
         covariates = Covariates(
-            static_covars=[],
             hist_covars=[hist_covar],
             future_covars=[future_covar],
         )

datasets/monash.py

@@ -20,9 +20,9 @@
                                                   positions in X_test
 y_test          : List[np.ndarray (n_cutoffs, H, C)]
                                                   ground-truth windows
-covariates      : dict                           {static_covars, hist_covars,
-                                                  future_covars} — all empty for
-                                                  Monash today
+covariates      : Covariates                     {static_covars, hist_covars,
+                                                  future_covars} — all absent
+                                                  (None) for Monash today
 task            : "forecasting"
 metrics         : ["mae", "mse", "rmse", "mase", "smape",
                    "crps", "wql", "mcis", "pinball", "skill_score_ratio"]