rename to optimization and define objective function

mlforecast/_modidx.py

@@ -332,6 +332,8 @@
                                    'mlforecast.lgb_cv._rmse': ('lgb_cv.html#_rmse', 'mlforecast/lgb_cv.py'),
                                    'mlforecast.lgb_cv._update': ('lgb_cv.html#_update', 'mlforecast/lgb_cv.py'),
                                    'mlforecast.lgb_cv._update_and_predict': ('lgb_cv.html#_update_and_predict', 'mlforecast/lgb_cv.py')},
+            'mlforecast.optimization': { 'mlforecast.optimization.mlforecast_objective': ( 'optimization.html#mlforecast_objective',
+                                                                                           'mlforecast/optimization.py')},
             'mlforecast.target_transforms': { 'mlforecast.target_transforms.AutoDifferences': ( 'target_transforms.html#autodifferences',
                                                                                                 'mlforecast/target_transforms.py'),
                                               'mlforecast.target_transforms.AutoDifferences.__init__': ( 'target_transforms.html#autodifferences.__init__',

mlforecast/core.py

@@ -126,7 +126,7 @@ def _as_tuple(x):
     return (x,)
 
 # %% ../nbs/core.ipynb 19
-Freq = Union[int, str, pd.offsets.BaseOffset]
+Freq = Union[int, str]
 Lags = Iterable[int]
 LagTransform = Union[Callable, Tuple[Callable, Any]]
 LagTransforms = Dict[int, List[LagTransform]]

mlforecast/optimization.py

@@ -0,0 +1,84 @@
+# AUTOGENERATED! DO NOT EDIT! File to edit: ../nbs/optimization.ipynb.
+
+# %% auto 0
+__all__ = ['mlforecast_objective']
+
+# %% ../nbs/optimization.ipynb 2
+import copy
+from typing import Callable, List, Optional
+
+import numpy as np
+import optuna
+import utilsforecast.processing as ufp
+from utilsforecast.compat import DataFrame
+from utilsforecast.losses import smape
+
+from . import MLForecast
+from .core import Freq
+
+# %% ../nbs/optimization.ipynb 3
+def mlforecast_objective(
+    df: DataFrame,
+    config_fn: Callable,
+    eval_fn: Callable,
+    model_constructor: Callable,
+    freq: Freq,
+    n_windows: int,
+    h: int,
+    id_col: str = "unique_id",
+    time_col: str = "ds",
+    target_col: str = "y",
+) -> Callable:
+    def objective(trial: optuna.Trial) -> float:
+        config = config_fn(trial)
+        trial.set_user_attr("config", copy.deepcopy(config))
+        splits = ufp.backtest_splits(
+            df,
+            n_windows=n_windows,
+            h=h,
+            id_col=id_col,
+            time_col=time_col,
+            freq=freq,
+        )
+        metrics = []
+        for i, (_, train, valid) in enumerate(splits):
+            mlf = MLForecast(
+                models={"model": model_constructor(**config["model_params"])},
+                freq=freq,
+                **config["mlf_init_params"],
+            )
+            mlf.fit(
+                train,
+                id_col=id_col,
+                time_col=time_col,
+                target_col=target_col,
+                **config["mlf_fit_params"],
+            )
+            static = [c for c in mlf.ts.static_features_.columns if c != id_col]
+            dynamic = [
+                c
+                for c in valid.columns
+                if c not in static + [id_col, time_col, target_col]
+            ]
+            if dynamic:
+                X_df: Optional[DataFrame] = ufp.drop_columns(
+                    valid, static + [target_col]
+                )
+            else:
+                X_df = None
+            preds = mlf.predict(h=h, X_df=X_df)
+            full = valid.merge(preds, on=[id_col, time_col], how="left")
+            if full.shape[0] < valid.shape[0]:
+                raise ValueError(
+                    "Cross validation result produced less results than expected. "
+                    "Please verify that the passed frequency (freq) matches your series' "
+                    "and that there aren't any missing periods."
+                )
+            metric = eval_fn(full)
+            metrics.append(metric)
+            trial.report(metric, step=i)
+            if trial.should_prune():
+                raise optuna.TrialPruned()
+        return np.mean(metrics)
+
+    return objective

nbs/core.ipynb

@@ -602,7 +602,7 @@
    "outputs": [],
    "source": [
     "#| exporti\n",
-    "Freq = Union[int, str, pd.offsets.BaseOffset]\n",
+    "Freq = Union[int, str]\n",
     "Lags = Iterable[int]\n",
     "LagTransform = Union[Callable, Tuple[Callable, Any]]\n",
     "LagTransforms = Dict[int, List[LagTransform]]\n",