[FIX] Allow CV without passing a df (#1545)

Author: marcopeix

neuralforecast/core.py

@@ -1939,6 +1939,23 @@ def _no_refit_cross_validation(
             fcsts_df = ufp.horizontal_concat([fcsts_df, fcsts])
 
             # Add original input df's y to forecasts DataFrame
+            if df is None:
+                # Reconstruct the target from the stored dataset. The dataset's
+                # temporal values are scaled, so undo any target scaling.
+                target_values = (
+                    self.dataset.temporal[:, self.dataset.y_idx].clone().numpy()
+                )
+                if self.scalers_:
+                    target_values = self._scalers_target_inverse_transform(
+                        target_values.reshape(-1, 1), self.dataset.indptr
+                    ).reshape(-1)
+                df = type(fcsts_df)(
+                    {
+                        id_col: ufp.repeat(self.uids, np.diff(self.dataset.indptr)),
+                        time_col: self.ds,
+                        target_col: target_values,
+                    }
+                )
             return ufp.join(
                 fcsts_df,
                 df[[id_col, time_col, target_col]],

tests/test_core.py

@@ -215,6 +215,32 @@ def test_neural_forecast_fit_cross_validation(setup_airplane_data):
     pd.testing.assert_frame_equal(init_cv, after_cv)
     pd.testing.assert_frame_equal(after_fcst, init_fcst)
 
+
+# cross_validation() with no `df` should reuse the stored dataset
+@pytest.mark.parametrize("use_polars", [False, True])
+def test_cross_validation_without_df_uses_stored_dataset(
+    use_polars, setup_airplane_data, setup_airplane_data_polars
+):
+    if use_polars:
+        df, _ = setup_airplane_data_polars
+        freq = "1mo"
+        col_kwargs = dict(id_col="uid", time_col="time", target_col="target")
+        assert_frame_equal = polars.testing.assert_frame_equal
+    else:
+        df, _ = setup_airplane_data
+        freq = "M"
+        col_kwargs = {}
+        assert_frame_equal = pd.testing.assert_frame_equal
+
+    models = [NHITS(h=12, input_size=24, max_steps=2, random_seed=0)]
+    nf = NeuralForecast(models=models, freq=freq)
+    nf.fit(df, **col_kwargs)
+    # use_init_models resets to the same seeded weights before each run
+    cv_with_df = nf.cross_validation(df, use_init_models=True, **col_kwargs)
+    cv_no_df = nf.cross_validation(use_init_models=True, **col_kwargs)
+    assert_frame_equal(cv_no_df, cv_with_df)
+
+
 # test cross_validation with refit
 def test_neural_forecast_refit(setup_airplane_data):
     AirPassengersPanel_train, _ = setup_airplane_data