Retrain Forecast Models (#1282)

Author: prasankh

ads/opctl/operator/lowcode/common/utils.py

@@ -218,17 +218,12 @@ def _write_file(local_filename, remote_filename, storage_options, **kwargs):
 
 
 def load_pkl(filepath):
-    return _safe_write(fn=_load_pkl, filepath=filepath)
-
-
-def _load_pkl(filepath):
     storage_options = {}
     if ObjectStorageDetails.is_oci_path(filepath):
         storage_options = default_signer()
 
     with fsspec.open(filepath, "rb", **storage_options) as f:
         return cloudpickle.load(f)
-    return None
 
 
 def write_pkl(obj, filename, output_dir, storage_options):

ads/opctl/operator/lowcode/forecast/model/arima.py

@@ -85,7 +85,11 @@ def _train_model(self, i, s_id, df, model_kwargs):
             X_pred = self.get_horizon(data).drop(target, axis=1)
 
             if self.loaded_models is not None and s_id in self.loaded_models:
-                model = self.loaded_models[s_id]
+                model = self.loaded_models[s_id]["model"]
+                order = model.order
+                seasonal_order = model.seasonal_order
+                model = pm.ARIMA(order=order, seasonal_order=seasonal_order)
+                model.fit(y=y, X=X_in)
             else:
                 # Build and fit model
                 model = pm.auto_arima(y=y, X=X_in, **model_kwargs)

ads/opctl/operator/lowcode/forecast/model/automlx.py

@@ -142,17 +142,20 @@ def _build_model(self) -> pd.DataFrame:
                 )
 
                 if self.loaded_models is not None and s_id in self.loaded_models:
-                    model = self.loaded_models[s_id]
-                else:
-                    model = Pipeline(
-                        task="forecasting",
-                        **model_kwargs,
-                    )
-                    model.fit(
-                        X=data_i.drop(target, axis=1),
-                        y=data_i[[target]],
-                        time_budget=time_budget,
-                    )
+                    model = self.loaded_models[s_id]["model"]
+                    model_kwargs["model_list"] = [model.selected_model_]
+                    model_kwargs["search_space"]={}
+                    model_kwargs["search_space"][model.selected_model_] = model.selected_model_params_
+
+                model = Pipeline(
+                    task="forecasting",
+                    **model_kwargs,
+                )
+                model.fit(
+                    X=data_i.drop(target, axis=1),
+                    y=data_i[[target]],
+                    time_budget=time_budget,
+                )
                 logger.debug(f"Selected model: {model.selected_model_}")
                 logger.debug(f"Selected model params: {model.selected_model_params_}")
                 summary_frame = model.forecast(

ads/opctl/operator/lowcode/forecast/model/prophet.py

@@ -9,6 +9,7 @@
 import matplotlib as mpl
 import numpy as np
 import optuna
+import inspect
 import pandas as pd
 from joblib import Parallel, delayed
 
@@ -39,6 +40,22 @@ def _add_unit(num, unit):
     return f"{num} {unit}"
 
 
+def _extract_parameter(model):
+    """
+    extract Prophet initialization parameters
+    """
+    from prophet import Prophet
+    sig = inspect.signature(Prophet.__init__)
+    param_names = list(sig.parameters.keys())
+    params = {}
+    for name in param_names:
+        if hasattr(model, name):
+            value = getattr(model, name)
+            if isinstance(value, (int, float, str, bool, type(None), dict, list)):
+                params[name] = value
+    return params
+
+
 def _fit_model(data, params, additional_regressors):
     from prophet import Prophet
 
@@ -96,16 +113,17 @@ def _train_model(self, i, series_id, df, model_kwargs):
             data = self.preprocess(df, series_id)
             data_i = self.drop_horizon(data)
             if self.loaded_models is not None and series_id in self.loaded_models:
-                model = self.loaded_models[series_id]
+                previous_model = self.loaded_models[series_id]["model"]
+                model_kwargs.update(_extract_parameter(previous_model))
             else:
                 if self.perform_tuning:
                     model_kwargs = self.run_tuning(data_i, model_kwargs)
 
-                model = _fit_model(
-                    data=data,
-                    params=model_kwargs,
-                    additional_regressors=self.additional_regressors,
-                )
+            model = _fit_model(
+                data=data,
+                params=model_kwargs,
+                additional_regressors=self.additional_regressors,
+            )
 
             # Get future df for prediction
             future = data.drop("y", axis=1)