Error on predict() using loaded model from MLFlow

[kusumy]

Hi, i have question about ML Flow in sktime. I created model as below:

pipe = TransformedTargetForecaster(steps=[
        ("imputer", Imputer(method="mean")),
        ("minmax", TabularToSeriesAdaptor(MinMaxScaler((1, 10)))),
        ("scaler", TabularToSeriesAdaptor(RobustScaler())),
        ("forecaster", make_reduction(svm.SVR(), window_length=lag, strategy="recursive")),   # Initially window_length=24
    ])

cv = SingleWindowSplitter(fh=fh_int)`

gscv = ForecastingGridSearchCV(
        forecaster=pipe,
        param_grid=param_grid,
        cv=cv,
        n_jobs=-1,
        verbose=1,
        scoring=MeanAbsolutePercentageError(symmetric=True),  # set custom scoring function
        error_score='raise',
        refit=True
    )
    
# Make ML Model
gscv.fit(train)

Then I saved the model to MLFlow :

# Log and register model to MLFlow
model_path = "model_dir"  
model_name = "engine_pcd"

mlflow_sktime.log_model(sktime_model=gscv, 
                            artifact_path=model_path,
                            registered_model_name=model_name)

The model was successfully logged and registered to MLFlow. Then I made the model at "Production" stage.
After that I loaded that model using code below:

stage = "Production"
client = mlflow.tracking.MlflowClient(mlflow.get_tracking_uri())
model_info = client.get_latest_versions(model_name, stages=['Production'])
run_id = model_info[0].run_id
run_uri = "runs:/{}/model_dir".format(run_id)
loaded_model = mlflow_sktime.load_model(run_uri)

The model was successfully loaded. Then I called predict on the loaded model

# Create future dataframe
date_range = pd.date_range(start=startDate, end=endDate, freq='H')  # Generate new dates with hourly frequency

# Create a new DataFrame with the generated date index
future_df = pd.DataFrame(index=date_range)

# Predict the future
fh = ForecastingHorizon(future_df.index, is_relative=False)
y_pred = loaded_model.predict(fh) 

There was an error when calling predict. The error said :

Exception has occurred: NotImplementedError
Generating in-sample predictions is not yet implemented for RecursiveTabularRegressionForecaster.
  File "forecast.py", line 262, in <module>
    y_pred = loaded_model.predict(fh)
NotImplementedError: Generating in-sample predictions is not yet implemented for RecursiveTabularRegressionForecaster.

Does this mean that to be able to log the model to MLFlow, we can not use any scikit-learn model (ie, SVR. Random forest, Linear Regression)?

I would greatly appreciate any help.

Regards

[fkiraly]

@kusumy, afaik this has nothing to do with mlflow, but is an issue with the reducer, this is a known limitation.
See here: #3224, also for the current rework.