AttributeError when logging a model with MLFlow

[ABerry057]

What happened + What you expected to happen

Description
When attempting to log a trained MLForecast model using MLFlow.flavor module, I encounter an AttributeError. It appears the MLForecast object somehow loses the models_ attribute in the process of performing cross validation; if I omit the cross validation step, there is no bug.

Expected behavior
Calling mlforecast.flavor.log_model() should successfully save the MLForecast model artifact after performing cross validation.

Logs

Traceback (most recent call last):
  File "/Users/ab/Desktop/mlflow-demo/src/train.py", line 63, in <module>
    mlforecast.flavor.log_model(model=fcst_model, artifact_path="model")
  File "/Users/ab/Desktop/mlflow-demo/.venv/lib/python3.11/site-packages/mlforecast/flavor.py", line 263, in log_model
    return Model.log(
           ^^^^^^^^^^
  File "/Users/ab/Desktop/mlflow-demo/.venv/lib/python3.11/site-packages/mlflow/models/model.py", line 796, in log
    flavor.save_model(path=local_path, mlflow_model=mlflow_model, **kwargs)
  File "/Users/ab/Desktop/mlflow-demo/.venv/lib/python3.11/site-packages/mlforecast/flavor.py", line 142, in save_model
    model.save(model_data_path)
  File "/Users/ab/Desktop/mlflow-demo/.venv/lib/python3.11/site-packages/mlforecast/forecast.py", line 1009, in save
    cloudpickle.dump(self.models_, f)
                     ^^^^^^^^^^^^
AttributeError: 'MLForecast' object has no attribute 'models_'. Did you mean: 'models'?

Versions / Dependencies

• Python 3.11.4
• mlforecast==1.0.1
• mlflow==2.20.1
• macOS 15.3

Reproduction script

import pandas as pd
import mlflow
from sklearn.linear_model import LinearRegression
from utilsforecast.losses import rmse, smape
from utilsforecast.evaluation import evaluate

import mlforecast.flavor
from mlforecast import MLForecast


data_url = "https://vincentarelbundock.github.io/Rdatasets/csv/AER/FrozenJuice.csv"
freq = 1
h = 2
data_series = (
    pd.read_csv(data_url, usecols=["rownames", "price"])
    .rename({"rownames": "ds", "price": "y"}, axis=1)
    .assign(unique_id="test")
)

# start MLFlow tracking
with mlflow.start_run():
    fcst_object = MLForecast(
        models=[LinearRegression()],
        freq=freq,
        lags=list(range(1, 5))
    )
    fcst_object.fit(data_series)
    if hasattr(fcst_object, "models_"):
        print("MLForecast object has models_ attribute")
    cv_metrics_df = fcst_object.cross_validation(
        df=data_series, h=h,
        n_windows=10
    )
    eval_result_dict = evaluate(
        cv_metrics_df.drop("cutoff", axis=1),
        metrics=[rmse, smape],
        agg_fn="mean"
    ).set_index("metric").squeeze().to_dict()
    mlflow.log_metrics(eval_result_dict)
    # the following print statement does not get executed
    if hasattr(fcst_object, "models_"):
        print("MLForecast object has models_ attribute")
    mlforecast.flavor.log_model(model=fcst_object, artifact_path="model")

Issue Severity

Medium: It is a significant difficulty but I can work around it.

[jmoralez]

You should perform cross validation first and then fit.

[ABerry057]

@jmoralez That worked, thank you for the quick suggestion!