Training MLflow models

Just like any Python packages, you can use MLflow and the related frameworks in any DSS recipe, notebook, scenario, …

This allows you to train and save a MLflow model directly in DSS, which you can then import as a saved model in order to leverage visual scoring, evaluation, drift analysis, …

Dataiku also features an integration of MLflow Experiment Tracking. When leveraging it, trained models are automatically stored in a configurable managed folder (see Deploying MLflow models).

Training a model

You can train the MLflow model either outside of DSS, in a python recipe, in a python notebook, or using Experiment Tracking

The list of frameworks supported by MLflow is available in the MLflow documentation. These include the most common libraries such as PyTorch, TensorFlow, Scikit-learn, etc.

Saving the MLflow model

You need to export your model in a standard format, provided by MLflow Models, compatible with DSS.

MLflow provides a save_model function for each supported machine learning framework.

For instance, saving a Keras model using MLflow in a model_directory will look like this:

... ommitted Keras model training code

import mlflow
mlflow.keras.save_model(model, model_directory)

You can then import the exported model in DSS as a Saved Model

Python recipe

The following snippet is a draft of a python recipe:

  • taking a train and an evaluation dataset as inputs

  • training a model

  • saving it in MLflow format

  • adding it as a new version to the saved model defined as output

import os
import shutil
import dataiku

from dataiku import recipe

client = dataiku.api_client()

project = client.get_project('PROJECT_ID')

# get train dataset
train_dataset = recipe.get_inputs_as_datasets()[0]
evaluation_dataset = recipe.get_inputs_as_datasets()[1]

# get output saved model
sm = project.get_saved_model(recipe.get_output_names()[0])

# get train dataset as a pandas dataframe
df = train_dataset.get_dataframe()

# get the path of a local managed folder where to temporarily save the trained model
mf = dataiku.Folder("local_managed_folder")
path = mf.get_path()

model_subdir = "my_subdir"
model_dir = os.path.join(path, model_subdir)

if os.path.exists(model_dir):

    # ...train your model...

    # it with package specific MLflow method (here, SKlearn)...
    mlflow.sklearn.save_model(my_model, model_dir)

    # import the model, creating a new version
    mlflow_version = sm.import_mlflow_version_from_managed_folder("version_name", "local_managed_folder", model_subdir, "code-env-with-mlflow-name")

# setting metadata (target name, classes,...)
mlflow_version.set_core_metadata(target_column, ["class0", "class1",...] ,

# evaluate the performance of this new version, to populate the performance screens of the saved model version in DSS


Experiment Tracking features logging of models in a configurable, and not necessarily local, managed folder.


local_managed_folder should be a filesystem managed folder, on the DSS host, as we use the dataiku.Folder.get_path() method to retrieve its path on the local filesystem then compute a directory path where the ML package can save the trained model.


As this recipe uses a local managed folder, it should not be executed in a container.


The 4th parameter of the dataikuapi.dss.savedmodel.DSSSavedModel.import_mlflow_version_from_managed_folder() is the name of the code environment to use when scoring the model. If not specified, the code environment of the project will be resolved and used.

This code environment must contain the mlflow package and the packages of the machine learning library of your choice.


A “Run checks” scenario step must be used to run the checks defined for the saved model on the metrics evaluated on the new version.


Recent versions of MLflow feature an `mlflow.evaluate` function. This function is different from dataikuapi.dss.savedmodel.MLFlowVersionHandler.evaluate(). Only the later will populate the interpretation screens of a saved model version in DSS.