Exporting models¶

DSS provides multiple capabilities for exporting models.

Deploy the model on Dataiku API nodes for real-time scoring
Export to Python
Export as a MLflow model
Export a Java class/JAR for a model
- Usage
- Limitations
Export a PMML file for a model
- Limitations
Export to Jupyter notebook

Deploy the model on Dataiku API nodes for real-time scoring ¶

See API Node & API Deployer: Real-time APIs, and its subsections Exposing a Visual Model and Exposing a MLflow model.

Export to Python ¶

For use cases where the API node may not be feasible (such as edge deployment, or very-low-latency deployment), it is possible to export the model to a zip file that can then be used to score the model in any Python code, fully outside of DSS.

This scoring uses the dataiku-scoring open source Python package.

Note that this feature is not available in all Dataiku licenses. You may need to reach out to your Dataiku Account Manager or Customer Success Manager.

Go to the trained model you wish to export (either a model trained in the Lab or a version of a saved model deployed in the Flow)
Click the Actions button on the top-right corner and select Export model as …
Select Python to download the export file

Requirements ¶

The model needs to be compatible with either Local (Optimized) scoring or be an MLflow imported model to be compatible with Python export.

The dataiku-scoring package is available on Pypi. It is compatible with Python 3.6 and above, and depends on NumPy.

To ensure to install the correct version, you can unzip the exported model file and run:

pip install -r requirements.txt

You can optionally use Pandas for easier interaction with the model. All Pandas versions above 0.23 are supported.

Usage ¶

The package exposes the load_model function, which allows you to load the model’s zip file as a python object. The loaded model will have a predict method and a predict_proba method for classification problems. Here are the signatures of these methods:

predict(
  data: [pandas.DataFrame, numpy.ndarray, List[Dict | List]]
) -> numpy.array

# Only exists for classification problems
predict_proba(
  data: [pandas.DataFrame, numpy.ndarray, List[Dict | List]]
) -> numpy.darray

Example ¶

We train a model in DSS on the iris dataset and download it using Python export. The iris dataset is a multiclass classification problem where the output classes are “Setosa,” “Virginica,” and “Versicolor.” Here is an example of using the downloaded zip file together with dataiku-scoring to score the first two rows of the iris dataset.

import dataikuscoring

model = dataikuscoring.load_model("path/to/model.zip")

data = [
    {'sepal.length': 5.1, 'sepal.width': 3.5, 'petal.length': 1.4, 'petal.width': 0.2},
    {'sepal.length': 4.9, 'sepal.width': 3.0, 'petal.length': 1.4, 'petal.width': 0.2}
]

model.predict(data)  # returns array(['Setosa', 'Setosa'], dtype=object)

Provided you have pandas in your environment, you can also directly use a pandas.DataFrame:

import pandas as pd

df = pd.DataFrame(data)
model.predict(df)

Or a List of List / NumPy array:

data = [[5.1, 3.5, 1.4, 0.2], [4.9, 3.0, 1.4, 0.2]]
model.predict(data)

import numpy as np
model.predict(np.array(data))

Note

A warning is displayed when you use a List of List or a NumPy array as input to the model.predict method. In these cases, the column names are not provided, so features are implicitly assumed to be in the same order as in the training dataset in DSS.

Since the iris dataset is a classification problem, we can also output probabilities:

model.predict_proba(data)

will output:

{'Virginica' : array([0.15511770783800324, 0.1360890057639774], dtype=object),
 'Versicolor': array([0.06850569755129679, 0.13872351067755467], dtype=object),
 'Setosa' : array([0.7763765946107, 0.725187483558468]}, dtype=object)}

Limitations ¶

The Python export feature does not support preparation scripts. In your Lab analysis where you trained your model, if the Script tab has steps, those steps are not included in the exported model. If your model has a preparation script, you must prepare the data yourself before scoring with the loaded python model. The expected input of the model is the output of your preparation script.

Export as a MLflow model ¶

MLflow is an open-source platform to manage machine learning models lifecycle. If your machine learning deployment pipelines already uses MLflow, you may want to use the MLflow export to benefit of DSS Visual ML tool.

You can export a zip file containing a MLflow model representing the trained in the DSS Visual ML tool. This can be used in any MLflow-compatible scoring system that supports the “python_function” flavor of MLflow

Note that this feature is not available in all Dataiku licenses. You may need to reach out to your Dataiku Account Manager or Customer Success Manager.

Warning

You will need to install the dataiku-scoring python package (see dataiku-scoring installation) in any python environment where you want to load and use the MLflow export of your model.

Note

The dataiku-scoring package defines a specific dss flavor for MLflow that enables using a model designed and exported with DSS with MLflow. You can find the documentation here if you are unfamiliar with MLflow flavors.

Warning

Tier 2 support: MLflow export is covered by Tier 2 support

Go to the trained model you wish to export (either a model trained in the Lab or a version of a saved model deployed in the Flow)
Click the Actions button on the top-right corner and select Export model as …
Select MLFlow to download the zip file
Unzip the zip file to use it with MLflow

Alternatively, you can export using the API with the following method dataikuapi.dss.ml.DSSTrainedPredictionModelDetails.get_scoring_MLflow().

Requirements ¶

The model needs to be compatible with either Local (Optimized) scoring or be an MLflow imported model to be compatible with Python export.

Like all MLflow models, the Dataiku-exported MLflow model contains its requirements.

You will need MLflow in version 1.21 at least.

Usage ¶

To load and use your model you can use the generic MLflow python_function flavor following:

import MLflow

model = MLflow.pyfunc.load_model("path/to/model_unzipped")

model.predict(input_data)

The model accepts the same input data as the one described in the Python export in Usage.

For a more extensive description on how to use an MLflow model you can refer to the official MLflow documentation.

If your model is a classification model, you can access the probabilities using model._model_impl.dss_model.predict_proba

“dss” MLflow flavor ¶

MLflow flavors are adapters designed for using various machine learning frameworks with a unified API.

In addition to the python_function flavor, MLflow models exported from Dataiku contain a dss flavor which is a simple wrapper on top of the dataiku-scoring package described in Export to Python, which allows you to use the dataiku-scoring API more directly.

The following describes the DSS flavor in the same way as other flavors in the official MLflow documentation.

The dataikuscoring.mlflow module defines the DSS flavor for MLflow. The dss model flavor enables logging DSS models in MLflow format via the dataikuscoring.mlflow.save_model() and dataikuscoring.mlflow.log_model() methods. These methods also add the standard MLflow python_function flavor to the MLflow Models that they produce, allowing the models to be interpreted as generic Python function for inference via MLflow.pyfunc.load_model(). This loaded PyFunc model can only score DataFrame input. You can also use the dataikuscoring.mlflow.load_model() method to load MLflow Models with the DSS Default format. You obtain a model similar to what is described in Export to Python.

Note

When exporting a Model using the MLflow format in DSS, the resulting zip file contains a zipped version of the output of the dataikuscoring.mlflow.save_model() method.

You can load your model using the dss flavor by doing model = dataikuscoring.mlflow.load_model("path/to/model_unzipped"). The loaded model is equivalent to the one loaded with the default Python export format and described in Usage.

Logging the model to an MLflow backend ¶

The model exported in MLflow format can later be imported into an MLflow Model registry backend or an MLflow Tracking backend.

The following example demonstrates how to import a model in a local MLflow backend using a local SQLite database and a DSS model exported in MLflow format:

import dataikuscoring.mlflow
import MLflow

model = MLflow.pyfunc.load_model("path/to/model_unzipped")
# alternatively:
# model = dataikuscoring.mlflow.load_model("path/to/model_unzipped")

MLflow.set_tracking_uri("sqlite:///MLflow.db")
MLflow.set_registry_uri("sqlite:///MLflow.db")
with MLflow.start_run() as run:
    dataikuscoring.mlflow.log_model(model, artifact_path)

You can then run MLflow ui --port 5001 --backend-store-uri sqlite:///MLflow.db in a shell to access the run and the model it contains, and deploy it to the model registry.

The model was loaded using MLflow, but its original flavor is the dss flavor. Therefore, we need to use the original dss flavor when logging the model to the backend.

Limitations ¶

The MLflow export feature does not support preparation scripts. In your Lab analysis, where you trained your model, if the Script tab has steps, those steps are not included in the exported model. If your model has a preparation script, you must prepare the data yourself before scoring with the loaded python model. The expected input of the model is the output of your preparation script.

Export a Java class/JAR for a model ¶

For use cases where the API node may not be feasible (such as edge deployment, or very-low-latency deployment), it is possible to export the model to a JAR file that can then be used to score the model in any Java code, fully outside of DSS.

Note that this feature is not available in all Dataiku licenses. You may need to reach out to your Dataiku Account Manager or Customer Success Manager.

The model needs to be compatible with Local (Optimized) scoring to be compatible with Java export.

Go to the trained model you wish to export (either a model trained in the Lab or a version of a saved model deployed in the Flow)
Click the Actions button on the top-right corner and select Export model as …
Select Java, indicate the full-qualified class name you want for your model

Add that JAR to the classpath of your Java application.

If you have several models you wish to use on the same JVM, you can export the “thin” JAR for each model, which only contains the class and resources for the model, and not the scoring libraries. In that case, you also need to download the scoring libraries (from the same dropdown menu) and add both JARs to the classpath.

Usage ¶

If you specified the name com.mycompany.myproject.MyModel at export time, you can use it like this once you’ve added the JAR to the classpath:

import com.mycompany.myproject.MyModel;
import com.dataiku.scoring.*;

// ...
MyModel model = new MyModel();
Observation.Builder obsBuilder = model.observationBuilder();
Observation obs = obsBuilder
    .with("myCategoricalFeature", "Some string value")
    .with("myNumericFeature", 42.0d)
    // other .with("featureName", <string or double value>)
    .build();
if (obs.hasError()) {
    System.err.println("Can't build observation: " + obs.getErrorMessage());
    // maybe throw here
}

// For a classification model
Try<ClassificationResult> prediction = model.predict(obs);
if (prediction.isError()) {
    System.err.println("Can't make a prediction: " + prediction.getMessage());
    // maybe throw here
} else {
    ClassificationResult result = prediction.get();
    // predictedClass is one of model.getClassLabels()
    String predictedClass = result.getPrediction();
     // probabilities has the same indices as model.getClassLabels()
     // i.e. 0 to (model.getNumClasses() - 1)
    double[] probabilities = result.getProbabilities();
    // Use result here
}

// For a regression model
Try<RegressionResult> prediction = model.predict(obs);
if (prediction.isError()) {
    System.err.println("Can't make a prediction: " + prediction.getMessage());
    // maybe throw here
} else {
    RegressionResult result = prediction.get();
    double predictedValue = result.getPrediction();
    // Use result here
}

You can find the javadoc for the com.dataiku.scoring package here: https://doc.dataiku.com/dss/api/11/scoring.

If you want to debug your model, you can run the “fat” jar version with -jar:

java -jar /path/to/dataiku-model-my-model-assembly.jar

… or the “thin” jar version, specifying you model class as the Main class to run:

java -cp /path/to/dataiku-model-my-model.jar:/path/to/dataiku-scoring-libs_DSS_VERSION.jar \
    com.mycompany.myproject.MyModel

This command will take JSON objects with feature values on standard input (one per line), and return predictions as JSON objects on standard output (one per line as well). For instance with a classification model trained on the classical Titanic dataset:

$ echo '{"Sex": "male", "Pclass": 3}' >titanic.txt
$ echo '{"Sex": "female", "Pclass": 1}' >>titanic.txt
$ java -jar dataiku-model-survived-on-titanic-assembly.jar <titanic.txt >out.txt
Nov 26, 2018 3:03:39 PM com.dataiku.scoring.pipelines.Normalization <init>
INFO: Normalize columns
Nov 26, 2018 3:03:39 PM com.dataiku.scoring.builders.Build binaryProbabilisticPipeline
INFO: Loaded model:
Nov 26, 2018 3:03:39 PM com.dataiku.scoring.builders.Build binaryProbabilisticPipeline
INFO: com.dataiku.scoring.models.ForestClassifier@3cd1f1c8
Nov 26, 2018 3:03:39 PM com.dataiku.scoring.builders.Build preprocessingPipeline
INFO: Loaded preprocessing pipeline:
Nov 26, 2018 3:03:39 PM com.dataiku.scoring.builders.Build preprocessingPipeline
INFO: PreprocessingPipeline(
    ImputeWithValue(Pclass -> 2.3099579242636747 ; Parch -> 0.364656381486676 ; SibSp -> 0.5105189340813464 ; Age -> 29.78608695652174 ; Fare -> 32.91587110799433 ; )
    Dummifier(Sex in [female, male, ] ; Embarked in [Q, S, C, ])
    Rescaler(Fare (shift, inv_scale)=(32.91587110799433, 0.01887857758669009) ; Age (shift, inv_scale)=(29.78608695652174, 0.07038582694206309) ; Parch (shift, inv_scale)=(0.364656381486676, 1.2618109015803536) ; Pclass (shift, inv_scale)=(2.3099579242636747, 1.2048162082648861) ; SibSp (shift, inv_scale)=(0.5105189340813464, 0.9172989588087348))
)

$cat out.txt
{"value":{"probabilities":{"died":0.6874695011372898,"survived":0.3125304988627102},"prediction":"died"},"isError":false}
{"value":{"probabilities":{"died":0.062296226501392105,"survived":0.9377037734986079},"prediction":"survived"},"isError":false}

com.dataiku.scoring uses java.util.logging for logging. If you wish to forward it to log4j or logback, you can use a SLF4J bridge.

Limitations ¶

The Java export feature does not support preparation scripts. In your Lab analysis where you trained your model, if the Script tab has steps then those steps are not included in the exported model. If your model has a preparation script, you will need to prepare the data yourself before scoring with the JAR. The expected input of the model (the features you add in an ObservationBuilder to build an Observation) is the output of your preparation script.

Export a PMML file for a model ¶

PMML is a XML-based language to define models, and score them using any PMML-compatible scoring system.

You can export a PMML file containing a Dataiku model, which can then be used in any PMML-compatible scoring system.

Note that this feature is not available in all Dataiku licenses. You may need to reach out to your Dataiku Account Manager or Customer Success Manager.

Warning

Tier 2 support: PMML export is covered by Tier 2 support

If your model is compatible with PMML export (see Limitations below):

Go to the trained model you wish to export (either a model trained in the Lab or a version of a saved model deployed in the Flow)
Click the Actions button on the top-right corner and select Export model as …
Select PMML

Limitations ¶

The following preprocessing options are compatible with PMML export:

Numeric features with regular handling
Categorical features with impact-coding or Dummy-encoding handling
“Impute with” and “Treat as regular” options when handling missing values. (“Drop rows” option is not compatible)
No Vector, Image, Binary or Text features
No feature generation (numerical derivatives, combination…)
No dimensionality reduction

The following algorithms are compatible with PMML export:

Logistic Regression
Extra trees
Linear Regression
Decision Tree
Random Forest
Gradient tree boosting, excluding XGBoost

The PMML export feature does not support preparation scripts. In your Lab analysis where you trained your model, if the Script tab has steps then those steps are not included in the exported model. If your model has a preparation script, you will need to prepare the data yourself before scoring. The expected input of the model is the output of your preparation script.

The PMML export outputs probabilities and will use the optimized threshold for binary classification.

The PMML export feature does not support probability calibration.

Computer vision models are not compatible with PMML export.

Export to Jupyter notebook ¶

Note

This only applies to models trained using the “In-memory (Python)” engine, both for prediction and clustering. Not all algorithms are supported by this feature.

Once a model has been trained, you can export it as a Jupyter notebook.

DSS will automatically generate a Jupyter (Python) notebook with code to reproduce a model similar to the model that you trained.

To generate a Jupyter notebook:

Go to the trained model you wish to export
Click the dropdown icon next to the “Deploy” button
Select “Export to Jupyter notebook”

Warning

This generated notebook is for educational and explanatory purposes only. In particular, this notebook does not reproduce all preprocessing capabilities of DSS, and is only a best-effort approximation of the model trained in DSS.

To use the exact model trained by DSS, deploy it to the Flow and use the API node, scoring recipes, or any of the export methods described above