mlflow.keras

The mlflow.keras module provides an API for logging and loading Keras models. This module exports Keras models with the following flavors:

Keras (native) format

This is the main flavor that can be loaded back into Keras.

mlflow.pyfunc

Produced for use by generic pyfunc-based deployment tools and batch inference.

mlflow.keras.autolog()[source]

Note

Experimental: This method may change or be removed in a future release without warning.

Enables automatic logging from Keras to MLflow. Autologging captures the following information:

Metrics and Parameters
  • Training loss; validation loss; user-specified metrics

  • Metrics associated with the EarlyStopping callbacks: stopped_epoch, restored_epoch, restore_best_weight, last_epoch, etc

  • fit() or fit_generator() parameters; optimizer name; learning rate; epsilon

  • fit() or fit_generator() parameters associated with EarlyStopping: min_delta, patience, baseline, restore_best_weights, etc

Artifacts
  • Model summary on training start

  • MLflow Model (Keras model) on training end

Example
import mlflow
import mlflow.keras
# Build, compile, enable autologging, and train your model
keras_model = ...
keras_model.compile(optimizer="rmsprop", loss="mse", metrics=["accuracy"])
# autolog your metrics, parameters, and model
mlflow.keras.autolog()
results = keras_model.fit(
    x_train, y_train, epochs=20, batch_size=128, validation_data=(x_val, y_val))

EarlyStopping Integration with Keras AutoLogging

MLflow will detect if an EarlyStopping callback is used in a fit() or fit_generator() call, and if the restore_best_weights parameter is set to be True, then MLflow will log the metrics associated with the restored model as a final, extra step. The epoch of the restored model will also be logged as the metric restored_epoch. This allows for easy comparison between the actual metrics of the restored model and the metrics of other models.

If restore_best_weights is set to be False, then MLflow will not log an additional step.

Regardless of restore_best_weights, MLflow will also log stopped_epoch, which indicates the epoch at which training stopped due to early stopping.

If training does not end due to early stopping, then stopped_epoch will be logged as 0.

MLflow will also log the parameters of the EarlyStopping callback, excluding mode and verbose.

mlflow.keras.get_default_conda_env(include_cloudpickle=False, keras_module=None)[source]
Returns

The default Conda environment for MLflow Models produced by calls to save_model() and log_model().

mlflow.keras.load_model(model_uri, **kwargs)[source]

Load a Keras model from a local file or a run.

Extra arguments are passed through to keras.load_model.

Parameters

model_uri

The location, in URI format, of the MLflow model. For example:

  • /Users/me/path/to/local/model

  • relative/path/to/local/model

  • s3://my_bucket/path/to/model

  • runs:/<mlflow_run_id>/run-relative/path/to/model

  • models:/<model_name>/<model_version>

  • models:/<model_name>/<stage>

For more information about supported URI schemes, see Referencing Artifacts.

Returns

A Keras model instance.

Example
# Load persisted model as a Keras model or as a PyFunc, call predict() on a pandas DataFrame
keras_model = mlflow.keras.load_model("runs:/96771d893a5e46159d9f3b49bf9013e2" + "/models")
predictions = keras_model.predict(x_test)
mlflow.keras.log_model(keras_model, artifact_path, conda_env=None, custom_objects=None, keras_module=None, registered_model_name=None, signature: mlflow.models.signature.ModelSignature = None, input_example: Union[pandas.core.frame.DataFrame, numpy.ndarray, dict, list] = None, **kwargs)[source]

Log a Keras model as an MLflow artifact for the current run.

Parameters
  • keras_model – Keras model to be saved.

  • artifact_path – Run-relative artifact path.

  • conda_env

    Either a dictionary representation of a Conda environment or the path to a Conda environment yaml file. If provided, this describes the environment this model should be run in. At minimum, it should specify the dependencies contained in get_default_conda_env(). If None, the default mlflow.keras.get_default_conda_env() environment is added to the model. The following is an example dictionary representation of a Conda environment:

    {
        'name': 'mlflow-env',
        'channels': ['defaults'],
        'dependencies': [
            'python=3.7.0',
            'keras=2.2.4',
            'tensorflow=1.8.0'
        ]
    }
    

  • custom_objects – A Keras custom_objects dictionary mapping names (strings) to custom classes or functions associated with the Keras model. MLflow saves these custom layers using CloudPickle and restores them automatically when the model is loaded with mlflow.keras.load_model() and mlflow.pyfunc.load_model().

  • keras_module – Keras module to be used to save / load the model (keras or tf.keras). If not provided, MLflow will attempt to infer the Keras module based on the given model.

  • registered_model_name – (Experimental) If given, create a model version under registered_model_name, also creating a registered model if one with the given name does not exist.

  • signature

    (Experimental) ModelSignature describes model input and output Schema. The model signature can be inferred from datasets with valid model input (e.g. the training dataset with target column omitted) and valid model output (e.g. model predictions generated on the training dataset), for example:

    from mlflow.models.signature import infer_signature
    train = df.drop_column("target_label")
    predictions = ... # compute model predictions
    signature = infer_signature(train, predictions)
    

  • input_example – (Experimental) Input example provides one or several instances of valid model input. The example can be used as a hint of what data to feed the model. The given example will be converted to a Pandas DataFrame and then serialized to json using the Pandas split-oriented format. Bytes are base64-encoded.

  • kwargs – kwargs to pass to keras_model.save method.

Example
from keras import Dense, layers
import mlflow
# Build, compile, and train your model
keras_model = ...
keras_model.compile(optimizer="rmsprop", loss="mse", metrics=["accuracy"])
results = keras_model.fit(
    x_train, y_train, epochs=20, batch_size = 128, validation_data=(x_val, y_val))
# Log metrics and log the model
with mlflow.start_run() as run:
    mlflow.keras.log_model(keras_model, "models")
mlflow.keras.save_model(keras_model, path, conda_env=None, mlflow_model=None, custom_objects=None, keras_module=None, signature: mlflow.models.signature.ModelSignature = None, input_example: Union[pandas.core.frame.DataFrame, numpy.ndarray, dict, list] = None, **kwargs)[source]

Save a Keras model to a path on the local file system.

Parameters
  • keras_model – Keras model to be saved.

  • path – Local path where the model is to be saved.

  • conda_env

    Either a dictionary representation of a Conda environment or the path to a Conda environment yaml file. If provided, this decsribes the environment this model should be run in. At minimum, it should specify the dependencies contained in get_default_conda_env(). If None, the default get_default_conda_env() environment is added to the model. The following is an example dictionary representation of a Conda environment:

    {
        'name': 'mlflow-env',
        'channels': ['defaults'],
        'dependencies': [
            'python=3.7.0',
            'keras=2.2.4',
            'tensorflow=1.8.0'
        ]
    }
    

  • mlflow_model – MLflow model config this flavor is being added to.

  • custom_objects – A Keras custom_objects dictionary mapping names (strings) to custom classes or functions associated with the Keras model. MLflow saves these custom layers using CloudPickle and restores them automatically when the model is loaded with mlflow.keras.load_model() and mlflow.pyfunc.load_model().

  • keras_module – Keras module to be used to save / load the model (keras or tf.keras). If not provided, MLflow will attempt to infer the Keras module based on the given model.

  • kwargs – kwargs to pass to keras_model.save method.

  • signature

    (Experimental) ModelSignature describes model input and output Schema. The model signature can be inferred from datasets with valid model input (e.g. the training dataset with target column omitted) and valid model output (e.g. model predictions generated on the training dataset), for example:

    from mlflow.models.signature import infer_signature
    train = df.drop_column("target_label")
    predictions = ... # compute model predictions
    signature = infer_signature(train, predictions)
    

  • input_example – (Experimental) Input example provides one or several instances of valid model input. The example can be used as a hint of what data to feed the model. The given example will be converted to a Pandas DataFrame and then serialized to json using the Pandas split-oriented format. Bytes are base64-encoded.

Example
import mlflow
# Build, compile, and train your model
keras_model = ...
keras_model_path = ...
keras_model.compile(optimizer="rmsprop", loss="mse", metrics=["accuracy"])
results = keras_model.fit(
    x_train, y_train, epochs=20, batch_size = 128, validation_data=(x_val, y_val))
# Save the model as an MLflow Model
mlflow.keras.save_model(keras_model, keras_model_path)