mlflow.tensorflow

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

TensorFlow (native) format

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

mlflow.pyfunc

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

mlflow.tensorflow.autolog(every_n_iter=100)[source]

Note

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

Enables automatic logging from TensorFlow to MLflow. Note that autologging for tf.keras is handled by mlflow.tensorflow.autolog(), not mlflow.keras.autolog(). As an example, try running the TensorFlow examples.

For each TensorFlow module, autologging captures the following information:

tf.keras
  • Metrics and Parameters

  • Training loss; validation loss; user-specified metrics

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

  • Artifacts

  • Model summary on training start

  • MLflow Model (Keras model)

  • TensorBoard logs on training end

tf.keras.callbacks.EarlyStopping
  • Metrics and Parameters

  • Metrics from the EarlyStopping callbacks: stopped_epoch, restored_epoch, restore_best_weight, etc

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

tf.estimator
  • Metrics and Parameters

  • TensorBoard metrics: average_loss, loss, etc

  • Parameters steps and max_steps

  • Artifacts

  • MLflow Model (TF saved model) on call to tf.estimator.export_saved_model

TensorFlow Core
  • Metrics

  • All tf.summary.scalar calls

Refer to the autologging tracking documentation for more information on TensorFlow workflows.

Parameters

every_n_iter – The frequency with which metrics should be logged. Defaults to 100. Ex: a value of 100 will log metrics at step 0, 100, 200, etc.

mlflow.tensorflow.get_default_conda_env()[source]
Returns

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

mlflow.tensorflow.load_model(model_uri, tf_sess=None)[source]

Load an MLflow model that contains the TensorFlow flavor from the specified path.

With TensorFlow version <2.0.0, this method must be called within a TensorFlow graph context.

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.

  • tf_sess – The TensorFlow session in which to load the model. If using TensorFlow version >= 2.0.0, this argument is ignored. If using TensorFlow <2.0.0, if no session is passed to this function, MLflow will attempt to load the model using the default TensorFlow session. If no default session is available, then the function raises an exception.

Returns

For TensorFlow < 2.0.0, a TensorFlow signature definition of type: tensorflow.core.protobuf.meta_graph_pb2.SignatureDef. This defines the input and output tensors for model inference. For TensorFlow >= 2.0.0, A callable graph (tf.function) that takes inputs and returns inferences.

Example
import mlflow.tensorflow
import tensorflow as tf
tf_graph = tf.Graph()
tf_sess = tf.Session(graph=tf_graph)
with tf_graph.as_default():
    signature_definition = mlflow.tensorflow.load_model(model_uri="model_uri",
                            tf_sess=tf_sess)
    input_tensors = [tf_graph.get_tensor_by_name(input_signature.name)
                        for _, input_signature in signature_definition.inputs.items()]
    output_tensors = [tf_graph.get_tensor_by_name(output_signature.name)
                        for _, output_signature in signature_definition.outputs.items()]
mlflow.tensorflow.log_model(tf_saved_model_dir, tf_meta_graph_tags, tf_signature_def_key, artifact_path, conda_env=None, signature: mlflow.models.signature.ModelSignature = None, input_example: Union[pandas.core.frame.DataFrame, numpy.ndarray, dict, list] = None, registered_model_name=None)[source]

Note

This method requires all argument be specified by keyword.

Log a serialized collection of TensorFlow graphs and variables as an MLflow model for the current run. This method operates on TensorFlow variables and graphs that have been serialized in TensorFlow’s SavedModel format. For more information about SavedModel format, see the TensorFlow documentation: https://www.tensorflow.org/guide/saved_model#save_and_restore_models.

This method saves a model with both python_function and tensorflow flavors. If loaded back using the python_function flavor, the model can be used to predict on pandas DataFrames, producing a pandas DataFrame whose output columns correspond to the TensorFlow model’s outputs. The python_function model will flatten outputs that are length-one, one-dimensional tensors of a single scalar value (e.g. {"predictions": [[1.0], [2.0], [3.0]]}) into the scalar values (e.g. {"predictions": [1, 2, 3]}), so that the resulting output column is a column of scalars rather than lists of length one. All other model output types are included as-is in the output DataFrame.

Parameters
  • tf_saved_model_dir – Path to the directory containing serialized TensorFlow variables and graphs in SavedModel format.

  • tf_meta_graph_tags – A list of tags identifying the model’s metagraph within the serialized SavedModel object. For more information, see the tags parameter of the tf.saved_model.builder.SavedModelBuilder method.

  • tf_signature_def_key – A string identifying the input/output signature associated with the model. This is a key within the serialized SavedModel signature definition mapping. For more information, see the signature_def_map parameter of the tf.saved_model.builder.SavedModelBuilder method.

  • artifact_path – The run-relative path to which to log model artifacts.

  • 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',
            'tensorflow=1.8.0'
        ]
    }
    

  • 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.

    param 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.

mlflow.tensorflow.save_model(tf_saved_model_dir, tf_meta_graph_tags, tf_signature_def_key, path, mlflow_model=None, conda_env=None, signature: mlflow.models.signature.ModelSignature = None, input_example: Union[pandas.core.frame.DataFrame, numpy.ndarray, dict, list] = None)[source]

Note

This method requires all argument be specified by keyword.

Save a serialized collection of TensorFlow graphs and variables as an MLflow model to a local path. This method operates on TensorFlow variables and graphs that have been serialized in TensorFlow’s SavedModel format. For more information about SavedModel format, see the TensorFlow documentation: https://www.tensorflow.org/guide/saved_model#save_and_restore_models.

Parameters
  • tf_saved_model_dir – Path to the directory containing serialized TensorFlow variables and graphs in SavedModel format.

  • tf_meta_graph_tags – A list of tags identifying the model’s metagraph within the serialized SavedModel object. For more information, see the tags parameter of the tf.saved_model.builder.savedmodelbuilder method.

  • tf_signature_def_key – A string identifying the input/output signature associated with the model. This is a key within the serialized savedmodel signature definition mapping. For more information, see the signature_def_map parameter of the tf.saved_model.builder.savedmodelbuilder method.

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

  • mlflow_model – MLflow model configuration to which to add the tensorflow flavor.

  • 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',
            'tensorflow=1.8.0'
        ]
    }
    

  • 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.