mlflow.models

The mlflow.models module provides an API for saving machine learning models in “flavors” that can be understood by different downstream tools.

The built-in flavors are:

For details, see MLflow Models.

class mlflow.models.EvaluationArtifact(uri, content=None)[source]

Bases: object

A model evaluation artifact containing an artifact uri and content.

property content: The content of the artifact (representation varies)

property uri: The URI of the artifact

class mlflow.models.EvaluationMetric(eval_fn, name, greater_is_better, long_name=None, version=None, metric_details=None, metric_metadata=None, genai_metric_args=None)[source]

Bases: object

An evaluation metric.

Parameters

eval_fn –

A function that computes the metric with the following signature:

def eval_fn(
    predictions: pandas.Series,
    targets: pandas.Series,
    metrics: Dict[str, MetricValue],
    **kwargs,
) -> Union[float, MetricValue]:
    """
    Args:
        predictions: A pandas Series containing the predictions made by the model.
        targets: (Optional) A pandas Series containing the corresponding labels
            for the predictions made on that input.
        metrics: (Optional) A dictionary containing the metrics calculated by the
            default evaluator.  The keys are the names of the metrics and the values
            are the metric values.  To access the MetricValue for the metrics
            calculated by the system, make sure to specify the type hint for this
            parameter as Dict[str, MetricValue].  Refer to the DefaultEvaluator
            behavior section for what metrics will be returned based on the type of
            model (i.e. classifier or regressor).
        kwargs: Includes a list of args that are used to compute the metric. These
            args could be information coming from input data, model outputs,
            other metrics, or parameters specified in the `evaluator_config`
            argument of the `mlflow.evaluate` API.

    Returns: MetricValue with per-row scores, per-row justifications, and aggregate
        results.
    """
    ...

name – The name of the metric.
greater_is_better – Whether a higher value of the metric is better.
long_name – (Optional) The long name of the metric. For example, "root_mean_squared_error" for "mse".
version – (Optional) The metric version. For example v1.
metric_details – (Optional) A description of the metric and how it is calculated.
metric_metadata – (Optional) A dictionary containing metadata for the metric.
genai_metric_args – (Optional) A dictionary containing arguments specified by users when calling make_genai_metric or make_genai_metric_from_prompt. Those args are persisted so that we can deserialize the same metric object later.

class mlflow.models.EvaluationResult(metrics, artifacts, baseline_model_metrics=None, run_id=None)[source]

Bases: object

Represents the model evaluation outputs of a mlflow.evaluate() API call, containing both scalar metrics and output artifacts such as performance plots.

property artifacts: A dictionary mapping standardized artifact names (e.g. “roc_data”) to artifact content and location information

property baseline_model_metrics: A dictionary mapping scalar metric names to scalar metric values for the baseline model

classmethod load(path)[source]: Load the evaluation results from the specified local filesystem path

property metrics: A dictionary mapping scalar metric names to scalar metric values

save(path)[source]: Write the evaluation results to the specified local filesystem path

property tables: A dictionary mapping standardized artifact names (e.g. “eval_results_table”) to corresponding table content as pandas DataFrame.

Note

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

class mlflow.models.FlavorBackend(config, **kwargs)[source]

Bases: object

Abstract class for Flavor Backend. This class defines the API interface for local model deployment of MLflow model flavors.

abstract build_image(model_uri, image_name, install_mlflow, mlflow_home, enable_mlserver, base_image=None)[source]

can_build_image()[source]

Returns: True if this flavor has a build_image method defined for building a docker container capable of serving the model, False otherwise.

abstract can_score_model()[source]

Check whether this flavor backend can be deployed in the current environment.

Returns: True if this flavor backend can be applied in the current environment.

abstract generate_dockerfile(model_uri, output_path, install_mlflow, mlflow_home, enable_mlserver, base_image=None)[source]

abstract predict(model_uri, input_path, output_path, content_type)[source]

Generate predictions using a saved MLflow model referenced by the given URI. Input and output are read from and written to a file or stdin / stdout.

Parameters

model_uri – URI pointing to the MLflow model to be used for scoring.
input_path – Path to the file with input data. If not specified, data is read from stdin.
output_path – Path to the file with output predictions. If not specified, data is written to stdout.
content_type – Specifies the input format. Can be one of {json, csv}

prepare_env(model_uri, capture_output=False)[source]: Performs any preparation necessary to predict or serve the model, for example downloading dependencies or initializing a conda environment. After preparation, calling predict or serve should be fast.

abstract serve(model_uri, port, host, timeout, enable_mlserver, synchronous=True, stdout=None, stderr=None)[source]

Serve the specified MLflow model locally.

Parameters

model_uri – URI pointing to the MLflow model to be used for scoring.
port – Port to use for the model deployment.
host – Host to use for the model deployment. Defaults to localhost.
timeout – Timeout in seconds to serve a request. Defaults to 60.
enable_mlserver – Whether to use MLServer or the local scoring server.
synchronous – If True, wait until server process exit and return 0, if process exit with non-zero return code, raise exception. If False, return the server process Popen instance immediately.
stdout – Redirect server stdout
stderr – Redirect server stderr

class mlflow.models.MetricThreshold(threshold=None, min_absolute_change=None, min_relative_change=None, greater_is_better=None, higher_is_better=None)[source]

Bases: object

This class allows you to define metric thresholds for model validation. Allowed thresholds are: threshold, min_absolute_change, min_relative_change.

Parameters

threshold –
(Optional) A number representing the value threshold for the metric.
- If higher is better for the metric, the metric value has to be >= threshold to pass validation.
- Otherwise, the metric value has to be <= threshold to pass the validation.
min_absolute_change –
(Optional) A positive number representing the minimum absolute change required for candidate model to pass validation with the baseline model.
- If higher is better for the metric, metric value has to be >= baseline model metric value + min_absolute_change to pass the validation.
- Otherwise, metric value has to be <= baseline model metric value - min_absolute_change to pass the validation.
min_relative_change –
(Optional) A floating point number between 0 and 1 representing the minimum relative change (in percentage of baseline model metric value) for candidate model to pass the comparison with the baseline model.
- If higher is better for the metric, metric value has to be >= baseline model metric value * (1 + min_relative_change)
- Otherwise, metric value has to be <= baseline model metric value * (1 - min_relative_change)
- Note that if the baseline model metric value is equal to 0, the threshold falls back performing a simple verification that the candidate metric value is better than the baseline metric value, i.e. metric value >= baseline model metric value + 1e-10 if higher is better; metric value <= baseline model metric value - 1e-10 if lower is better.
greater_is_better – A required boolean representing whether higher value is better for the metric.
higher_is_better –

Deprecated since version 2.3.0: Use greater_is_better instead.

A required boolean representing whether higher value is better for the metric.

property greater_is_better: Boolean value representing whether higher value is better for the metric.

property higher_is_better: Warning

mlflow.models.evaluation.validation.MetricThreshold.higher_is_better is deprecated. This method will be removed in a future release. Use The attribute `higher_is_better` is deprecated. Use `greater_is_better` instead. instead.

Boolean value representing whether higher value is better for the metric.

property min_absolute_change: Value of the minimum absolute change required to pass model comparison with baseline model.

property min_relative_change: Float value of the minimum relative change required to pass model comparison with baseline model.

property threshold: Value of the threshold.

class mlflow.models.Model(artifact_path=None, run_id=None, utc_time_created=None, flavors=None, signature=None, saved_input_example_info: Optional[Dict[str, Any]] = None, model_uuid: Optional[Union[str, Callable]] = <function Model.<lambda>>, mlflow_version: Optional[str] = '2.15.2.dev0', metadata: Optional[Dict[str, Any]] = None, model_size_bytes: Optional[int] = None, resources: Optional[Union[str, List[mlflow.models.resources.Resource]]] = None, **kwargs)[source]

Bases: object

An MLflow Model that can support multiple model flavors. Provides APIs for implementing new Model flavors.

add_flavor(name, **params)[source]: Add an entry for how to serve the model in a given format.

classmethod from_dict(model_dict)[source]: Load a model from its YAML representation.

get_input_schema()[source]: Retrieves the input schema of the Model iff the model was saved with a schema definition.

get_model_info()[source]: Create a ModelInfo instance that contains the model metadata.

get_output_schema()[source]: Retrieves the output schema of the Model iff the model was saved with a schema definition.

get_params_schema()[source]: Retrieves the parameters schema of the Model iff the model was saved with a schema definition.

get_serving_input(path: str)[source]

Load serving input example from a model directory. Returns None if there is no serving input example.

Parameters: path – Path to the model directory.
Returns: Serving input example or None if the model has no serving input example.

classmethod load(path)[source]

Load a model from its YAML representation.

Parameters: path – A local filesystem path or URI referring to the MLmodel YAML file representation of the Model object or to the directory containing the MLmodel YAML file representation.
Returns: An instance of Model.

example

from mlflow.models import Model

# Load the Model object from a local MLmodel file
model1 = Model.load("~/path/to/my/MLmodel")

# Load the Model object from a remote model directory
model2 = Model.load("s3://mybucket/path/to/my/model")

load_input_example(path: str)[source]

Load the input example saved along a model. Returns None if there is no example metadata (i.e. the model was saved without example). Raises FileNotFoundError if there is model metadata but the example file is missing.

Parameters: path – Path to the model directory.
Returns: Input example (NumPy ndarray, SciPy csc_matrix, SciPy csr_matrix, pandas DataFrame, dict) or None if the model has no example.

load_input_example_params(path: str)[source]

Load the params of input example saved along a model. Returns None if there are no params in the input_example.

Parameters: path – Path to the model directory.
Returns: params (dict) or None if the model has no params.

classmethod log(artifact_path, flavor, registered_model_name=None, await_registration_for=300, metadata=None, run_id=None, resources=None, **kwargs)[source]

Log model using supplied flavor module. If no run is active, this method will create a new active run.

Parameters

artifact_path – Run relative path identifying the model.
flavor – Flavor module to save the model with. The module must have the save_model function that will persist the model as a valid MLflow model.
registered_model_name – 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 – {{ signature }}
input_example – {{ input_example }}
await_registration_for – Number of seconds to wait for the model version to finish being created and is in READY status. By default, the function waits for five minutes. Specify 0 or None to skip waiting.
metadata – {{ metadata }}
resources – {{ resources }}
kwargs – Extra args passed to the model flavor.

Returns

A ModelInfo instance that contains the metadata of the logged model.

property metadata

Custom metadata dictionary passed to the model and stored in the MLmodel file.

Getter: Retrieves custom metadata that have been applied to a model instance.
Setter: Sets a dictionary of custom keys and values to be included with the model instance
Type: Optional[Dict[str, Any]]
Returns: A Dictionary of user-defined metadata iff defined.

Example

# Create and log a model with metadata to the Model Registry
from sklearn import datasets
from sklearn.ensemble import RandomForestClassifier
import mlflow
from mlflow.models import infer_signature

with mlflow.start_run():
    iris = datasets.load_iris()
    clf = RandomForestClassifier()
    clf.fit(iris.data, iris.target)
    signature = infer_signature(iris.data, iris.target)
    mlflow.sklearn.log_model(
        clf,
        "iris_rf",
        signature=signature,
        registered_model_name="model-with-metadata",
        metadata={"metadata_key": "metadata_value"},
    )

# model uri for the above model
model_uri = "models:/model-with-metadata/1"

# Load the model and access the custom metadata
model = mlflow.pyfunc.load_model(model_uri=model_uri)
assert model.metadata.metadata["metadata_key"] == "metadata_value"

Note

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

property model_size_bytes

An optional integer that represents the model size in bytes

Getter: Retrieves the model size if it’s calculated when the model is saved
Setter: Sets the model size to a model instance
Type: Optional[int]

property resources

An optional dictionary that contains the resources required to serve the model.

Getter: Retrieves the resources required to serve the model
Setter: Sets the resources required to serve the model
Type: Dict[str, Dict[ResourceType, List[Dict]]]

Note

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

save(path)[source]: Write the model as a local YAML file.

property saved_input_example_info: A dictionary that contains the metadata of the saved input example, e.g., {"artifact_path": "input_example.json", "type": "dataframe", "pandas_orient": "split"}.

property signature

An optional definition of the expected inputs to and outputs from a model object, defined with both field names and data types. Signatures support both column-based and tensor-based inputs and outputs.

Getter: Retrieves the signature of a model instance iff the model was saved with a signature definition.
Setter: Sets a signature to a model instance.
Type: Optional[ModelSignature]

to_dict()[source]: Serialize the model to a dictionary.

to_json()[source]: Write the model as json.

to_yaml(stream=None)[source]: Write the model as yaml string.

class mlflow.models.ModelConfig(*, development_config: Optional[Union[str, Dict[str, Any]]] = None)[source]

Bases: object

ModelConfig used in code to read a YAML configuration file, and this configuration file can be overridden when logging a model.

get(key)[source]: Gets the value of a top-level parameter in the configuration.

class mlflow.models.ModelSignature(inputs: Optional[Union[mlflow.types.schema.Schema, dataclasses.dataclass]] = None, outputs: Optional[Union[mlflow.types.schema.Schema, dataclasses.dataclass]] = None, params: Optional[mlflow.types.schema.ParamSchema] = None)[source]

Bases: object

ModelSignature specifies schema of model’s inputs, outputs and params.

ModelSignature can be inferred from training dataset, model predictions using and params for inference, or constructed by hand by passing an input and output Schema, and params ParamSchema.

classmethod from_dict(signature_dict: Dict[str, Any])[source]

Deserialize from dictionary representation.

Parameters: signature_dict – Dictionary representation of model signature. Expected dictionary format: {‘inputs’: <json string>, ‘outputs’: <json string>, ‘params’: <json string>” }
Returns: ModelSignature populated with the data form the dictionary.

to_dict() → Dict[str, Any][source]

Serialize into a ‘jsonable’ dictionary.

Input and output schema are represented as json strings. This is so that the representation is compact when embedded in an MLmodel yaml file.

Returns: dictionary representation with input and output schema represented as json strings.

class mlflow.models.Resource(type: mlflow.models.resources.ResourceType, target_uri: str)[source]

Bases: abc.ABC

Base class for defining the resources needed to serve a model.

Parameters

type (ResourceType) – The resource type.
target_uri (str) – The target URI where these resources are hosted.

classmethod from_dict(data)[source]: Convert the dictionary to a Resource. Subclasses must implement this method.

target_uri: str

abstract to_dict()[source]: Convert the resource to a dictionary. Subclasses must implement this method.

type: mlflow.models.resources.ResourceType

class mlflow.models.ResourceType(value)[source]

Bases: enum.Enum

Enum to define the different types of resources needed to serve a model.

SERVING_ENDPOINT = 'serving_endpoint'

SQL_WAREHOUSE = 'sql_warehouse'

UC_FUNCTION = 'uc_function'

VECTOR_SEARCH_INDEX = 'vector_search_index'

mlflow.models.add_libraries_to_model(model_uri, run_id=None, registered_model_name=None)[source]

Note

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

Given a registered model_uri (e.g. models:/<model_name>/<model_version>), this utility re-logs the model along with all the required model libraries back to the Model Registry. The required model libraries are stored along with the model as model artifacts. In addition, supporting files to the model (e.g. conda.yaml, requirements.txt) are modified to use the added libraries.

By default, this utility creates a new model version under the same registered model specified by model_uri. This behavior can be overridden by specifying the registered_model_name argument.

Parameters

model_uri – A registered model uri in the Model Registry of the form models:/<model_name>/<model_version/stage/latest>
run_id – The ID of the run to which the model with libraries is logged. If None, the model with libraries is logged to the source run corresponding to model version specified by model_uri; if the model version does not have a source run, a new run created.
registered_model_name – The new model version (model with its libraries) is registered under the inputted registered_model_name. If None, a new version is logged to the existing model in the Model Registry.

Note

This utility only operates on a model that has been registered to the Model Registry.

Note

The libraries are only compatible with the platform on which they are added. Cross platform libraries are not supported.

Example

# Create and log a model to the Model Registry
import pandas as pd
from sklearn import datasets
from sklearn.ensemble import RandomForestClassifier
import mlflow
import mlflow.sklearn
from mlflow.models import infer_signature

with mlflow.start_run():
    iris = datasets.load_iris()
    iris_train = pd.DataFrame(iris.data, columns=iris.feature_names)
    clf = RandomForestClassifier(max_depth=7, random_state=0)
    clf.fit(iris_train, iris.target)
    signature = infer_signature(iris_train, clf.predict(iris_train))
    mlflow.sklearn.log_model(
        clf, "iris_rf", signature=signature, registered_model_name="model-with-libs"
    )

# model uri for the above model
model_uri = "models:/model-with-libs/1"

# Import utility
from mlflow.models.utils import add_libraries_to_model

# Log libraries to the original run of the model
add_libraries_to_model(model_uri)

# Log libraries to some run_id
existing_run_id = "21df94e6bdef4631a9d9cb56f211767f"
add_libraries_to_model(model_uri, run_id=existing_run_id)

# Log libraries to a new run
with mlflow.start_run():
    add_libraries_to_model(model_uri)

# Log libraries to a new registered model named 'new-model'
with mlflow.start_run():
    add_libraries_to_model(model_uri, registered_model_name="new-model")

mlflow.models.build_docker(model_uri=None, name='mlflow-pyfunc', env_manager='virtualenv', mlflow_home=None, install_java=False, install_mlflow=False, enable_mlserver=False, base_image=None)[source]

Builds a Docker image whose default entrypoint serves an MLflow model at port 8080, using the python_function flavor. The container serves the model referenced by model_uri, if specified. If model_uri is not specified, an MLflow Model directory must be mounted as a volume into the /opt/ml/model directory in the container.

Important

Since MLflow 2.10.1, the Docker image built with --model-uri does not install Java for improved performance, unless the model flavor is one of ["johnsnowlabs", "h2o", "mleap", "spark"]. If you need to install Java for other flavors, e.g. custom Python model that uses SparkML, please specify install-java=True to enforce Java installation. For earlier versions, Java is always installed to the image.

Warning

If model_uri is unspecified, the resulting image doesn’t support serving models with the RFunc or Java MLeap model servers.

NB: by default, the container will start nginx and gunicorn processes. If you don’t need the nginx process to be started (for instance if you deploy your container to Google Cloud Run), you can disable it via the DISABLE_NGINX environment variable:

docker run -p 5001:8080 -e DISABLE_NGINX=true "my-image-name"

See https://www.mlflow.org/docs/latest/python_api/mlflow.pyfunc.html for more information on the ‘python_function’ flavor.

Parameters

model_uri – URI to the model. A local path, a ‘runs:/’ URI, or a remote storage URI (e.g., an ‘s3://’ URI). For more information about supported remote URIs for model artifacts, see https://mlflow.org/docs/latest/tracking.html#artifact-stores”
name – Name of the Docker image to build. Defaults to ‘mlflow-pyfunc’.
install_java – If specified, install Java in the image. Default is False in order to reduce both the image size and the build time. Model flavors requiring Java will enable this setting automatically, such as the Spark flavor. (This argument is only available in MLflow 2.10.1 and later. In earlier versions, Java is always installed to the image.)
enable_mlserver – If specified, the image will be built with the Seldon MLserver as backend.
env_manager – If specified, create an environment for MLmodel using the specified environment manager. The following values are supported: (1) virtualenv (default): use virtualenv and pyenv for Python version management (2) conda: use conda (3) local: use the local environment without creating a new one.
install_mlflow – If specified and there is a conda or virtualenv environment to be activated mlflow will be installed into the environment after it has been activated. The version of installed mlflow will be the same as the one used to invoke this command.
mlflow_home – Path to local clone of MLflow project. Use for development only.
base_image – Base image for the Docker image. If not specified, the default image is either UBUNTU_BASE_IMAGE = “ubuntu:20.04” or PYTHON_SLIM_BASE_IMAGE = “python:{version}-slim” Note: If custom image is used, there are no guarantees that the image will work. You may find greater compatibility by building your image on top of the ubuntu images. In addition, you must install Java and virtualenv to have the image work properly.

mlflow.models.convert_input_example_to_serving_input(input_example) → Optional[str][source]

Note

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

Helper function to convert a model’s input example to a serving input example that can be used for model inference in the scoring server.

Parameters: input_example – model input example. Supported types are pandas.DataFrame, numpy.ndarray, dictionary of (name -> numpy.ndarray), list, scalars and dicts with json serializable values.
Returns: serving input example as a json string

mlflow.models.evaluate(model=None, data=None, *, model_type=None, targets=None, predictions=None, dataset_path=None, feature_names=None, evaluators=None, evaluator_config=None, custom_metrics=None, extra_metrics=None, custom_artifacts=None, validation_thresholds=None, baseline_model=None, env_manager='local', model_config=None, baseline_config=None, inference_params=None)[source]

Evaluate the model performance on given data and selected metrics.

This function evaluates a PyFunc model or custom callable on the specified dataset using specified evaluators, and logs resulting metrics & artifacts to MLflow tracking server. Users can also skip setting model and put the model outputs in data directly for evaluation. For detailed information, please read the Model Evaluation documentation.

Default Evaluator behavior:

The default evaluator, which can be invoked with evaluators="default" or evaluators=None, supports model types listed below. For each pre-defined model type, the default evaluator evaluates your model on a selected set of metrics and generate artifacts like plots. Please find more details below.
For both the "regressor" and "classifier" model types, the default evaluator generates model summary plots and feature importance plots using SHAP.
For regressor models, the default evaluator additionally logs:
- metrics: example_count, mean_absolute_error, mean_squared_error, root_mean_squared_error, sum_on_target, mean_on_target, r2_score, max_error, mean_absolute_percentage_error.
For binary classifiers, the default evaluator additionally logs:
- metrics: true_negatives, false_positives, false_negatives, true_positives, recall, precision, f1_score, accuracy_score, example_count, log_loss, roc_auc, precision_recall_auc.
- artifacts: lift curve plot, precision-recall plot, ROC plot.
For multiclass classifiers, the default evaluator additionally logs:
- metrics: accuracy_score, example_count, f1_score_micro, f1_score_macro, log_loss
- artifacts: A CSV file for “per_class_metrics” (per-class metrics includes true_negatives/false_positives/false_negatives/true_positives/recall/precision/roc_auc, precision_recall_auc), precision-recall merged curves plot, ROC merged curves plot.
For question-answering models, the default evaluator logs:
- metrics: exact_match, token_count, toxicity (requires evaluate, torch, flesch_kincaid_grade_level (requires textstat) and ari_grade_level.
- artifacts: A JSON file containing the inputs, outputs, targets (if the targets argument is supplied), and per-row metrics of the model in tabular format.
For text-summarization models, the default evaluator logs:
- metrics: token_count, ROUGE (requires evaluate, nltk, and rouge_score to be installed), toxicity (requires evaluate, torch, transformers), ari_grade_level (requires textstat), flesch_kincaid_grade_level (requires textstat).
- artifacts: A JSON file containing the inputs, outputs, targets (if the targets argument is supplied), and per-row metrics of the model in the tabular format.
For text models, the default evaluator logs:
- metrics: token_count, toxicity (requires evaluate, torch, transformers), ari_grade_level (requires textstat), flesch_kincaid_grade_level (requires textstat).
- artifacts: A JSON file containing the inputs, outputs, targets (if the targets argument is supplied), and per-row metrics of the model in tabular format.
For retriever models, the default evaluator logs:
- metrics: precision_at_k(k), recall_at_k(k) and ndcg_at_k(k) - all have a default value of retriever_k = 3.
- artifacts: A JSON file containing the inputs, outputs, targets, and per-row metrics of the model in tabular format.
For sklearn models, the default evaluator additionally logs the model’s evaluation criterion (e.g. mean accuracy for a classifier) computed by model.score method.
The metrics/artifacts listed above are logged to the active MLflow run. If no active run exists, a new MLflow run is created for logging these metrics and artifacts. Note that no metrics/artifacts are logged for the baseline_model.
Additionally, information about the specified dataset - hash, name (if specified), path (if specified), and the UUID of the model that evaluated it - is logged to the mlflow.datasets tag.
The available evaluator_config options for the default evaluator include:
- log_model_explainability: A boolean value specifying whether or not to log model explainability insights, default value is True.
- explainability_algorithm: A string to specify the SHAP Explainer algorithm for model explainability. Supported algorithm includes: ‘exact’, ‘permutation’, ‘partition’, ‘kernel’. If not set, shap.Explainer is used with the “auto” algorithm, which chooses the best Explainer based on the model.
- explainability_nsamples: The number of sample rows to use for computing model explainability insights. Default value is 2000.
- explainability_kernel_link: The kernel link function used by shap kernal explainer. Available values are “identity” and “logit”. Default value is “identity”.
- max_classes_for_multiclass_roc_pr: For multiclass classification tasks, the maximum number of classes for which to log the per-class ROC curve and Precision-Recall curve. If the number of classes is larger than the configured maximum, these curves are not logged.
- metric_prefix: An optional prefix to prepend to the name of each metric and artifact produced during evaluation.
- log_metrics_with_dataset_info: A boolean value specifying whether or not to include information about the evaluation dataset in the name of each metric logged to MLflow Tracking during evaluation, default value is True.
- pos_label: If specified, the positive label to use when computing classification metrics such as precision, recall, f1, etc. for binary classification models. For multiclass classification and regression models, this parameter will be ignored.
- average: The averaging method to use when computing classification metrics such as precision, recall, f1, etc. for multiclass classification models (default: 'weighted'). For binary classification and regression models, this parameter will be ignored.
- sample_weights: Weights for each sample to apply when computing model performance metrics.
- col_mapping: A dictionary mapping column names in the input dataset or output predictions to column names used when invoking the evaluation functions.
- retriever_k: A parameter used when model_type="retriever" as the number of top-ranked retrieved documents to use when computing the built-in metric precision_at_k(k), recall_at_k(k) and ndcg_at_k(k). Default value is 3. For all other model types, this parameter will be ignored.
Limitations of evaluation dataset:
- For classification tasks, dataset labels are used to infer the total number of classes.
- For binary classification tasks, the negative label value must be 0 or -1 or False, and the positive label value must be 1 or True.
Limitations of metrics/artifacts computation:
- For classification tasks, some metric and artifact computations require the model to output class probabilities. Currently, for scikit-learn models, the default evaluator calls the predict_proba method on the underlying model to obtain probabilities. For other model types, the default evaluator does not compute metrics/artifacts that require probability outputs.
Limitations of default evaluator logging model explainability insights:
- The shap.Explainer auto algorithm uses the Linear explainer for linear models and the Tree explainer for tree models. Because SHAP’s Linear and Tree explainers do not support multi-class classification, the default evaluator falls back to using the Exact or Permutation explainers for multi-class classification tasks.
- Logging model explainability insights is not currently supported for PySpark models.
- The evaluation dataset label values must be numeric or boolean, all feature values must be numeric, and each feature column must only contain scalar values.
Limitations when environment restoration is enabled:
- When environment restoration is enabled for the evaluated model (i.e. a non-local env_manager is specified), the model is loaded as a client that invokes a MLflow Model Scoring Server process in an independent Python environment with the model’s training time dependencies installed. As such, methods like predict_proba (for probability outputs) or score (computes the evaluation criterian for sklearn models) of the model become inaccessible and the default evaluator does not compute metrics or artifacts that require those methods.
- Because the model is an MLflow Model Server process, SHAP explanations are slower to compute. As such, model explainaibility is disabled when a non-local env_manager specified, unless the evaluator_config option log_model_explainability is explicitly set to True.

Parameters

model –
Optional. If specified, it should be one of the following:
- A pyfunc model instance
- A URI referring to a pyfunc model
- A URI referring to an MLflow Deployments endpoint e.g. "endpoints:/my-chat"
- A callable function: This function should be able to take in model input and return predictions. It should follow the signature of the predict method. Here’s an example of a valid function:
  
  ..code-block:: python
  
  model = mlflow.pyfunc.load_model(model_uri)
  
  def fn(model_input):
  return model.predict(model_input)
If omitted, it indicates a static dataset will be used for evaluation instead of a model. In this case, the data argument must be a Pandas DataFrame or an mlflow PandasDataset that contains model outputs, and the predictions argument must be the name of the column in data that contains model outputs.
data –
One of the following:
- A numpy array or list of evaluation features, excluding labels.
- A Pandas DataFrame containing evaluation features, labels, and optionally model
  outputs. Model outputs are required to be provided when model is unspecified. If feature_names argument not specified, all columns except for the label column and model_output column are regarded as feature columns. Otherwise, only column names present in feature_names are regarded as feature columns.
- A Spark DataFrame containing evaluation features and labels. If
  feature_names argument not specified, all columns except for the label column are regarded as feature columns. Otherwise, only column names present in feature_names are regarded as feature columns. Only the first 10000 rows in the Spark DataFrame will be used as evaluation data.
- A mlflow.data.dataset.Dataset instance containing evaluation
  features, labels, and optionally model outputs. Model outputs are only supported with a PandasDataset. Model outputs are required when model is unspecified, and should be specified via the predictions prerty of the PandasDataset.
targets – If data is a numpy array or list, a numpy array or list of evaluation labels. If data is a DataFrame, the string name of a column from data that contains evaluation labels. Required for classifier and regressor models, but optional for question-answering, text-summarization, and text models. If data is a mlflow.data.dataset.Dataset that defines targets, then targets is optional.

predictions –

Optional. The name of the column that contains model outputs.

When model is specified and outputs multiple columns, predictions can be used to specify the name of the column that will be used to store model outputs for evaluation.
When model is not specified and data is a pandas dataframe, predictions can be used to specify the name of the column in data that contains model outputs.

Example usage of predictions

# Evaluate a model that outputs multiple columns
data = pd.DataFrame({"question": ["foo"]})


def model(inputs):
    return pd.DataFrame({"answer": ["bar"], "source": ["baz"]})


results = evaluate(model=model, data=data, predictions="answer", ...)

# Evaluate a static dataset
data = pd.DataFrame({"question": ["foo"], "answer": ["bar"], "source": ["baz"]})
results = evaluate(data=data, predictions="answer", ...)

model_type –
(Optional) A string describing the model type. The default evaluator supports the following model types:
- 'classifier'
- 'regressor'
- 'question-answering'
- 'text-summarization'
- 'text'
- 'retriever'
If no model_type is specified, then you must provide a a list of metrics to compute via the extra_metrics param.

Note

'question-answering', 'text-summarization', 'text', and 'retriever' are experimental and may be changed or removed in a future release.
inference_params – (Optional) A dictionary of inference parameters to be passed to the model when making predictions, such as {"max_tokens": 100}. This is only used when the model is an MLflow Deployments endpoint URI e.g. "endpoints:/my-chat"
dataset_path – (Optional) The path where the data is stored. Must not contain double quotes (“). If specified, the path is logged to the mlflow.datasets tag for lineage tracking purposes.
feature_names – (Optional) A list. If the data argument is a numpy array or list, feature_names is a list of the feature names for each feature. If feature_names=None, then the feature_names are generated using the format feature_{feature_index}. If the data argument is a Pandas DataFrame or a Spark DataFrame, feature_names is a list of the names of the feature columns in the DataFrame. If feature_names=None, then all columns except the label column and the predictions column are regarded as feature columns.
evaluators – The name of the evaluator to use for model evaluation, or a list of evaluator names. If unspecified, all evaluators capable of evaluating the specified model on the specified dataset are used. The default evaluator can be referred to by the name "default". To see all available evaluators, call mlflow.models.list_evaluators().
evaluator_config – A dictionary of additional configurations to supply to the evaluator. If multiple evaluators are specified, each configuration should be supplied as a nested dictionary whose key is the evaluator name.
extra_metrics –
(Optional) A list of EvaluationMetric objects. These metrics are computed in addition to the default metrics associated with pre-defined model_type, and setting model_type=None will only compute the metrics specified in extra_metrics. See the mlflow.metrics module for more information about the builtin metrics and how to define extra metrics.
Example usage of extra metrics
```
import mlflow
import numpy as np


def root_mean_squared_error(eval_df, _builtin_metrics):
    return np.sqrt((np.abs(eval_df["prediction"] - eval_df["target"]) ** 2).mean)


rmse_metric = mlflow.models.make_metric(
    eval_fn=root_mean_squared_error,
    greater_is_better=False,
)
mlflow.evaluate(..., extra_metrics=[rmse_metric])
```

custom_artifacts –

(Optional) A list of custom artifact functions with the following signature:

def custom_artifact(
    eval_df: Union[pandas.Dataframe, pyspark.sql.DataFrame],
    builtin_metrics: Dict[str, float],
    artifacts_dir: str,
) -> Dict[str, Any]:
    """
    Args:
        eval_df:
            A Pandas or Spark DataFrame containing ``prediction`` and ``target``
            column.  The ``prediction`` column contains the predictions made by the
            model.  The ``target`` column contains the corresponding labels to the
            predictions made on that row.
        builtin_metrics:
            A dictionary containing the metrics calculated by the default evaluator.
            The keys are the names of the metrics and the values are the scalar
            values of the metrics. Refer to the DefaultEvaluator behavior section
            for what metrics will be returned based on the type of model (i.e.
            classifier or regressor).
        artifacts_dir:
            A temporary directory path that can be used by the custom artifacts
            function to temporarily store produced artifacts. The directory will be
            deleted after the artifacts are logged.

    Returns:
        A dictionary that maps artifact names to artifact objects
        (e.g. a Matplotlib Figure) or to artifact paths within ``artifacts_dir``.
    """
    ...

Object types that artifacts can be represented as:

A string uri representing the file path to the artifact. MLflow will infer the type of the artifact based on the file extension.

A string representation of a JSON object. This will be saved as a .json artifact.

Pandas DataFrame. This will be resolved as a CSV artifact.

Numpy array. This will be saved as a .npy artifact.

Matplotlib Figure. This will be saved as an image artifact. Note that matplotlib.pyplot.savefig is called behind the scene with default configurations. To customize, either save the figure with the desired configurations and return its file path or define customizations through environment variables in matplotlib.rcParams.

Other objects will be attempted to be pickled with the default protocol.

Example usage of custom artifacts

import mlflow
import matplotlib.pyplot as plt


def scatter_plot(eval_df, builtin_metrics, artifacts_dir):
    plt.scatter(eval_df["prediction"], eval_df["target"])
    plt.xlabel("Targets")
    plt.ylabel("Predictions")
    plt.title("Targets vs. Predictions")
    plt.savefig(os.path.join(artifacts_dir, "example.png"))
    plt.close()
    return {"pred_target_scatter": os.path.join(artifacts_dir, "example.png")}


def pred_sample(eval_df, _builtin_metrics, _artifacts_dir):
    return {"pred_sample": pred_sample.head(10)}


mlflow.evaluate(..., custom_artifacts=[scatter_plot, pred_sample])

validation_thresholds –

(Optional) A dictionary of metric name to mlflow.models.MetricThreshold used for model validation. Each metric name must either be the name of a builtin metric or the name of a metric defined in the extra_metrics parameter.

Example of Model Validation

from mlflow.models import MetricThreshold

thresholds = {
    "accuracy_score": MetricThreshold(
        # accuracy should be >=0.8
        threshold=0.8,
        # accuracy should be at least 5 percent greater than baseline model accuracy
        min_absolute_change=0.05,
        # accuracy should be at least 0.05 greater than baseline model accuracy
        min_relative_change=0.05,
        greater_is_better=True,
    ),
}

with mlflow.start_run():
    mlflow.evaluate(
        model=your_candidate_model,
        data,
        targets,
        model_type,
        dataset_name,
        evaluators,
        validation_thresholds=thresholds,
        baseline_model=your_baseline_model,
    )

See the Model Validation documentation for more details.

baseline_model – (Optional) A string URI referring to an MLflow model with the pyfunc flavor. If specified, the candidate model is compared to this baseline for model validation purposes.
env_manager –
Specify an environment manager to load the candidate model and baseline_model in isolated Python environments and restore their dependencies. Default value is local, and the following values are supported:
- virtualenv: (Recommended) Use virtualenv to restore the python environment that was used to train the model.
- conda: Use Conda to restore the software environment that was used to train the model.
- local: Use the current Python environment for model inference, which may differ from the environment used to train the model and may lead to errors or invalid predictions.
model_config – the model configuration to use for loading the model with pyfunc. Inspect the model’s pyfunc flavor to know which keys are supported for your specific model. If not indicated, the default model configuration from the model is used (if any).
baseline_config – the model configuration to use for loading the baseline model. If not indicated, the default model configuration from the baseline model is used (if any).

Returns

An mlflow.models.EvaluationResult instance containing metrics of candidate model and baseline model, and artifacts of candidate model.

mlflow.models.get_model_info(model_uri: str) → mlflow.models.model.ModelInfo [source]

Get metadata for the specified model, such as its input/output signature.

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>
mlflow-artifacts:/path/to/model

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

Returns

A ModelInfo instance that contains the metadata of the logged model.

Example usage of get_model_info

import mlflow.models
import mlflow.sklearn
from sklearn.ensemble import RandomForestRegressor

with mlflow.start_run() as run:
    params = {"n_estimators": 3, "random_state": 42}
    X, y = [[0, 1]], [1]
    signature = mlflow.models.infer_signature(X, y)
    rfr = RandomForestRegressor(**params).fit(X, y)
    mlflow.log_params(params)
    mlflow.sklearn.log_model(rfr, artifact_path="sklearn-model", signature=signature)

model_uri = f"runs:/{run.info.run_id}/sklearn-model"
# Get model info with model_uri
model_info = mlflow.models.get_model_info(model_uri)
# Get model signature directly
model_signature = model_info.signature
assert model_signature == signature

mlflow.models.infer_pip_requirements(model_uri, flavor, fallback=None, timeout=None)[source]

Infers the pip requirements of the specified model by creating a subprocess and loading the model in it to determine which packages are imported.

Parameters

model_uri – The URI of the model.
flavor – The flavor name of the model.
fallback – If provided, an unexpected error during the inference procedure is swallowed and the value of fallback is returned. Otherwise, the error is raised.
timeout – If specified, the inference operation is bound by the timeout (in seconds).

Returns

A list of inferred pip requirements (e.g. ["scikit-learn==0.24.2", ...]).

mlflow.models.infer_signature(model_input: Any = None, model_output: MlflowInferableDataset = None, params: Optional[Dict[str, Any]] = None) → mlflow.models.signature.ModelSignature[source]

Infer an MLflow model signature from the training data (input), model predictions (output) and parameters (for inference).

The signature represents model input and output as data frames with (optionally) named columns and data type specified as one of types defined in mlflow.types.DataType. It also includes parameters schema for inference, . This method will raise an exception if the user data contains incompatible types or is not passed in one of the supported formats listed below.

The input should be one of these:

pandas.DataFrame
pandas.Series
dictionary of { name -> numpy.ndarray}
numpy.ndarray
pyspark.sql.DataFrame
scipy.sparse.csr_matrix
scipy.sparse.csc_matrix
dictionary / list of dictionaries of JSON-convertible types

The element types should be mappable to one of mlflow.types.DataType.

For pyspark.sql.DataFrame inputs, columns of type DateType and TimestampType are both inferred as type datetime, which is coerced to TimestampType at inference.

Parameters

model_input – Valid input to the model. E.g. (a subset of) the training dataset.
model_output – Valid model output. E.g. Model predictions for the (subset of) training dataset.

params –

Valid parameters for inference. It should be a dictionary of parameters that can be set on the model during inference by passing params to pyfunc predict method.

An example of valid parameters:

from mlflow.models import infer_signature
from mlflow.transformers import generate_signature_output

# Define parameters for inference
params = {
    "num_beams": 5,
    "max_length": 30,
    "do_sample": True,
    "remove_invalid_values": True,
}

# Infer the signature including parameters
signature = infer_signature(
    data,
    generate_signature_output(model, data),
    params=params,
)

# Saving model with model signature
mlflow.transformers.save_model(
    model,
    path=model_path,
    signature=signature,
)

pyfunc_loaded = mlflow.pyfunc.load_model(model_path)

# Passing params to `predict` function directly
result = pyfunc_loaded.predict(data, params=params)

Returns

ModelSignature

mlflow.models.list_evaluators()[source]: Return a name list for all available Evaluators.

mlflow.models.make_metric(*, eval_fn, greater_is_better, name=None, long_name=None, version=None, metric_details=None, metric_metadata=None, genai_metric_args=None)[source]

A factory function to create an EvaluationMetric object.

Parameters

eval_fn –

A function that computes the metric with the following signature:

def eval_fn(
    predictions: pandas.Series,
    targets: pandas.Series,
    metrics: Dict[str, MetricValue],
    **kwargs,
) -> Union[float, MetricValue]:
    """
    Args:
        predictions: A pandas Series containing the predictions made by the model.
        targets: (Optional) A pandas Series containing the corresponding labels
            for the predictions made on that input.
        metrics: (Optional) A dictionary containing the metrics calculated by the
            default evaluator.  The keys are the names of the metrics and the values
            are the metric values.  To access the MetricValue for the metrics
            calculated by the system, make sure to specify the type hint for this
            parameter as Dict[str, MetricValue].  Refer to the DefaultEvaluator
            behavior section for what metrics will be returned based on the type of
            model (i.e. classifier or regressor).  kwargs: Includes a list of args
            that are used to compute the metric. These args could information coming
            from input data, model outputs or parameters specified in the
            `evaluator_config` argument of the `mlflow.evaluate` API.
        kwargs: Includes a list of args that are used to compute the metric. These
            args could be information coming from input data, model outputs,
            other metrics, or parameters specified in the `evaluator_config`
            argument of the `mlflow.evaluate` API.

    Returns: MetricValue with per-row scores, per-row justifications, and aggregate
        results.
    """
    ...

greater_is_better – Whether a higher value of the metric is better.
name – The name of the metric. This argument must be specified if eval_fn is a lambda function or the eval_fn.__name__ attribute is not available.
long_name – (Optional) The long name of the metric. For example, "mean_squared_error" for "mse".
version – (Optional) The metric version. For example v1.
metric_details – (Optional) A description of the metric and how it is calculated.
metric_metadata – (Optional) A dictionary containing metadata for the metric.
genai_metric_args – (Optional) A dictionary containing arguments specified by users when calling make_genai_metric or make_genai_metric_from_prompt. Those args are persisted so that we can deserialize the same metric object later.