Skip to content

Latest commit

 

History

History
123 lines (83 loc) · 5.68 KB

ref_model_artifacts_format.md

File metadata and controls

123 lines (83 loc) · 5.68 KB
Raw

Legion Model Artifact format

This format declares, how you can store ML models, built on different languages (Python, Java, etc) with wide range of machine-learning libs used (scikit-learn, tensorflow, keras, etc.).

Legion Model Artifact is representative as file-system folder packed to ZIP file using Deflate ZIP compression algorithm.

Legion Model Artifact folder contains of next objects:

  • legion.model.yaml file in root folder - contains meta information about type of binary model and other model related information (e.g. language, import endpoints, dependencies).

  • Additional folders and files, depend on meta information declared in legion.model.yaml.

Meta information file legion.model.yaml

This file, as previously stated, contains all information about model and additional folders and files. This file is in YAML format.

Structure of file:

  • binaries - section that declares what environment (language and dependencies mechanisms) should be used for loading this model binaries.

  • binaries.type - name of supported Legion Model Environments. Please see section Legion Model Environments.

  • binaries.dependencies - name of dependency management system, that is compatible with Legion Model Environments chosen in binaries.type.

  • binaries.<additional> - additional values for Model Environment, dependency management system (such as path to requirements file).

  • model - section that describes where model artifacts is. Model artifact format depends on Legion Model Environment.

  • model.name - name of model, [a-Z0-9\-]+ string that does not start from digit.

  • model.version - version of model. Format is <Apache Version>-<Additional suffix>, where Additional suffix is a [a-Z0-9\-\.]+ string.

  • model.workDir - working directory to start model from.

  • model.entrypoint - name of model artifact (e.g. Python module or Java JAR file).

  • legionVersion - version of Legion Model Artifact format.

  • toolchain - section that describes toolchain used for training and preparing Legion Model Artifact.

  • toolchain.name - name of used toolchain.

  • toolchain.version - version of used toolchain.

  • toolchain.<additional> - additional fields, related to used toolchain (e.g. used submodule of toolchain).

Examples:

Example with GPPI using conda for dependency management, mlflow toolchain.

binaries:
  type: python
  dependencies: conda
  conda_path: mlflow/model/mlflow_env.yml
model:
  name: wine-quality
  version: 1.0.0-12333122
  workDir: mlflow/model
  entrypoint: entrypoint
legionVersion: '1.0'
toolchain:
  name: mlflow
  version: 1.0.0

Legion Model Environments

Nowadays, legion supports next kinds of model environments:

  • General Python Prediction Interface (GPPI), that provides an ability to import trained model as a python module and use one of predefined function for predicting. Value for binaries.type is python.

  • General Java Prediction Interface (GJPI), that provides an ability to import trained model as a Java Library and use one of predefined interfaces for predicting. Value for binaries.type is java.

Legion's General Python Prediction Interface (GPPI)

General information

Field Value
Name General Python Prediction Interface (GPPI)
Supported languages Python 3.6+
binaries.type "python"
binaries.dependencies "conda"
binaries.conda_path Path to conda env, from artifact root
model.workDir Working directory, PYTHON PATH
model.entrypoint Python import, relative to model.workDir

Description

This interface is representable as importable Python module with declared interface (functions with arguments and return types). Toolchains, that want to save models in this format, have to provide entrypoint with this interface as a python module, or, they may provide a wrapper around their interface for this interface.

ENV. variables required

  • MODEL_LOCATION -- path to model's file, relative to working directory.

Interface declaration

Interface consists of next functions:

Function Description
init Required. Is being invoked during service boot. Returns prediction mode: object-based or matrix-based.
predict_on_objects Optional. Make prediction based on input objects. Return type is configurable.
get_object_input_type Optional. Get type of input for predict_on_objects. List of dicts used if is not provided.
get_object_output_type Optional. Get output type of predict_on_objects. Otherwise they returns JSON-serializable list of dicts.
predict_on_matrix Optional. Make prediction based on matrix with values (tuple of tuples). Accepts names of columns. Returns matrix.
get_output_json_serializer Optional. Is used for output serialization if declared. Default is used otherwise.
get_info Optional. Returns OpenAPI description of input and output types if it is possible.