IMPORT_MODELS

Imports models into Vertica, either Vertica models that were exported with EXPORT_MODELS, or models in Predictive Model Markup Language (PMML) or TensorFlow format. You can use this function to move models between Vertica clusters, or to import PMML and TensorFlow models trained elsewhere.

Other Vertica model management operations such as GET_MODEL_SUMMARY and GET_MODEL_ATTRIBUTE support imported models.

This is a meta-function. You must call meta-functions in a top-level SELECT statement.

Behavior type

Syntax

IMPORT_MODELS ( 'source'
           [ USING PARAMETERS [ new_schema = 'schema-name' ] [, category = 'model-category' ] ] )

Arguments

source

The absolute path of the location from which to import models, one of the following:

• The directory of a single model:

  path/model-directory
  

• The parent directory of multiple model directories:

  parent-dir-path/*
  

Parameters

new_schema

An existing schema where the machine learning models are imported. If omitted, models are imported to the default schema.

IMPORT_MODELS extracts the name of the imported model from its metadata.json file, if it exists. Otherwise, the function uses the name of the model directory.

category

Specifies the category of the model to import, one of the following:

• VERTICA_MODELS

• PMML

• TENSORFLOW

This parameter is required if the model directory has no metadata.json file. IMPORT_MODELS returns with an error if one of the following cases is true:

• No category is specified and the model directory has no metadata.json.

• The specified category does not match the model type.

Note

If the category is TENSORFLOW, IMPORT_MODELS only imports the following files from the model directory:

• model-name.pb

• model-name.json

• model-name.pbtxt (optional)

Privileges

Superuser

Requirements and restrictions

The following requirements and restrictions apply:

• If you export a model, then import it again, the export and import model directory names must match. If naming conflicts occur, import the model to a different schema by using the new_schema parameter, and then rename the model.

• The machine learning configuration parameter MaxModelSizeKB sets the maximum size of a model that can be imported into Vertica.

• Some PMML features and attributes are not currently supported. See PMML features and attributes for details.

• If you import a PMML model with both metadata.json and crc.json files, the CRC file must contain the metadata file's CRC value. Otherwise, the import operation returns with an error.

Examples

Import models into the specified schema:

In both examples no model category is specified, so IMPORT_MODEL uses the model's metadata.json file to determine its category:

• Import a single model mykmeansmodel into the newschema schema:

  => SELECT IMPORT_MODELS ('/home/dbadmin/myschema/mykmeansmodel' USING PARAMETERS new_schema='newschema')
  IMPORT_MODELS
  ----------------
  Success
  (1 row)
  

• Import all models in the myschema directory into the newschema schema:

  => SELECT IMPORT_MODELS ('/home/dbadmin/myschema/*' USING PARAMETERS new_schema='newschema')
  IMPORT_MODELS
  ----------------
  Success
  (1 row)
  

Specify the category of models to import:

In the first two examples, IMPORT_MODELS returns with success only if the specified model and category match; otherwise, it returns an error:

• Import kmeans_pmml as a PMML model:

  SELECT IMPORT_MODELS ('/root/user/kmeans_pmml' USING PARAMETERS category='PMML')
   import_models
  ---------------
   Success
  (1 row)
  

• Import tf_mnist_estimator as a TensorFlow model:

  => SELECT IMPORT_MODELS ( '/path/tf_models/tf_mnist_estimator' USING PARAMETERS category='TENSORFLOW');
   import_models
  ---------------
   Success
  (1 row)
  

• Import all TensorFlow models from the specified directory:

  => SELECT IMPORT_MODELS ( '/path/tf_models/*' USING PARAMETERS category='TENSORFLOW');
   import_models
  ---------------
   Success
  (1 row)
  

See also