EXPORT_MODELS

Exports machine learning models.

Exports machine learning models. Vertica supports three model formats:

  • Native Vertica (VERTICA_MODELS)

  • PMML

  • TensorFlow

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

Behavior type

Volatile

Syntax

EXPORT_MODELS ( 'output-dir', 'export-target' [ USING PARAMETERS category = 'model-category' ] )

Arguments

output-dir
Absolute path of an output directory to store the exported models, either an absolute path on the initiator node file system or a URI for a supported file system or object store.
export-target
Models to export: 
 [schema.]{model-name | * }

schema specifies the schema from which models are exported. If omitted, EXPORT_MODELS uses the default schema. Supply * (asterisk) to batch export all models from the schema.

If a model in a batch fails to export, the function issues a warning and then continues to export any remaining models in the batch. Details about any failed model exports are available in the log file generated at the output-dir location.

Parameters

category
The category of models to export, one of the following:

  • VERTICA_MODELS

  • PMML

  • TENSORFLOW

EXPORT_MODELS exports models of the specified category according to the scope of the export operation—that is, whether it applies to a single model, or to all models within a schema. See Export Scope and Category Processing below.

Exported Files below describes the files that EXPORT_MODELS exports for each category.

If you omit this parameter, EXPORT_MODELS exports the model, or models in the specified schema, according to their model type.

Privileges

Superuser or MLSUPERVISOR

Export scope and category processing

EXPORT_MODELS executes according to the following parameter settings:

  • Scope of the export operation: single model, or all models within a given schema

  • Category specified or omitted

The following table shows how these two parameters control the export process:

Export scope If category specified... If category omitted...
Single model Convert the model to the specified category, provided the model and category are compatible; otherwise, return with a mismatch error. Export the model according to model type.
All models in schema Export only models that are compatible with the specified category and issue mismatch warnings on all other models in the schema. Export all models in the schema according to model type.

Exported files

EXPORT_MODELS exports the following files for each model category:

VERTICA_MODELS

  • Multiple binary files (exact number dependent on model type)

  • metadata.json: Metadata file with model information, including model name, category, type, and Vertica version on export.

  • crc.json: Used on import to validate other files of this model.

PMML

  • XML file with the same name as the model and complying with PMML standard.

  • metadata.json: Metadata file with model information, including model name, category, type, and Vertica version on export.

  • crc.json: Used on import to validate other files of this model.

TENSORFLOW

  • model-name.pb: Contains the TensorFlow model, saved in 'frozen graph' format.

  • metadata.json: Metadata file with model information, including model name, category, type, and Vertica version on export.

  • tf_model_desc.json: Summary model description.

  • model.json: Verbose model description.

  • crc.json: Used on import to validate other files of this model.

Categories and compatible models

If EXPORT_MODELS specifies a single model and also sets the category parameter, the function succeeds if the model type and category are compatible; otherwise, it returns with an error. The following model types are compatible with the listed categories:

PMML
PMML
TensorFlow
TENSORFLOW
VERTICA_MODELS
PMML
VERTICA_MODELS

If EXPORT_MODELS specifies to export all models from a schema and sets a category, it issues a warning message on each model that is incompatible with that category. The function then continues to process remaining models in that schema.

EXPORT_MODELS logs all errors and warnings in output-dir/export_log.json.

Examples

Export models without changing their category

Export model myschema.mykmeansmodel without changing its category:

=> SELECT EXPORT_MODELS ('/home/dbadmin', 'myschema.mykmeansmodel');
EXPORT_MODELS
----------------
Success
(1 row)

Export all models in schema myschema without changing their categories:

=> SELECT EXPORT_MODELS ('/home/dbadmin', 'myschema.*');
EXPORT_MODELS
----------------
Success
(1 row)

Export all the models in schema models to an S3 bucket without changing the model categories:

SELECT export_models('s3://vertica/ml_models', 'models.*');
EXPORT_MODELS 
---------------
Success
(1 row)

Export models that are compatible with the specified category

The category is set to PMML. Models of type PMML and VERTICA_MODELS are compatible with the PMML category, so the export operation succeeds if my_keans is of either type:

=> SELECT EXPORT_MODELS ('/tmp/', 'my_kmeans' USING PARAMETERS category='PMML');

The category is set to VERTICA_MODELS. Only models of type VERTICA_MODELS are compatible with the VERTICA_MODELS category, so the export operation succeeds only if my_keans is of that type:

=> SELECT EXPORT_MODELS ('/tmp/', 'public.my_kmeans' USING PARAMETERS category='VERTICA_MODELS');

The category is set to TENSORFLOW. Only models of type TensorFlow are compatible with the TENSORFLOW category, so the model tf_mnist_keras must be of type TensorFlow:

=> SELECT EXPORT_MODELS ('/tmp/', 'tf_mnist_keras', USING PARAMETERS category='TENSORFLOW');
export_models
---------------
Success
(1 row)

After exporting the TensorFlow model tf_mnist_keras, list the exported files:

$ ls tf_mnist_keras/
crc.json  metadata.json  mnist_keras.pb  model.json  tf_model_desc.json

See also

IMPORT_MODELS