TensorFlow integration and directory structure

This page covers importing Tensorflow models into Vertica, making predictions on data in the Vertica database, and exporting the model for use on another Vertica cluster or a third-party platform.

For a start-to-finish example through each operation, see TensorFlow example.

Vertica supports models created with either TensorFlow 1.x and 2.x, but 2.x is strongly recommended.

To use TensorFlow with Vertica, install the TFIntegration UDX package on any node. You only need to do this once:

$ /opt/vertica/bin/admintools -t install_package -d database_name -p 'password' --package TFIntegration

Directory and file structure for TensorFlow models

Before importing your models, you should have a separate directory for each model that contains one of each of the following. Note that Vertica uses the directory name as the model name when you import it:

• model_name.pb: a trained model in frozen graph format

• tf_model_desc.json: a description of the model

You can generate both files for a given TensorFlow 2 model with the script freeze_tf2_model.py included in Machine-Learning-Examples/TensorFlow repository (or in opt/vertica/packages/TFIntegration/examples), though you may need to make modifications to the description based on your use-case and how you want your Vertica database to interact with your model:

$ python3 Tensorflow/freeze_tf2_model.py your/model/directory

For example, the tf_models directory contains two models, tf_mnist_estimator and tf_mnist_keras:

tf_models/
├── tf_mnist_estimator
│   ├── mnist_estimator.pb
│   └── tf_model_desc.json
└── tf_mnist_keras
    ├── mnist_keras.pb
    └── tf_model_desc.json

tf_model_desc.json

The tf_model_desc.json file forms the bridge between TensorFlow and Vertica. It describes the structure of the model so that Vertica can correctly match up its inputs and outputs to input/output tables.

Notice that the freeze_tf2_model.py script automatically generates this file for your model, and this generated file can often be used as-is. For more complex models or use cases, you might have to edit this file. For a detailed breakdown of each field, see tf_model_desc.json overview.

Importing TensorFlow models into Vertica

To import TensorFlow models, use IMPORT_MODELS with the category 'TENSORFLOW'.

Import a single model. Keep in mind that the Vertica database uses the directory name as the model name:

select IMPORT_MODELS ( '/path/tf_models/tf_mnist_keras' USING PARAMETERS category='TENSORFLOW');
 import_models
---------------
 Success
(1 row)

Import all models in the directory (where each model has its own directory) with a wildcard (*):

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

Predicting with an imported TensorFlow model

After importing your TensorFlow model, you can use the model to predict on data in a Vertica table.

The PREDICT_TENSORFLOW function is different from the other predict functions in that it does not accept any parameters that affect the input columns such as "exclude_columns" or "id_column"; rather, the function always predicts on all the input columns provided. However, it does accept a num_passthru_cols parameter which allows the user to "skip" some number of input columns, as shown below.

The OVER(PARTITION BEST) clause tells Vertica to parallelize the operation across multiple nodes. See Window partition clause for details:

=> select PREDICT_TENSORFLOW (*
                   USING PARAMETERS model_name='tf_mnist_keras', num_passthru_cols=1)
                   OVER(PARTITION BEST) FROM tf_mnist_test_images;

--example output, the skipped columns are displayed as the first columns of the output
 ID | col0 | col1 | col2 | col3 | col4 | col5 | col6 | col7 | col8 | col9
----+------+------+------+------+------+------+------+------+------+------
  1 |    0 |    0 |    1 |    0 |    0 |    0 |    0 |    0 |    0 |    0
  3 |    1 |    0 |    0 |    0 |    0 |    0 |    0 |    0 |    0 |    0
  6 |    0 |    0 |    0 |    0 |    1 |    0 |    0 |    0 |    0 |    0
...

Exporting TensorFlow models

Vertica exports the model as a frozen graph, which can then be re-imported at any time. Keep in mind that models that are saved as a frozen graph cannot be trained further.

Use EXPORT_MODELS to export TensorFlow models. For example, to export the tf_mnist_keras model to the /path/to/export/to directory:

=> SELECT EXPORT_MODELS ('/path/to/export/to', 'tf_mnist_keras');
 export_models
---------------
 Success
(1 row)

When you export a TensorFlow model, the Vertica database creates and uses the specified directory to store files that describe the model:

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

The .pb and tf_model_desc.json files describe the model, and the rest are organizational files created by the Vertica database.

See also

• TensorFlow models
• TensorFlow example
• IMPORT_MODELS
• EXPORT_MODELS
• PREDICT_TENSORFLOW