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 each of the following files. 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

For example, a tf_models directory that contains two models, tf_mnist_estimator and tf_mnist_keras, has the following layout:

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

You can generate both of these files for a given TensorFlow 2 (TF2) model with the freeze_tf2_model.py script included in the Machine-Learning-Examples GitHub repository and in the opt/vertica/packages/TFIntegration/examples directory in the Vertica database. The script accepts three arguments:

model-path

Path to a saved TF2 model directory.

folder-name

(Optional) Name of the folder to which the frozen model is saved; by default, frozen_tfmodel.

column-type

(Optional) Integer, either 0 or 1, that signifies whether the input and output columns for the model are primitive or complex types. Use a value of 0 (default) for primitive types, or 1 for complex.

For example, the following call outputs the frozen tf_autoencoder model, which accepts complex input/output columns, into the frozen_autoencoder folder:

$ python3 ./freeze_tf2_model.py path/to/tf_autoencoder frozen_autoencoder 1

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 TensorFlow 2 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)

Make predictions with an imported TensorFlow model

After importing your TensorFlow model, you can use the model to predict on data in a Vertica table. Vertica provides two functions for making predictions with imported TensorFlow models: PREDICT_TENSORFLOW and PREDICT_TENSORFLOW_SCALAR.

The function you choose depends on whether you specified a column-type of 0 or 1 when calling the freeze_tf2_model.py script. If column-type was 0, meaning the model accepts primitive input and output types, use PREDICT_TENSORFLOW to make predictions; otherwise, use PREDICT_TENSORFLOW_SCALAR, as your model should accept complex input and output types.

Using PREDICT_TENSORFLOW

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
...

Using PREDICT_TENSORFLOW_SCALAR

The PREDICT_TENSORFLOW_SCALAR function accepts one input column of type ROW, where each field corresponds to an input tensor. It returns one output column of type ROW, where each field corresponds to an output tensor. This complex type support can simplify the process for making predictions on data with many input features.

For instance, the MNIST handwritten digit classification dataset contains 784 input features for each input row, one feature for each pixel in the images of handwritten digits. The PREDICT_TENSORFLOW function requires that each of these input features are contained in a separate input column. By encapsulating these features into a single ARRAY, the PREDICT_TENSORFLOW_SCALAR function only needs a single input column of type ROW, where the pixel values are the array elements for an input field:

--Each array for the "image" field has 784 elements. 
=> SELECT * FROM mnist_train;
id |                   inputs
---+---------------------------------------------
 1 | {"image":[0, 0, 0,..., 244, 222, 210,...]}
 2 | {"image":[0, 0, 0,..., 185, 84, 223,...]}
 3 | {"image":[0, 0, 0,..., 133, 254, 78,...]}
 ...

In this case, the function output consists of a single opeartion with one tensor. The value of this field is an array of ten elements, which are all zero except for the element whose index is the predicted digit:

=> SELECT id, PREDICT_TENSORFLOW_SCALAR(inputs USING PARAMETERS model_name='tf_mnist_ct') FROM mnist_test;
 id |                   PREDICT_TENSORFLOW_SCALAR                      
----+-------------------------------------------------------------------
  1 | {"prediction:0":["0", "0", "0", "0", "1", "0", "0", "0", "0", "0"]} 
  2 | {"prediction:0":["0", "1", "0", "0", "0", "0", "0", "0", "0", "0"]} 
  3 | {"prediction:0":["0", "0", "0", "0", "0", "0", "0", "1", "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