TensorFlow example

Vertica uses the TFIntegration UDX package to integrate with TensorFlow.

Vertica uses the TFIntegration UDX package to integrate with TensorFlow. You can train your models outside of your Vertica database, then import them to Vertica and make predictions on your data.

TensorFlow scripts and datasets are included in the GitHub repository under Machine-Learning-Examples/TensorFlow.

The example below creates a Keras (a TensorFlow API) neural network model trained on the MNIST handwritten digit classification dataset, the layers of which are shown below.

The data is fed through each layer from top to bottom, and each layer modifies the input before returning a score. In this example, the data passed in is a set of images of handwritten Arabic numerals and the output would be the probability of the input image being a particular digit:

inputs = keras.Input(shape=(28, 28, 1), name="image")
x = layers.Conv2D(32, 5, activation="relu")(inputs)
x = layers.MaxPooling2D(2)(x)
x = layers.Conv2D(64, 5, activation="relu")(x)
x = layers.MaxPooling2D(2)(x)
x = layers.Flatten()(x)
x = layers.Dense(10, activation='softmax', name='OUTPUT')(x)
tfmodel = keras.Model(inputs, x)

For more information on how TensorFlow interacts with your Vertica database and how to import more complex models, see TensorFlow integration and directory structure.

Prepare a TensorFlow model for Vertica

The following procedures take place outside of Vertica.

Train and save a TensorFlow model

  1. Install TensorFlow 2.

  2. Train your model. For this particular example, you can run train_simple_model.py to train and save the model. Otherwise, and more generally, you can manually train your model in Python and then save it:

    $ mymodel.save('my_saved_model_dir')

  3. Run the freeze_tf2_model.py script included in the Machine-Learning-Examples repository or in opt/vertica/packages/TFIntegration/examples, specifying your model and an output directory (optional, defaults to frozen_tfmodel).

    This script transforms your saved model into the Vertica-compatible frozen graph format and creates the tf_model_desc.json file, which describes how Vertica should translate its tables to TensorFlow tensors:

    $ ./freeze_tf2_model.py path/to/tf/model frozen_model_dir

Import TensorFlow models and make predictions in Vertica

  1. If you haven't already, as dbadmin, 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

  2. Copy the the directory to any node in your Vertica cluster and import the model:

    => SELECT IMPORT_MODELS('path/to/frozen_model_dir' USING PARAMETERS category='TENSORFLOW');

  3. Import the dataset you want to make a prediction on. For this example:

    1. Copy the Machine-Learning-Examples/TensorFlow/data directory to any node on your Vertica cluster.

    2. From that data directory, run the SQL script load_tf_data.sql to load the MNIST dataset:

      $ vsql -f load_tf_data.sql

  4. Make a prediction with your model on your dataset with PREDICT_TENSORFLOW. In this example, the model is used to classify the images of handwritten numbers in the MNIST dataset:

    => 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
...

  5. To export the model with EXPORT_MODELS:

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

TensorFlow 1 (deprecated)

  1. Install TensorFlow 1.15 with Python 3.7 or below.

  2. Run train_save_model.py in Machine-Learning-Examples/TensorFlow/tf1 to train and save the model in TensorFlow and frozen graph formats.

See also