How to setup Memgraph TensorFlow Op?

TensorFlow is an open-source software library for high-performance numerical computation. A TensorFlow op (operation) is a fundamental building block of all TensorFlow models.

Memgraph TensorFlow op wraps the high-performance Memgraph client for use with TensorFlow, allowing natural data transfer between Memgraph and TensorFlow at any point in the model.

At this point, we strongly advise you to read the TensorFlow section of our Reference guide.

In this article, we assume you have installed Python 3 and the TensorFlow pip package. See the link for more information. We also assume that you have installed and are running Memgraph (see more)

Memgraph Tensorflow Op usage

Memgraph TensorFlow op is a shared library (.so file). Library name is libmemgraph_op.so.

Load op library:

import tensorflow as tf
memgraph_op_module = tf.load_op_library('libmemgraph_op.so')

Create Memgraph TensorFlow op:

# Create Memgraph op, and put placeholders for input
memgraph_op = memgraph_op_module.memgraph_op(query_holder,
                                             input_list_holder,
                                             output_dtype=tf.int64)

Where query_holder and input_list_holder are TensorFlow placeholders.

Computation:

# Run Memgraph op
output = sess.run(memgraph_op, {query_holder: query,
                                input_list_holder: input_list})

The output is a tuple, where the first element is the header and the second element is a result matrix.

Example

Here is a simple example. You can use the movie dataset or you can use this simple dataset:

CREATE (:User {id: 1})-[:Rating {score:5.0}]->(:Movie {id: 1});
CREATE (:User {id: 2})-[:Rating {score:2.5}]->(:Movie {id: 3});
CREATE (:User {id: 3})-[:Rating {score:4.5}]->(:Movie {id: 8});
CREATE (:User {id: 4})-[:Rating {score:1.0}]->(:Movie {id: 12});
CREATE (:User {id: 5})-[:Rating {score:4.5}]->(:Movie {id: 33});
CREATE (:User {id: 42})-[:Rating {score:1.0}]->(:Movie {id: 42});

This example assumes that Memgraph is running on 127.0.0.1:7687 without ssl. If you want to change this, use op attributes.

import tensorflow as tf

# Load libmemgraph_op.so
memgraph_op_module = tf.load_op_library('libmemgraph_op.so')


def main():
    query = """match (u :User)-->(m :Movie)
              where u.id in $input_list
              return u.id, m.id;"""

    # Input list used in query
    input_list = [1, 2, 3, 4, 5]

    # Create tensorflow session
    with tf.Session() as sess:

        # Query placeholder
        query_holder = tf.placeholder(tf.string)

        # Input list placeholder
        input_list_holder = tf.placeholder(tf.int64)

        # Create Memgraph op, and put placeholders for input
        memgraph_op = memgraph_op_module.memgraph_op(query_holder,
                                                     input_list_holder,
                                                     output_dtype=tf.int64)

        # Run Memgraph op
        output = sess.run(memgraph_op, {query_holder: query,
                                        input_list_holder: input_list})

        # First output is list of headers
        print("Headers:")
        for i in output[0]:
            print(i)

        # Output matrix (rows), query results
        print("Rows: ")
        for i in output[1]:
            print(i)

if __name__ == "__main__":
    main()

Memgraph Parallel Tensorflow Op Usage

Load op library:

import tensorflow as tf
memgraph_op_module = tf.load_op_library('libmemgraph_op.so')

Create Memgraph TensorFlow op:

# Create Memgraph op, and put placeholders for input
memgraph_op = memgraph_op_module.parallel_memgraph_op(query_holder,
                                                      input_list_holder,
                                                      output_dtype=tf.int64,
                                                      num_workers=4)

Where query_holder and input_list_holder are TensorFlow placeholders.

Computation:

# Run Memgraph op
output = sess.run(memgraph_op, {query_holder: query,
                                input_list_holder: input_list})

The output is a tuple, where the first element is the header and the second element is a result matrix.

Example

This example shows one of the archetypal patterns of using the parallel op. We will find nodes by ids and return each of their features.

We will query this example dataset:

CREATE (n:Node {id: 1, features: [100, 115, 121, 95, 72, 142]});
CREATE (n:Node {id: 2, features: [45, 125, 212, 46, 25, 92]});
CREATE (n:Node {id: 3, features: [34, 74, 261, 194, 142, 37]});
CREATE (n:Node {id: 4, features: [76, 92, 11, 16, 78, 261]});
CREATE (n:Node {id: 5, features: [175, 63, 111, 192, 58, 91]});
CREATE (n:Node {id: 6, features: [251, 184, 43, 57, 243, 231]});
CREATE (n:Node {id: 7, features: [187, 136, 37, 33, 76, 145]});
CREATE (n:Node {id: 8, features: [193, 195, 200, 74, 28, 127]});

This example assumes that Memgraph is running on 127.0.0.1:7687 without ssl. If you want to change this, use op attributes.

import tensorflow as tf

# Load libmemgraph_op.so
memgraph_op_module = tf.load_op_library('libmemgraph_op.so')


def main():
    query = """
            UNWIND $input_list AS idx
            MATCH (n:Node {id: idx})
            RETURN n.features
            """

    # Input list used in query
    input_list = [1, 2, 3, 4, 5]

    # Create tensorflow session
    with tf.Session() as sess:

        # Query placeholder
        query_holder = tf.placeholder(tf.string)

        # Input list placeholder
        input_list_holder = tf.placeholder(tf.int64)

        # Create Memgraph op, and put placeholders for input
        memgraph_op = memgraph_op_module.parallel_memgraph_op(query_holder,
                                                     input_list_holder,
                                                     output_dtype=tf.int64)

        # Run Memgraph op
        output = sess.run(memgraph_op, {query_holder: query,
                                        input_list_holder: input_list})

        # First output is list of headers
        print("Headers:")
        for i in output[0]:
            print(i)

        # Output matrix (rows), query results
        print("Rows: ")
        for i in output[1]:
            print(i)

if __name__ == "__main__":
    main()

Where to next?

To learn more about Memgraph's functionalities, visit the Reference guide. For real-world examples of how to use Memgraph, we strongly suggest going through one of the available Tutorials.