TensorFlow is an open-source framework for Machine Learning intelligence created by Google. It supports deep-learning and general numerical computations on CPUs, GPUs, and clusters of GPUs. It is subject to the terms and conditions of the Apache 2.0 License.
In the sections below, we provide guidance on installing TensorFlow on Databricks and give an example of running TensorFlow programs. See Integrating Deep Learning Libraries with Apache Spark for an example of integrating a deep learning library with Spark.
This guide is not a comprehensive guide on TensorFlow. Refer to the TensorFlow website.
TensorFlow is included in Databricks Runtime ML (Beta), a machine learning runtime that provides a ready-to-go environment for machine learning and data science. Instead of installing TensorFlow using the instructions below, you can simply create a cluster using Databricks Runtime ML. See Databricks Runtime ML (Beta).
- On CPUs, use the
- On GPUs, use the
TensorBoard is TensorFlow’s suite of visualization tools for debugging, optimizing, and understanding TensorFlow programs.
TensorBoard is not supported on Community Edition accounts.
To run TensorBoard on your Databricks cluster, you must update the Databricks security group in your AWS account to give ingress access to incoming TensorBoard connections. You will need to specify which IP addresses are allowed to connect to TensorBoard. You can give access to an individual IP address or provide a range that represents your entire office IP range. You or your admin only need to complete this step once. To set it up:
In your AWS console, find the Databricks security group. It will have a label similar to
<databricks-instance>-worker-unmanaged. For example,
Edit the security group and add an inbound TCP rule to allow port
6006to worker machines. It can be a single IP address of your machine or a range. Make sure your laptop and office allows sending TCP traffic on port
Once you’ve enabled incoming connections to TensorBoard, you can start TensorBoard directly from your notebook using a single command.
log_dir = "/tmp/tensorflow_log_dir" dbutils.tensorboard.start(log_dir)
This command displays a link that, when clicked, opens TensorBoard in a new tab. Make sure
to use the same log directory when you start TensorBoard and when you run your TensorFlow program.
We recommend you use a local directory on the driver, for example
store your log files for the best performance (and copy to persistent storage as needed).
TensorBoard will continue to run until you either use
dbutils.tensorboard.stop() to stop it or until you shut down your cluster.
To test and migrate single-machine TensorFlow workflows, you can start with a driver-only cluster on Databricks by setting the number of workers to zero. Though Apache Spark is not functional under this setting, it is a cost-effective way to run single-machine TensorFlow workflows. This example shows how you can run TensorFlow, with TensorBoard monitoring on a driver-only cluster.
spark-tensorflow-connector is a library within the TensorFlow ecosystem that enables conversion between Spark DataFrames and TFRecords (a popular format for storing data for TensorFlow). With spark-tensorflow-connector, you can use Spark DataFrame APIs to read TFRecords files into DataFrames and write DataFrames as TFRecords.
The spark-tensorflow-connector library is included in Databricks Runtime ML (Beta), a machine learning runtime that provides a ready-to-go environment for machine learning and data science. Instead of installing the library using the instructions below, you can simply create a cluster using Databricks Runtime ML. See Databricks Runtime ML (Beta).
To use spark-tensorflow-connector on Databricks, you’ll need to build the project JAR locally, upload it to Databricks, and attach it to your cluster as a library.
Ensure you have Maven in your PATH (see the Maven installation instructions if needed).
Clone the TensorFlow ecosystem repository and cd into the
git clone https://github.com/tensorflow/ecosystem cd ecosystem/spark/spark-tensorflow-connector
Follow the instructions in the README to build the project locally. For the build to succeed, you may need to modify the test configuration so that tests run serially. You can do this by adding a
<configuration>tag to the
<configuration> <parallel>false</parallel> </configuration>
The build command prints the path of the spark-tensorflow-connector JAR, for example:
Installing /Users/<yourusername>/ecosystem/spark/spark-tensorflow-connector/target/spark-tensorflow-connector_2.11-1.6.0.jar to /Users/<yourusername>/.m2/repository/org/tensorflow/spark-tensorflow-connector_2.11/1.6.0/spark-tensorflow-connector_2.11-1.6.0.jar
Upload this JAR to Databricks as a library and attach it to your cluster. You should now be able to run the example notebook (adapted from the spark-tensorflow-connector usage examples):