Keras is a high-level neural networks application programming interface (API), written in Python and capable of running on top of TensorFlow, CNTK, or Theano. On Rivanna, we provide TensorFlow containers that include the Keras API. Since version 1.12.0, TensorFlow contains its own Keras API implementation as described on the TensorFlow website.

Using Keras with TensorFlow containers

Like TensorFlow itself, Python code that utlizes the Keras package can be run interactively as Jupyter Notebooks, in interactive shell jobs, or non-interctively as Slurm batch jobs. Rivanna provides several nodes with graphics processing units (GPUs) that should be used when running Keras code.

As best practice, it is recommended to create a local copy of the TensorFlow container before using the Keras package in Slurm or interactive jobs.

Example Python Script

The follwing example script can be found on the TensorFlow Tutorials website. Note that this script requires TensorFlow 1.12.0.

import tensorflow as tf
mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

model = tf.keras.models.Sequential([
  tf.keras.layers.Dense(512, activation=tf.nn.relu),
  tf.keras.layers.Dense(10, activation=tf.nn.softmax)

              metrics=['accuracy']), y_train, epochs=5)
model.evaluate(x_test, y_test)

Keras Jupyter Notebooks

The Keras package can be used for interactive code development in Jupyter Notebooks as described in our TensorFlow Jupyter Notebook documentation.

Keras Slurm jobs

The Keras package can be used for Slurm job batch processing as described in our TensorFlow Slurm Job documentation.

Keras Interactive Jobs (ijob)

The Keras package can be used for code development in an interactive shell environment as described in our TensorFlow Interactive Jobs documentation. The ijob is for command-line usage; if you require a graphical user interface, including some form of IDE, see the instructions for Tensorboard or you may use Jupyterlab.

Interaction with the Host File System

The following user directories are overlayed onto each container by default on Rivanna:

  • /home
  • /scratch
  • /nv
  • /project

Due to the overlay, these directories are by default the same inside and outside the container with the same read, write, and execute permissions. This means that file modifications in these directories (e.g. in /home) via processes running inside the container are persistent even after the container instance exits. The /nv and /project directories refer to leased storage locations that may not be available to all users.

For more details review our software and Singularity container documentation.