TensorFlow
TensorFlow is an open source software library for numerical computation using data flow graphs. Nodes in the graph represent mathematical operations, while the graph edges represent the multidimensional data arrays (tensors) communicated between them. The flexible architecture allows you to deploy computation to one or more CPUs or GPUs in a desktop, server, or mobile device with a single API. TensorFlow was originally developed by researchers and engineers working on the Google Brain Team within Google’s Machine Intelligence research organization for the purposes of conducting machine learning and deep neural networks research, but the system is general enough to be applicable in a wide variety of other domains as well.
About TensorFlow on Bessemer
Note
A GPU-enabled worker node must be requested in order to use the GPU version of this software. See Using GPUs on Bessemer for more information.
As TensorFlow and all its dependencies are written in Python, it can be installed locally in your home directory. The use of Anaconda (Python) is recommended as it is able to create a virtual environment in your home directory, allowing the installation of new Python packages without needing admin permissions.
Installation in Home Directory - CPU Version
In order to to install to your home directory, Conda is used to create a virtual python environment for installing your local version of TensorFlow.
First request an interactive session, e.g. with Interactive Jobs.
Then TensorFlow can be installed by the following:
# Load the conda module
module load Anaconda3/5.3.0
# Create an conda virtual environment called 'tensorflow'
conda create -n tensorflow python=3.6
# Activate the 'tensorflow' environment
source activate tensorflow
pip install tensorflow
Every Session Afterwards and in Your Job Scripts
Every time you use a new session or within your job scripts, the modules must be loaded and Conda must be activated again. Use the following command to activate the Conda environment with TensorFlow installed:
module load Anaconda3/5.3.0
source activate tensorflow
Installation in Home Directory - GPU Version
The GPU version of TensorFlow is a distinct Pip package and is also dependent on CUDA and cuDNN libraries, making the installation procedure slightly different.
Warning
You will need to ensure you load CUDA and cuDNN modules which are compatible with the version of TensorFlow used (see table).
First request an interactive session, e.g. see Interactive use of the GPUs.
Then GPU version of TensorFlow can be installed by the following
# Load the conda module
module load Anaconda3/5.3.0
# Load the CUDA and cuDNN module
module load cuDNN/7.6.4.38-gcccuda-2019b
# Create an conda virtual environment called 'tensorflow-gpu'
conda create -n tensorflow-gpu python=3.6
# Activate the 'tensorflow-gpu' environment
source activate tensorflow-gpu
# Install GPU version of TensorFlow
pip install tensorflow-gpu
To install a version of tensorflow-gpu other than the latest version
you should specify a version number when running pip install i.e.
pip install tensorflow-gpu==<version_number>
Every Session Afterwards and in Your Job Scripts
Every time you use a new session or within your job scripts, the modules must be loaded and Conda must be activated again. Use the following command to activate the Conda environment with TensorFlow installed:
module load Anaconda3/5.3.0
module load cuDNN/7.6.4.38-gcccuda-2019b
source activate tensorflow-gpu
Testing your TensorFlow installation
You can test that TensorFlow is running on the GPU with the following Python code (requires TensorFlow >= 2):
import tensorflow as tf
tf.debugging.set_log_device_placement(True)
# Creates a graph
# (ensure tensors placed on the GPU)
with tf.device('/device:GPU:0'):
a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')
b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')
c = tf.matmul(a, b)
# Runs the op.
print(c)
Which when run should give the following results:
[[ 22. 28.]
[ 49. 64.]]
CUDA and cuDNN Import Errors
TensorFlow releases depend on specific versions of both CUDA and cuDNN.
If the wrong cuDNN module is loaded, you may receive ImportError runtime errors such as:
ImportError: libcublas.so.10.0: cannot open shared object file: No such file or directory
This indicates that TensorFlow was expecting to find CUDA 10.0 (and an appropriate version of cuDNN) but was unable to do so.
The following table shows the which module to load for the various versions of TensorFlow, based on the tested build configurations.
TensorFlow |
CUDA |
cuDNN |
cuDNN module to load |
|---|---|---|---|
2.12.0 |
11.8 |
>= 8.6 |
|
2.4.0 |
11.0 |
>= 8.0 |
|
2.3.0 |
10.1 |
>= 7.6 |
|
2.1.0 |
10.1 |
>= 7.6 |
|
2.0.0 |
10.0 |
>= 7.4 |
|
1.14.0 |
10.0 |
>= 7.4 |
|
1.13.1 |
10.0 |
>= 7.4 |
|
>= 1.5.0 |
9.0 |
7 |
N/A |
>= 1.3.0 |
8.0 |
6 |
N/A |
>= 1.0.0 |
8.0 |
5.1 |
N/A |
Training
The Research Software Engineering team has an introductory workshop on deep learning with the TensorFlow Keras framework <https://rses-dl-course.github.io/>__.