Faster Machine Learning with Custom Built TensorFlow on Eagle

TensorFlow is a widely used and powerful symbolic math library commonly used for a variety of machine learning techniques. TensorFlow has built in API support for regression, clustering, classification, hidden Markov models, neural networks, reinforcement learning, as well as, a variety of activation functions, loss function, and optimizers. TensorFlow has received growing adoption among scientists, researchers, and industry professionals for its broad applicability and flexibility.

TensorFlow versions obtained from pip or conda installs may not be optimized for the CPU and GPU architectures found on Eagle. To address this, pre-compiled versions which are optimized both for the CPU and GPU architectures have been created and offer computational benefits compared to other installation approaches. These versions can easily be installed from the wheels provided in /nopt/nrel/apps/wheels/ which contains different TensorFlow versions.

Here is an example of how you can install an optimized version of TensorFlow to your environment.

pip install --upgrade --no-deps --force-reinstall /nopt/nrel/apps/wheels/tensorflow-2.4.0-cp38-cp38-linux_x86_64.whl
These builds provide a significant advantage as illustrated below over the standard conda install of TensorFlow. Benchmark image

A recent tutorial was given on this topic, for more information see the recording or checkout the tutorial materials