One of the most challenging tasks for a data scientist is optimizing their choice of model hyperparameters, the knobs they can tune to pick the best model within a model class. This optimization can be resource- and labour-intensive, and often relies on time-consuming hand-crafted or brute-force approaches. Cray is adding hyperparameter optimization (HPO) algorithms, capable of running in a distributed fashion, to help data scientists find the optimal model for production AI deployments.
"Developing AI models can be a complex, time-consuming process. By offering intelligent hyperparameter optimization support within our Urika-CS and Urika-XC suites, we're giving data scientists pre-set algorithms to quickly identify the most favorable machine learning model designs", stated Per Nyberg, vice president market development, AI and cloud at Cray. "Offering data scientists this support has several benefits, including increased productivity and workflow efficiencies."
The new Urika suites are augmented with four HPO strategies - two commonly used strategies and two strategies developed by Cray to take advantage of the parallelism available in a distributed system. Taken together, these strategies simplify the task of finding and tuning the optimal model for a given application.
The four strategies are:
Cray has added four popular analytics and deep learning frameworks to its Urika AI and Analytics suites: PyTorch, Keras and Horovod for model development and training, as well as the highly-scalable Programming Big Data in R (pbdR) package. The upgraded software suites provide researchers and data science teams the right tools and more choice for how they perform their analytics, machine learning and deep learning workflows.
The new versions of the Urika-CS AI and Analytics software suite and Urika-XC software suite are expected to be available within 90 days.
To learn more and to see a live demo of these new capabilities, you can stop by the Cray booth #2413 at SC18.