New PowerAI Developer Tools Make Deep Learning Easier to Use

Today, we announced a set of new additions to IBM’s PowerAI deep learning toolkit. PowerAI started off as package of software distributions of many of the major deep learning software frameworks for model training like TensorFlow, Caffe, Torch, Theano, and the associated libraries like cuDNN. The PowerAI software has always been optimized for performance using the NVLink-based Power server, the IBM Power 822LC for HPC ("Minsky"). 

The new additions expand the scope of PowerAI with:

• New data preparation & transformation tools for deep learning
• Cluster orchestration, virtualization & distribution with Apache Spark
• Tools and GUIs to enhance the development experience for data scientists
• Faster training times by distributing deep learning and model tuning across a cluster

AI Software Stack

The figure below shows what the typical AI software and infrastructure stack looks like.  

The AI stack foundation starts with the right hardware: servers with accelerators and the right storage. GPU accelerators are extremely well suited for the compute-intensive nature of deep learning training, and servers with the highest CPU to GPU bandwidth, like IBM’s NVLink server, enable the high performance data transfer required for larger and more complex deep learning models. But all of this first starts with getting the right data.

It Starts with Data

Finding good, labelled data as input is a big challenge for data scientists and deep learning developers. In most cases, data is replicated from existing databases or streamed directly into a data lake from sensors or social media. The data scientist must then get the data ready for machine learning, often by transforming it into the right format that the deep learning distribution can recognize. For example, if you want to train a model with 100 million images, you will have to resize the images to the size that TensorFlow or Caffe require. This data preparation step, classically called ETL, or extract, transform, and load, has become a major pain point for data scientists, since data is coming in from multiple sources and in many different formats.

For deep learning training, data scientists will first choose a good neural network model as a starting point and then tune its hyper-parameters for their selected input data set. This model tuning is an iterative process, wherein the data scientist will try lots of hyper-parameters till they find a trained model with the best accuracy. Each model training run can take days, if not weeks, so improving efficiency and quickly finding the correct hyper-parameters can save a tremendous amount of time.

Building New PowerAI Features for Deep Learning Developers

Part of what you get with PowerAI is enterprise-class support from IBM and our team of deep learning experts. This team works daily with deep learning developers around the world, and we have been listening to their feedback and input to help maintain PowerAI as the leading deep learning toolkit for the enterprise.

As a result, today we are announcing four major additions to PowerAI designed to make it easier for developers and data scientists:

•   AI Vision: A custom application development tool aimed at computer vision workloads. AI Vision enables application developers with little or no experience with deep learning to build a trained deep learning model for different input data sets.

Video demos of AI Vision are available on YouTube. (edit added on July 7th)

•   Apache Spark-based Data Extraction, Transformation and Preparation tool: We enhanced IBM Spectrum Conductor with Spark with a GUI-based set of tools that enable the data scientist to create functions that transform an input data set to the format required by frameworks like TensorFlow or Caffe, making it a cinch to match your data set to your framework. For example, you can have a transform function that resizes images for Caffe, and then simply transform and load any input data set into the right format to use with Caffe. IBM Spectrum Conductor with Spark automatically launches a whole set of Spark jobs on the cluster, each of which resizes a portion of the input data set.

•   DL Insight: Model tuning software that automatically tunes hyper-parameters for models based on input data sets using Spark-based distributed computing. We enhanced IBM Spectrum Conductor with Spark so that it automatically launches multiple model training runs with different hyper-parameters using a subset of the data. It then monitors the training progress and searches and identifies the best hyper-parameters using several different search methods such as random and Bayesian search. To improve usability, DL Insight comes with a powerful and intuitive GUI that visualizes training and provides continuous feedback to quickly create and optimize deep learning models.

•   Distributed Deep Learning: To accelerate the training time, we are adding methods to scale a single training job across a cluster of servers. We have both a MPI-based scaling approach, inspired by high-performance computing methods, as well as a Spark and HPC converged distributed computing model, for clusters with either an ethernet or Infiniband network.

As illustrated by the figure below, several of these enhancements are built on top of the Spectrum Conductor with Spark. It orchestrates a cluster running Apache Spark, and is based on more than a decade of software development on tools like Spectrum Symphony that IBM developed for cluster management and high-performance computing clusters.

These software enhancements to PowerAI will roll-out for general release over the next few months. Contact us or post below if you want early access.

Technical preview for DL Insight in Spectrum Conductor with Spark for deep learning is available now.

You can download PowerAI for your Power Systems S822LC for HPC server, try PowerAI today on Nimbix’s deep learning cloud with NVIDIA NVLink and NVIDIA Tesla P100 GPUs.

If you are a data scientist using deep learning, we are looking for feedback on other software tools that we can add to make your experience better. Please post your comments below.