We are excited to announce availability of PowerAI 3.4. The focus in this release was to provide upgraded applications to stay current with the evolving Deep Learning frameworks; and to increase ease of use, simplify installation and help developers get started with developing their own cognitive applications.
We’re also publishing the initial version of the PowerAI Cluster Deployment Kit. In this release, we are introducing cluster hardware deployment kits to simplify and accelerate PowerAI deployments in a cluster. This guide provides detailed hardware configuration, deployment, and provisioning.
PowerAI 3.4 continues to expand the TensorFlow environment on POWER with the release of TensorFlow 1.0.1. Building on the rich environment we released with TensorFlow 1.0, this release adds several new features:
- Hadoop HDFS support for improved high-performance access to distributed data;
- Integration with the NCCL communications library to take even more advantage of the high-performance NVLink connections available in the Minsky server; and
- many enhancements to the experimental TensorFlow XLA model compiler.
PowerAI includes a major upgrade of Theano to Theano 0.9. Theano 0.9 adds support for the gpuarray backend (e.g.
THEANO_FLAGS=device=cuda0). The new Theano build continues to support the deprecated old GPU backend (e.g.
THEANO_FLAGS=device=gpu), but support may be phased out in the future by the Theano developer community.
DIGITS now comes preconfigured to work with both Caffe and Torch out of the box. In addition, the DIGITS package supports the use of python installed plugins to provide additional features when using the DIGITS server page. These plugins are included in the PowerAI distribution and can be configured with the Python pip installer.
To help developers jumpstart the process of building cognitive applications, this release includes many sample applications to give developers a head start with building their own applications.
TensorFlow: The TensorFlow team provides example models on GitHub at https://github.com/tensorflow/models. TensorFlow 1.0 marked a major milestone in the evolution of TensorFlow with the release of a new developer API. Not all TensorFlow models in the TensorFlow repo have been upgraded to the TF 1.0 API. For the popular Inception network, we have upgraded and published
inception/imagenet_train on github to use the new API. This model can be found at https://github.com/ibmsoe/tensorflow-models in the branch
Caffe: Each Caffe package includes example scripts and sample models as part of the distribution.
Torch: The Torch package includes example scripts and samples models as part of the distribution. These are customized version of the popular Imagenet examples from https://github.com/soumith/imagenet-multiGPU.torch.
For more details, access to the deployment kit and to download the latest code, check out the IBM PowerAI Release 3.4 page. To support the rapidly growing deep learning developer community using PowerAI, we are launching an expanded IBM Deep Learning and PowerAI Developer site.
So, get started with PowerAI to develop cognitive applications on Power! Share how you are unleashing the power of deep learning to transform the future of computing in the comments section.