Excited about MXNet joining Apache!
Post by Dr. Matt Wood
From Alexa to Amazon Go, we use deep learning extensively across all areas of Amazon, and we’ve tried a lot of deep learning engines along the way. One has emerged as the most scalable, efficient way to perform deep learning, and for these reasons, we have selected MXNet as our engine of choice at Amazon.
MXNet is an open source, state of the art deep learning engine, which allows developers to build sophisticated, custom artificial intelligence systems. Training these systems is significantly faster in MXNet, due to its scale and performance. For example, for the popular image recognition network, Resnet, MXNet has 2X the throughput compared to other engines, letting you train equivalent models in half the time. MXNet also shows close to linear scaling across hundreds of GPUs, while the performance of other engines show diminishing returns at scale.
We have a significant team at Amazon working with the MXNet community to continue to evolve it. The team proposed MXNet joining the Apache Incubator to take advantage of the Apache Software Foundation’s process, stewardship, outreach, and community events. We’re excited to announce that it has been accepted.
We’re at the start what we’ll be investing in Apache MXNet, and look forward to partnering with the community to keep extending its already significant utility.
If you’d like to get started with MXNet, take a look at the keynote presentation I gave at AWS Re:Invent, and fire up an instance (or an entire cluster), of the AWS Deep Learning AMI which includes MXNet with example code, pre-compiled and ready to rock. You should also watch Leo’s presentation and tutorial on recommendation modeling.