Highlights from this release include: 

  • Update Data Loader to aeon for flexible, multi-threaded data loading and transformations. More information can be found in the docs, but in brief, aeon:
    • provides an easy interface to adapt existing models to your own, custom, datasets
    • supports images, video and audio and is easy to extend with your own providers for custom modalities
    • is designed to be efficient handling large datasets and loading and augmenting data with minimal latency
  • Neural Machine Translation model
  • Remove Fast RCNN model (use Faster RCNN model instead)
  • Fix super blocking for small N with 1D conv
  • Fix update-direct conv kernel for small N
  • Add gradient clipping to Adam optimizer
  • Documentation updates and bug fixes

As always, you can grab this release from github at: https://github.com/NervanaSystems/neon