DALI Release 0.6.1 Beta

The DALI 0.6.1 is a beta release. Hence, for all the features, the functionality and performance will likely be limited.

Key Features and Enhancements

This DALI release includes the following key features and enhancements.

  • Added the DALI TensorFlow plugin to improve out-of-the-box forward compatibility.
  • Added a new package nvidia-dali-tf-plugin that allows user to compile DALI TensorFlow plugin using the locally installed version.
  • Added the below examples:
    • VideoReader
    • PyTorch and MXNet examples with various readers
  • Added a new operator variant gpu box encoder.
  • Added a new JSON parser—now COCO reader start up time is significantly faster (~10x).
  • Added the ability to automatically reset MXNet and PyTorch Python iterators.
  • Enhanced the Video reader to link dynamically with the CUDA Driver API and nvcuvid.
  • Added the ability to deliver exactly 1 epoch from DALIGenericIterator (PyTorch and MXNet).

Breaking API Changes

  • PyTorch iterator returns the exact number of samples per epoch, so the final batch could be smaller if epoch size is not divisible by the batch size. To keep the old behaviour when data is wrapped up, use stop_at_epoch argument.

Using DALI 0.6.1 Beta

The 19.01 NVIDIA GPU Cloud (NGC) optimized container for MXNet, PyTorch, and TensorFlow includes an older version of DALI.

To upgrade to DALI 0.6.1 beta, follow the installation instructions in the DALI Quick Start Guide.

Refer to the DALI Developer Guide for usage details.

Known Issues

  • The New Video reader operator requires NVIDIA VIDEO CODEC SDK support in the platform. The NVIDIA GPU Cloud (NGC) optimized containers lack this functionality in the default configuration prior to the 19.01 version, resulting in problem with loading libnvcuvid.so. To enable it, run the container with the ‘video’ capability enabled, as below: 
    -e "NVIDIA_DRIVER_CAPABILITIES=compute,utility,video"