DALI Release 0.2 Beta

This DALI 0.2 release is a beta release.

Key Features and Enhancements

This DALI release includes the following key features and enhancements.
  • Added Sphinx based documentation that is in sync with the code on GitHub. For more information, see DALI Master Branch User Guide.

  • Build system has been refined, common errors have meaningful messages, improved localization and version detection of key dependency packages, such as, nvJPEG, libturbo-jpeg, and LMDB.

  • Added Unfused Crop and CropCastPermute operators.

  • Added improvements for TensorFlow plugin (polymorphism and shape argument for the output).

  • Expanded examples of TensorFlow working with different readers, such as, MXNetReader, FileReader, and TFRecordReader.

  • Updated nvJPEG to 0.1.4

  • Added fallback to host decoder when image is not JPEG but PNG instead. For example, n02105855_2933.JPEG from ImageNet.

Breaking API Changes

  • The API for the Resize operator changed to match other similar operators like ResizeCropMirror.

  • The API for the TensorFlow plugin changed to allow specifying the whole shape of the tensor instead of N, H, and W separately; which enables handling both NCHW and NHWC outputs.

  • The type of labels produced by the TensorFlow plugin have changed. In DALI version 0.1.2, it was always tf.float32. In this release, a new optional parameter called label_type is introduced to the TensorFlow plugin to control the type of label. The default value for label_type is tf.int64 to better align with the label type in TFRecord.

Using DALI 0.2 Beta

The 18.08 NVIDIA GPU Cloud (NGC) optimized container for MXNet, PyTorch, and TensorFlow, includes an older version of DALI. To upgrade to DALI 0.2 beta, follow the installation instructions in the DALI Quick Start Guide.

Refer to the DALI Developer Guide for usage details.

Known Issues

  • This is a beta release, therefore, not all functionality is fully supported and working. This beta release is meant for testing and research.

  • The DALI integrated ResNet-50 samples in the 18.08 NGC TensorFlow and PyTorch containers have lower than expected accuracy and performance results. We are working to address the issue in the next release.

  • This is a beta release. All features are expected to be available, however, some aspects of functionality and performance will likely be limited compared to a non-beta release.