MXNet Release 18.08

The NVIDIA container image of MXNet, release 18.08, is available.

Contents of MXNet

This container image contains the complete source of the version of MXNet in /opt/mxnet. It is pre-built and installed to the Python path.

The container also includes the following:

Driver Requirements

Release 18.08 is based on CUDA 9, which requires NVIDIA Driver release 384.xx.

Key Features and Enhancements

This MXNet release includes the following key features and enhancements.
  • MXNet container image version 18.08 is based on 1.2.0, with all upstream changes from the MXNet master branch up to and including PR 11545.
  • Latest version of cuDNN 7.2.1.
  • Latest version of DALI 0.1.2 Beta.
  • New demonstrator of increased mixed-precision ResNet-50 training speeds on Volta when processed end-to-end in the NHWC data layout. We are working to PR the code improvements to upstream MXNet. To evaluate in the meantime, type /opt/mxnet/examples/image_classification/train_imagenet_runner --batch-size N . Substitute 256 for N on systems with GPUs having 32GB global memory (or 192 with 16GB GPUs) and prepare the imagenet database as directed in nvidia-examples/imagenet_preparations. Training images should be pre-resized to 480px shorter side and validation should be pre-resized to 256px shorter side. The script expects RecordIO files to be present in the /data/imagenet/train-480-val-256-recordio/ directory.
  • Ubuntu 16.04 with July 2018 updates

Announcements

Starting with the next major version of the CUDA release, we will no longer provide updated Python 2 containers and will only update Python 3 containers.

Known Issues

The multi-threaded nature of MXNet model execution may result in a variable maximum usage of GPU global memory. Users that experience sporadic out-of-GPU-memory errors should experiment with setting the environment variable MXNET_GPU_WORKER_NTHREADS=1 as a possible remedy. We anticipate the need for this experimentation will be removed in a subsequent release.