Microsoft Cognitive Toolkit Release 18.01

The NVIDIA container image of Microsoft Cognitive Toolkit, release 18.01, is available.

Microsoft Cognitive Toolkit container image version 18.01 is based on Microsoft Cognitive Toolkit 2.3.1.

Contents of the Microsoft Cognitive Toolkit

This container image contains the source files that are used to build the Microsoft Cognitive Toolkit in the /opt/cntk directory. The Microsoft Cognitive Toolkit binaries are pre-built and installed into the /usr/local/cntk directory.

A Python 3.4 Conda environment (including the Microsoft Cognitive Toolkit module) is installed in the /opt/conda/envs/cntk-py34 directory. This is included in PATH, so by default the python command executes a Python 3.4 interpreter. To use the base Ubuntu Python 2.7 binary, you can either adjust PATH or explicitly call /usr/bin/python. This image of the Microsoft Cognitive Toolkit 2.3.1, is built with the following options enabled:

• Async SGD
• Python 3 API
• OpenCV
• Kaldi

The container also includes the following:

• Ubuntu 16.04 including Python 3.4 environment
• NVIDIA CUDA 9.0.176 including CUDA^® Basic Linear Algebra Subroutines library™ (cuBLAS) 9.0.282
• NVIDIA CUDA^® Deep Neural Network library™ (cuDNN) 7.0.5
• NCCL 2.1.2 (optimized for NVLink™ )
• OpenMPI™ 3.0.0

Driver Requirements

Release 18.01 is based on CUDA 9, which requires NVIDIA Driver release 384.xx.

Key Features and Enhancements

This Microsoft Cognitive Toolkit release includes the following key features and enhancements.

• Preview mixed precision training support and ResNet-50 ImageNet example
• Latest version of cuBLAS
• Latest version of cuDNN
• Latest version of NCCL
• Ubuntu 16.04 with December 2017 updates

Known Issues

cuBLAS 9.0.282 regresses RNN seq2seq FP16 performance for a small subset of input sizes. As a workaround, revert back to the 11.12 container.