1. Abstract
The NVIDIA® DIGITS DevBox Software Image is tuned and optimized to deliver maximum performance of the deep learning frameworks on the DIGITS DevBox system, built and shipped by NVIDIA® Corporation.
2. Updates in Refresh 3 (May 2017)
This section lists the changes in the NVIDIA® DIGITS™ DevBox software, Refresh 3 (May 2017):
- NVIDIA DIGITS Software updated to version 5.0
- NVIDIA Linux Display Driver updated to version 375.26
- NVIDIA CUDA toolkit updated to CUDA 8.0.61
- NVIDIA cuDNN updated to cuDNN v5 (version 5.1.10)
- NVIDIA Caffe updated to version 0.15.13 https://github.com/NVIDIA/caffe/releases/tag/v0.15.13
- NVIDIA Torch7 is updated to version 0.9.99
- Theano is updated to version 0.9.0
See the DIGITS DevBox Update Guide for instructions on how to perfrom an over-the-network update.
3. Updates in Refresh 2 (Feb 2016)
This section lists the changes in the NVIDIA DIGITS DevBox software, Refresh 2 (Sept 2016):
- NVIDIA DIGITS Software updated to version 3.0.0-1
- NVIDIA cuDNN library updated to libcuDNN v4 (version 4.0.7)
- NVIDIA Caffe updated to version 0.14.2-1 https://github.com/NVIDIA/caffe/releases/tag/v0.14.2
- NVIDIA Torch7 is updated to version 0.9.92
4. Updates in Refresh 1 (Sept 2015)
This section lists the changes in the NVIDIA DIGITS DevBox software, Refresh 1 (Sept 2015):
- NVIDIA DIGITS Software updated to version 2.0
- NVIDIA Linux Display Driver updated to version 352.41
- NVIDIA CUDA toolkit updated to CUDA 7.5
- NVIDIA cuDNN updated to cuDNN v3 (version 7.0.64)
- Caffe updated to version 0.13.1 https://github.com/NVIDIA/caffe/releases/tag/v0.13.1
- Theano updated to version 0.7.1 https://github.com/Theano/Theano/releases/tag/rel-0.7.1a1
- Torch is updated: https://github.com/torch/distro/tree/master/extra and https://github.com/torch/nn/tree/getParamsByDevice, with cuDNN v3 integration from https://github.com/soumith/cudnn.torch
In-the-Box Software Components
This section provides a list of all software components included in the in-the-box package with the associated installation locations:
- Ubuntu 14.04.2 http://releases.ubuntu.com/14.04/ from Canonical Ltd. Ubuntu and Canonical are registered trademarks of Canonical Ltd.
- NVIDIA Linux Display Driver version 346.63.
- NVIDIA CUDA toolkit versions 7.0 and 6.5. Production toolkit release notes can be found at: http://docs.nvidia.com/cuda/index.html The CUDA toolkits are installed under /usr/local/cuda-7.0/ and /usr/local/cuda-6.5/ folders.
- NVIDIA cuDNN v2 production (version 6.5.48) GPU-accelerated library of primitives for deep neural networks.
- NVIDIA DIGITS which is an interactive environment for training, evaluating, and experimenting with neural networks. Version 1.0.3 is installed under /usr/share/digits, with a web interface accessible at http://localhost. The packaged version of DIGITS is from: https://github.com/NVIDIA/DIGITS/releases/tag/v1.0.3
- Caffe is installed under /usr/bin and the Caffe python interface is under /usr/lib/python2.7/dist-packages. This installation of Caffe is built from source at: https://github.com/NVIDIA/caffe/releases/tag/v0.10.0
- Torch is installed system-wide, with the Torch Lua modules under /usr/local. Torch is packaged from the following commit: https://github.com/torch/distro/commit/e0c565120622f99ef6e1ca7fccca66cfe2da34fc
- Theano 0.7.0 is installed under /usr/lib/python2.7/dist-packages. The packaged Theano is from https://pypi.python.org/pypi/Theano/0.7.0
- BIDMach version 1.0.0 is installed under /usr/share/bidmach and the `bidmach` helper script is located in /usr/bin. The packaged BIDMach is from the 1.0.0 release tarball at: http://bid2.berkeley.edu/bid-data-project/download/
6. Known Issues
At the time of the release no known defects have been found that affect the use of this software.
Acknowledgements
MNIST DATA: The DIGITS tutorial uses the "MNIST" dataset from: http://yann.lecun.com/exdb/mnist/
The original MNIST dataset was transformed by computing pixel_value = 255 - pixel_value to convert from black background to white background, and encoded in PNG format. During training in DIGITS, each pixel is normalized as follows: normalized_pixel = (original_pixel - m)/255 where m is the global mean across all the pixels in the training set, and 255 is the range of pixel values.
Notices
Notice
ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, "MATERIALS") ARE BEING PROVIDED "AS IS." NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE.
Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all other information previously supplied. NVIDIA Corporation products are not authorized as critical components in life support devices or systems without express written approval of NVIDIA Corporation.