2. GPU Nvidia support
• Find the your Nvidia GPU “Compute
Capability” or architecture from
https://developer.nvidia.com/cuda-gpus
• CUDA SDK 6.5: Last Version with support
for Tesla with Compute Capability 1.x
• CUDA SDK 7.5 support for Compute
Capability 2.0 – 5.x (Fermi, Kepler,
Maxwell)
• CUDA SDK 8.0 support for Compute
Capability 2.0 – 6.x (Fermi, Kepler,
Maxwell, Pascal)
3. Installation deep learning for DIGITS
• Ubuntu 14
• Nvidia driver (375)
• CUDA 8 (do not install driver)
• cuDNN 5.1
• DIGITS
• If you want Update DIGITS 5.1
4. Install Ubuntu 14
• In windows you can create bootable usb flash by “rufus-2.11.exe”
• You can install Ubuntu alongside windows
• You should used Ubuntu as a training then when the model is trained you
can use it in another platform
• If you have one partition (“C”) you need to shrink and make around
100GB free space then install Ubuntu on the free space in your hard
disk. If your windows not show in the booting menu used below
command
• Go to terminal by “Alt+ ctrl+F1”, put user and password
• Sudo update-grub
• Restart.
5. Nvidia driver
• Alt+ctrl+F1 (if you have two GPU used none nvidia for monitor)
• sudo stop lightdm
• sudo add-apt-repository ppa:graphics-drivers/ppa
• sudo apt-get update
• sudo apt-get –y install nvidia-375
6. Install CUDA 8
• https://developer.nvidia.com/cuda-downloads
• Download the latest version of CUDA
• The NVIDIA CUDA Toolkit: A development environment for building GPU-accelerated
applications. This toolkit includes a compiler specifically designed for NVIDIA GPUs and
associated math libraries + optimization routines.
7. cuDNN
• Register free account in nvidia website
• Download cuDNN
install and configure to Ubuntu library
• The cuDNN library: A GPU-accelerated library of primitives for deep neural networks.
Using the cuDNN package, you can increase training speeds by upwards of 44%, with over
6x speedups in Torch and Caffe.
8. Install DIGITS
• Search “nvidia digits github”
• Fallowing the instructions
Deb
packages
Ubuntu
14.04
14.04 repo
docs/Ubunt
uInstall.md
10. Datasets for Computer vision + Deep Learning
Google Research: Computer vision + Deep Learning
1. Images: Open Images Dataset
2. Videos: YouTube-8M: A Large and Diverse Labeled Video Dataset
for Video Understanding Research
11. Datasets Images
• Deep learning needs large amount of inputs for training. detecting
and classifying objects in static images
• Open Images Dataset
automatically caption images
natural language replies in response to shared photos
~9 million URLs to images
6000 categories
each image has about 8 labels assigned
Inception v3 model
12. Datasets Video I
1. video is much more time-consuming to annotate manually than
images
• video annotation system, which identifies relevant Knowledge Graph topics
• video metadata and content analysis
• only public videos with more than 1000 views
• frequency analysis, automated filtering, verification by human raters
• 24 top-level verticals
2. video is very computationally expensive to process and store
• extracted frame-level features
• Inception-V3 image annotation model
13. Datasets Video II
• YouTube-8M: A Large and Diverse Labeled Video Dataset for Video
Understanding Research
• 8 million YouTube video URLs (representing over 500,000 hours of video)
• 4800 Knowledge Graph entities (classes)
14. For research
• Reduce the size of model
• Deep compression
• Pruning deep learning
• Hash table neural network