Steps to set up a Deep Learning machine

deep-learning-setup.md

Create a 16.04 bootable install USB drive

http://unetbootin.github.io/

Install Ubuntu 16.04

• sudo apt-get update && sudo apt-get upgrade && sudo apt-get dist-upgrade
• sudo apt-get install make cmake gcc g++

Install NVIDIA drivers

• sudo apt-get install dkms build-essential linux-headers-generic

Create the following file /etc/modprobe.d/blacklist-nouveau.conf and add the following lines:

• blacklist nouveau
• blacklist lbm-nouveau
• options nouveau modeset=0
• alias nouveau off
• alias lbm-nouveau off

Disable the Kernel nouveau by typing the following commands:

• echo options nouveau modeset=0 | sudo tee -a /etc/modprobe.d/nouveau-kms.conf
• sudo update-initramfs -u

Disable the desktop by running the following command:

• ctrl-alt-f1

• sudo add-apt-repository ppa:graphics-drivers/ppa

• sudo apt-get update

• sudo apt-get install nvidia-367

Install SSH

• sudo apt-get install openssh-server

Turn off secure boot in the BIOS setup

• Advanced Mode | Security | Boot/Secure Boot | OS Type = Other OS

Install CUDA

• https://developer.nvidia.com/cuda-downloads
• Choose the .deb file for Ubuntu 16.04 and save it.
• Follow the installation instructions.

Install cuDNN

• Download from https://developer.nvidia.com/cudnn
• Choose latest for CUDA 8.0, “Library for Linux"
• Untar it with -zxf
• cd cuda
• sudo cp lib64/* /usr/local/cuda/lib64/
• sudo cp include/* /usr/local/cuda/include/

Install NCCL

Clone and install NVIDIA nccl

• git clone https://github.com/NVIDIA/nccl.git
• cd nccl
• make CUDA_HOME=/usr/local/cuda test
• sudo make PREFIX=/usr/local install Add the following to your ~/.profile file
• export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
• export NCCL_INCLUDE_DIR=~/nccl/include
• export NCCL_LIBRARY=~/nccl/lib
• sudo cp ~/nccl/build/include/nccl.h /usr/include

Install Caffe

• https://github.com/NVIDIA/DIGITS/blob/digits-4.0/docs/BuildCaffe.md
• cp Makefile.config.example Makefile.config
• Uncomment the USE_CUDNN := 1 line from Makefile.config
• cmake -DUSE_NCCL=ON -DNCCL_ROOT_DIR=~/nccl/build .. (necessary to use parameters because of a bug in NCCL detection)
• Note: should be able to find cuDNN and NCCL.
• make all -j8
• make runtest (to make sure it all worked)
• Add 'export CAFFE_HOME=~/caffe' to .profile

Install DIGITS

• sudo pip install —upgrade pip
• sudo pip install bs4
• sudo pip install mock
• Take the most recent tagged release (e.g. "v5.1-dev")
• https://github.com/NVIDIA/DIGITS/blob/digits-4.0/docs/BuildDigits.md
• Test that it’s OK by:
• sudo pip install -e .
• sudo pip install -e plugins/data/imageGradients
• sudo pip install -e plugins/view/imageGradients
• ./digits-test

Running

• cd ~/digits
• ./digits-devserver

Monitoring

• watch -n1 nvidia-smi