Install Caffe on EC2 from scratch (Ubuntu, CUDA 7, cuDNN)

Install Caffe on AWS from scratch

Overview

By the end of this tutorial, you will have successfully installed CUDA 7 and cuDNNv2 working with Caffe on AWS using a g2.2xlarge or g2.8xlarge instance using Ubuntu 14.04.

This guide was tested in May 2015.

Keywords: AWS, GPU, amazon, caffe, install, how, to

Begin Tutorial

This guide also applies to standard desktop Ubuntu installations.

Start up one of Amazon GPU instances (g2.2xlarge or g2.8xlarge) using Ubuntu 64 bit (HVM) and NOT Amazon's AMI. Make sure to attach both instance store 0 and instance store 1 in the "Add Storage" step. Also increase the Root /dev/sda1 device size to something larger than 8 GiB.

Installing the NVIDIA Drivers

Update and install the preliminaries:

sudo apt-get update && sudo apt-get upgrade
sudo apt-get install build-essential

Note: Amazon says you must use the 340.46 driver (see the official GPU documentation here) but this guide works while using the most recent NVIDIA driver 346.46.

Download the "run" CUDA installer (which includes the NVIDIA driver) from NVIDIA's website. The link is usually here.

wget http://developer.download.nvidia.com/compute/cuda/7_0/Prod/local_installers/cuda_7.0.28_linux.run

Extract all the installers:

chmod +x cuda_7.0.28_linux.run
mkdir nvidia_installers
./cuda_7.0.28_linux.run -extract=`pwd`/nvidia_installers

Then update the linux image to be compatible with NVIDIA's drivers:

sudo apt-get install linux-image-extra-virtual

Important: While installing the linux-image-extra-virtual, you may be prompted "What would you like to do about menu.lst?" I selected "keep the local version currently installed"

We now need to disable nouveau since it conflicts with NVIDIA's kernel module:

sudo vi /etc/modprobe.d/blacklist-nouveau.conf

And add the following lines to this file:

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

Back in the terminal/shell, execute the commands:

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

After the reboot is complete, we have a few more steps:

sudo apt-get install linux-source
sudo apt-get install linux-headers-`uname -r`

Now we can finally install the driver:

cd nvidia_installers
sudo ./NVIDIA-Linux-x86_64-346.46.run

1. Accept the license agreement.
2. If you see: "nvidia-installer was forced to guess the X library path '/usr/lib' and X module path ..." go ahead anc click OK.
3. If you see "The CC version check failed" then click "Ignore CC version check".
4. It may ask you about 32-bit libraries, I selected to yes, install them.
5. It will ask you about running nvidia-xconfig to update your X configuration file. I selected no.
6. Run nvidia-smi to view the installed GPUs.

Installing CUDA

Now we can install CUDA and optionally the examples. Make sure to run sudo modprobe nvidia first.

sudo modprobe nvidia
sudo apt-get install build-essential
sudo ./cuda-linux64-rel-7.0.28-19326674.run
sudo ./cuda-samples-linux-7.0.28-19326674.run

1. Sometimes it is not necessary to reinstall build-essential.
2. When the license agreement appears, press "q" so you don't have to scroll down.
3. Accept the EULA.
4. Use the default path by pressing enter.
5. Would you like to add desktop menu shortcuts? Answer depends on your preference.
6. Would you like to create a symbolic link? Enter yes.
7. It will now install CUDA.

Finally, update your path variables. Open your ~/.bashrc file and ad the following lines:

export PATH=$PATH:/usr/local/cuda-7.0/bin
export LD_LIBRARY_PATH=:/usr/local/cuda-7.0/lib64

Remember to run source ~/.bashrc after saving .bashrc and run ldconfig as root ($ sudo ldconfig)

Installing cuDNN

After registering with NVIDA, download cuDNN. Extract the tar and copy the headers and libraries to the CUDA directory.

Update: Caffe now requires cuDNN v3. Get it here.

tar -zxf cudnn-7.0-linux-x64-v3.0-prod.tgz
cd cuda
sudo cp lib64/* /usr/local/cuda/lib64/
sudo cp include/cudnn.h /usr/local/cuda/include/

Installing Caffe

Install the dependencies:

sudo apt-get install -y libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libboost-all-dev libhdf5-serial-dev protobuf-compiler gfortran libjpeg62 libfreeimage-dev libatlas-base-dev git python-dev python-pip libgoogle-glog-dev libbz2-dev libxml2-dev libxslt-dev libffi-dev libssl-dev libgflags-dev liblmdb-dev python-yaml python-numpy

Then run:

sudo easy_install pillow

Now we can download Caffe. Navigate to the directory of your choice for the cloning.

cd ~
git clone https://github.com/BVLC/caffe.git

We now install more dependencies. Warning: This takes 10-30 minutes.

cd caffe
cat python/requirements.txt | xargs -L 1 sudo pip install

Now we update the Makefile:

cp Makefile.config.example Makefile.config
vi Makefile.config

1. Uncomment the line: USE_CUDNN := 1
2. Make sure the CUDA_DIR correctly points to our CUDA installation.
3. If you want the Matlab wrapper, uncomment the appropriate MATLAB_DIR line.

Now we build Caffe. Set X to the number of CPU threads (or cores) on your machine. Use the command htop to check how many CPU threads you have.

make pycaffe -jX
make all -jX
make test -jX

Now to quickly test Caffe, from the CAFFE_ROOT (wherever the Caffe code resides)

./data/mnist/get_mnist.sh
./examples/mnist/create_mnist.sh
./examples/mnist/train_lenet.sh

You may get errors for create_mnist.sh but run train_lenet.sh anyway. Chances are it will still work. If you see the network training, then everything has been successfully set up.

You may get the following error :

libdc1394 error: Failed to initialize libdc1394

To fix it, disable the corresponding driver:

sudo ln /dev/null /dev/raw1394