Install GPU TensorFlow on AWS Ubuntu 16.04

 TensorFlow™ is an open source software library for numerical computation using data flow graphs. Nodes in the graph represent mathematical operations, while the graph edges represent the multidimensional data arrays (tensors) communicated between them.

On a typical system, there are multiple computing devices. In TensorFlow, the supported device types are CPU and GPU.  GPUs offer 10 to 100 times more computational power than traditional CPUs, which is one of the main reasons why graphics cards are currently being used to power some of the most advanced neural networks responsible for deep learning.

The environment setup is often the hardest part of getting a deep learning setup going, so hopefully you will find this step-by-step guide helpful.

Launch a GPU-enabled Ubuntu 16.04 AWS instance

Choose an Amazon Machine Image (AMI) – Ubuntu Server 16.04 LTS


Choose an instance type

The smallest GPU-enabled machine is p2.xlarge


You can find more details here.

Configure Instance Details, Add Storage (choose storage size), Add Tags, Configure Security Group and Review Instance Launch and Launch.


Open the terminal on your local machine and connect to the remote machine (ssh -i)

Update the package lists for upgrades for packages that need upgrading, as well as new packages that have just come to the repositories

sudo apt-get –assume-yes update

Install the newer versions of the packages

sudo apt-get –assume-yes  upgrade

Install the CUDA 8 drivers

CUDA is a parallel computing platform and application programming interface (API) model created by Nvidia. GPU-accelerated CUDA libraries enable drop-in acceleration across multiple domains such as linear algebra, image and video processing, deep learning and graph analytics.

Verify that you have a CUDA-Capable GPU

lspci | grep -i nvidia
00:1e.0 3D controller: NVIDIA Corporation GK210GL [Tesla K80] (rev a1)

Verify You Have a Supported Version of Linux

uname -m && cat /etc/*release


The x86_64 line indicates you are running on a 64-bit system. The remainder gives information about your distribution.

 Verify the System Has gcc Installed

gcc –version

If the message is “The program ‘gcc’ is currently not installed. You can install it by typing: sudo apt install gcc”

sudo apt-get install gcc

gcc –version

gcc (Ubuntu 5.4.0-6ubuntu1~16.04.5) 5.4.0 20160609


Verify the System has the Correct Kernel Headers and Development Packages Installed

uname –r


CUDA support

Download the CUDA-8 driver (CUDA 9 is not yet supported by TensorFlow 1.4)

The driver can be downloaded from here:



Or, downloaded directly to the remote machine:

wget -O ./cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb

Downloading patch 2 as well:

wget -O ./cuda-repo-ubuntu1604-8-0-local-cublas-performance-update_8.0.61-1_amd64.deb

Install the CUDA 8 driver and patch 2

Extract, analyse, unpack and install the downloaded .deb files

sudo dpkg -i cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb

sudo dpkg -i cuda-repo-ubuntu1604-8-0-local-cublas-performance-update_8.0.61-1_amd64.deb

apt-key is used to manage the list of keys used by apt to authenticate packages. Packages which have been authenticated using these keys will be considered trusted.

sudo apt-key add /var/cuda-repo-8-0-local-ga2/
sudo apt-key add /var/cuda-repo-8-0-local-cublas-performance-update/

sudo apt-get update

Once completed (~10 min), reboot the system to load the NVIDIA drivers.

sudo shutdown -r now

Install cuDNN v6.0

The NVIDIA CUDA® Deep Neural Network library (cuDNN) is a GPU-accelerated library of primitives for deep neural networks. cuDNN provides highly tuned implementations for standard routines such as forward and backward convolution, pooling, normalization, and activation layers.

Download the cuDNN v6.0 driver

The driver can be downloader from here: please note that you will need to register first.


Copy the driver to the AWS machine (scp -r -i)

Extract the cuDNN files and copy them to the target directory

tar xvzf cudnn-8.0-linux-x64-v6.0.tgz  

sudo cp -P cuda/include/cudnn.h /usr/local/cuda/includesudo

cp -P cuda/lib64/libcudnn* /usr/local/cuda/lib64

sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

Update your bash file

nano ~/.bashrc

Add the following lines to the end of the bash file:

export CUDA_HOME=/usr/local/cuda


export PATH=${CUDA_HOME}/bin:${PATH}


Save the file and exit.

Install TensorFlow

Install the libcupti-dev library

The libcupti-dev library is the NVIDIA CUDA Profile Tools Interface. This library provides advanced profiling support. To install this library, issue the following command:

sudo apt-get install libcupti-dev

Install pip

Pip is a package management system used to install and manage software packages written in Python which can be found in the Python Package Index (PyPI).

sudo apt-get install python-pip

sudo pip install –upgrade pip

Install TensorFlow

sudo pip install tensorflow-gpu

Test the installation

Run the following within the Python command line:

from tensorflow.python.client import device_lib

def get_available_gpus():

    local_device_protos = device_lib.list_local_devices()

    return [ for x in local_device_protos if x.device_type == ‘GPU’]


The output should look similar to that:

2017-11-22 03:18:15.187419: I tensorflow/core/platform/] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA

2017-11-22 03:18:17.986516: I tensorflow/stream_executor/cuda/] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero

2017-11-22 03:18:17.986867: I tensorflow/core/common_runtime/gpu/] Found device 0 with properties:

name: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235

pciBusID: 0000:00:1e.0

totalMemory: 11.17GiB freeMemory: 11.10GiB

2017-11-22 03:18:17.986896: I tensorflow/core/common_runtime/gpu/] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: Tesla K80, pci bus id: 0000:00:1e.0, compute capability: 3.7)




Agile Development of Data Products with R and Shiny: A Practical Approach

Many companies are interested in turning their data assets into products and services.  It’s not limited anymore to online firms like LinkedIn or Facebook, but there is a variety of companies in offline industries (GM, Apple etc) who have started to develop products and services based on analytics.

But how do you succeed at developing and launching data products?

I would like to suggest a framework building on the idea of Lean Startup and the Minimum Viable Product (MVP), to support the rapid modelling and development of innovative data products. These principles can be applied when launching a new tech start-up, starting a small business, or when starting a new initiative within a large corporation.

Agile Development of Data Products

A minimum viable product has just those core features that allow the product to be deployed, and no more. The product is typically deployed to a subset of possible customers, such as early adopters that are thought to be more forgiving, more likely to give feedback, and able to grasp a product vision from an early prototype or marketing information

Some of the benefits of prototyping and developing MVP:

1.       You can get valuable feedback from the users early in the project.

2.       Different stakeholders can compare if the model matches the specification.

3.       It allows the model developer some insight into the accuracy of initial project estimates and whether the deadlines and milestones proposed can be successfully met.

Our fully functional model will look like that:

Simulator Screenshot

Before we dive into the details, feel free to play around with the prototype and get familiar with the model

Ready to go?

Let’s go through the different stages of the process, following a simple example and using R and Shiny for modelling and prototyping. I’ve also published the code in my github repository , feel free to fork it and to play with it.


This is where your creativity should kick in! What problem do you try to solve by using data?

Are you aware of a gap in the industry that you currently work in?

Let’s follow a simple example:

Effective employee benefits will significantly reduce staff turnover and companies with the most effective benefits are using them to influence the behavior of their staff and their bottom line, as opposed to simply being competitive

How can you find this optimal point, balancing between the increasing cost of employee benefits and the need to retain staff and reduce staff turnover?


We will assume a (simplified) model that links the attrition rates to the benefits provided to the employee.

I know, it’s simplified. I’m also aware that there are many other relevant parameters in a real-life scenario.

But it’s just an example, so let’s move on.


Our simplified model depends on the following parameters:

  1. Number of Employees
  2. Benefits Saturation ($): we assume a linear dependency between the attrition rate and the benefits provided by the company. As the benefits increase, attrition rate drops, to 0% attrition at the point of Benefits Saturation. Any increase of benefits above the Benefits Saturation point will not have an impact on the attrition rates.
  3. Benefits ($): benefits provided by the company
  4. Max Attrition (%): maximal attrition rates at lowest benefits (100$)
  5. Training Period (months): number of months required to train a new employee
  6. Salary ($)

This model demonstrates the balance between increasing the benefits and the overall cost, and reducing the attrition rate and the associated cost related to hiring and training of new staff.

We will use R to implement our model:

R is an open-source software environment for statistical computing and graphics. The R language is widely used among statisticians and data miners for developing statistical software and data analysis.

We will use R-Studio, a powerful and productive user interface for R. It’s free and open source, and works great on Windows, Mac, and Linux.

Let’s create a file named server.R and write some code. You can find the full source code in my github repository:

Cutoff is a function that models the attrition vs. benefits, where cutoffx is the value of Benefits Saturation:

cutoff <- function(minx, maxx,maxy,cutoffx,x)


ysat <- (x>=cutoffx)*0

slope <- ( 0 – maxy ) / ( cutoffx – minx )

yslope <- (maxy + (x-minx)*slope)*(x

return(ysat + yslope)


Calculating the different cost components:

benefitsCost <- input$numberee * input$benefits

attritionCost <- input$salary * nTrainingMonths * input$numberee * (currentAttrition / 100)

overallCost <- benefitsCost + attritionCost

The value of the variables starting with input$ is retrieved from the sliding bars, which are part of the User Interface (UI):


When changing the value of the slider named “Number Of Empolyees” from a value of 200 to a value of 300, the value of the variable input$numberee will change from 200 to 300 accordingly.

Benefits is a sequence of 20 numbers from 100 to 500, covering the range of benefits:

benefits <- round(seq(from = 100, to = 500, length.out = 20))

Let’s plot the benefits cost, the attrition cost and the overall cost as a function of the benefits:

Cost Vs Benefits

The different cost components are calculated below:

benefitsCostV <- input$numberee * benefits

attritionCostV <- input$salary * nTrainingMonths * input$numberee * (attrition / 100) totalCostV <- benefitsCostV + attritionCostV

And now we can plot the different cost components:

plot(benefitsCostV ~ benefits,col = “red”, main = “Cost vs. Benefits”, xlab = “benefits($)”, ylab = “cost($)”)

lines(benefits,benefitsCostV, col = “red”, lwd = 3)

points(benefits,attritionCostV, col = “blue”)

lines(benefits,attritionCostV, col = “blue”,lwd = 3)

points(benefits,totalCostV, col = “purple”)

lines(benefits,totalCostV, col = “purple”,lwd = 3)

Let’s find the minimal cost, and draw a nice orange circle around this optimal point:

minBenefitsIndex <- which.min(totalCostV)

minBenefits <- benefits[minBenefitsIndex]

minBenefitsCost <- totalCostV[minBenefitsIndex]

abline(v=minBenefits,col = “cyan”, lty = “dashed”, lwd = 1)

symbols(minBenefits,minBenefitsCost,circles=20, fg = “darkorange”, inches = FALSE, add=TRUE, lwd = 2)

Tip: Don’t spend too much time on writing the perfect R code; your model might change a lot once your stakeholders will provide their feedback.


Shiny is web application framework for R that will turn your analyses into interactive web applications.

Let’s install Shiny and import the library functions:



Once we have the model in place, we will create the user interface and link it back to the model.

Create a new file named ui.R with the User Interface (UI) elements.

For example, let’s write some text:

titlePanel(“Cost Optimization (Benefits and Talent) – Simulation”),

h5(“This interactive application simulates the impact of multiple …..

And let’s add a slider:


“Number of Enployees:”,

min = 100,

max = 1000,

value = 200,

step = 100),

The inputId of the slider (numberee) is linking the value of the UI control (number of Employees) to the server side computation engine.

Create a Shiny account, copy-paste the token and a secret to the command line and execute in R-Studio:

shinyapps::setAccountInfo(name=’benefits’, token=’xxxx’, secret=’yyyy’)

And, deploy your code to the Shiny server:


Your prototype is live!

Send the URL to your stakeholders and collect their feedback. Iterate quickly and improve the model and the user interface.


Once your stakeholders are happy with your prototype, it’s time to move on to the next stage and develop your data product.

The good news is that at this stage you should have a pretty good understanding of the requirements and the priorities, based on the feedback provided by your stakeholders.

It’s also the perfect time for you (or for your product development group) to focus more on hosting, architecture, design, the Software Development Life-cycle (SDLC), quality assurance, release management and more.

There are different technologies to consider when developing data products, which I will cover in future posts.

For now I will just mention an interesting option, where you can reuse your server-side R code. Using yhat you can expose your server-side R functionality via a set of web services, and consume these services from a client-side JavaScript libraries, like d3js.

Comments, questions?

Let me know.