Deploying deep learning models

Platform agnostic approach for production with docker+Kubernetes

About this presentation
This was originally created to explain the basics of
code deployment both in academia and startup
environments.
ACADEMIA Especially in academia outside computer science
departments it is typical that the code developed is very unstructured
without much thought on reproducibility or legibility.

STARTUPS In smaller startups it is beneficial for all the team members

to understand at least something on all the various aspects involved in
building a tech product.

Based on personal experience, these sort of knowledge gaps between a

clinician/biologist and the technical person both in technology and
biology can make the teamwork painfully slow.
General “Data Science” Architecture
Typical data scientist or researcher may use the
“tech stack” on the left.
In academia, the code is typically not deployed anywhere, e.g.
installing a custom app on the smartphone of a subject in a clinical
trial. The researcher simply gather the data with some 3rd party
software and then writes some research-grade code to analyze it.

In industry, the data science team can consist of multiple roles, and
it becomes essential for the organization to have a smooth operation
between different roles. In other words, the research and models
done by the data scientist can be put in production quickly without
major re-writing of the code.

http://101.datascience.community/2016/11/28/data-scientists-data-eng https://www.slideshare.net/continuumio/journey-to-open-data-science
ineers-software-engineers-the-difference-according-to-linkedin/
General “BIG DATA/ML” Architecture
For example the model developed in TensorFlow might look like this when deployed as product (e.g.
as an app for your phone to tell whether your image contains a cat or a dog).
DOCKER

https://www.docker.com/survey-2016

https://www.docker.com/what-docker EXACTLY THE SAME MODEL

Both locally at the office and in the cloud

e.g. ASUS ESC8000 G3

local Server 'Lock-in'less
Cloud Service

For inference, i.e.

process customer
queries via API
DOCKER Deep learning?

http://www.somatic.io/blog/docker-and-deep-learning-a-bad-match

Unfortunately that is wrong for deep learning applications. For any serious deep learning application, you need NVIDIA
graphics cards, otherwise it could take months to train your models. NVIDIA requires both the host driver and the docker
image's driver to be exactly the same. If the version is off by a minor number, you will not be able to use the NVIDIA card, it
will refuse to run. I don't know how much of the binary code changes between minor versions, but I would rather have the card
try to run instructions and get a segmentation fault then die because of a version mismatch.

We build our docker images based off the NVIDIA card and driver along with the software needed. We essentially have the
same docker image for each driver version. To help stay manage this, we have a test platform that makes sure all of
our code runs on all the different docker images.

This issue is mostly in NVIDIA's court, they can modify their drivers to be able to work across different versions. I'm not
sure if there is anything that Docker can do on their side. I think its something they should figure out though, the
combination of docker and deep learning could help a lot more people get started faster, but right now its
an empty promise.

The biggest impact on data science right now is not coming from a new
algorithm or statistical method. It’s coming from Docker containers.
Containers solve a bunch of tough problems simultaneously: they make it easy to
use libraries with complicated setups; they make your output reproducible; they
make it easier to share your work; and they can take the pain out of the Python data
science stack.

The wonderful triad of Docker : “Isolation! Portability!

Repeatability!” There are numerous use cases where Docker might just be what
you need, be it Data Analytics, Machine Learning or AI
DOCKERize everything as microservices

.pwc.com/us/en/technology-forecast/2014

(ARC401) Cloud First: New Architecture for New Infrastructure

Amazon Web Services, slideshare.net/AmazonWebServices
http://www.slideshare.net/RichardHarvey7/micro-services-and-containers
Why Microservices?
Why run microservices using
Docker and Kubernetes?
Posted by: Seth Lakowske events.linuxfoundation.org, Frank Zhao
Published: 2016-04-25
http://sethlakowske.com/articles/why-run-docker-containers-and-kubernetes/

Benefits of microservices
1) Code can be broken out into smaller microservices that are easier to learn, release
and update.
2) Individual microservices can be written using the best tools for the job.
3) Releasing a new service doesn't require synchronization across a whole company.
4) New technology stacks have lower risk since the service is relatively small.
5) Developers can run containers locally, rebuilding and verifying after each commit on a
system that mirrors production.
6) Both Docker and Kubernetes are open source and free to use.
7) Access to Docker hub leverages the work of the opensource community.
8) Service isolation without the heavyweight VM. Adding a service to a server does not
affect other services on the server.
9) Services can be more easily run on a large cluster of nodes making it more reliable.
10) Some clients will only host in private and not on public clouds.
11) Lends itself to immutable infrastructure, so services are reloadable without missing
state when a server goes down.
12) Immutable containers improve security since data can only be mutated in specified
volumes, root kits often can't be installed even if the system is penetrated.
13) Increasing support for new hardware, like the GPU in a container means even gpgpu
tasks like deep learning can be containerized.
14) There is a cost for running microservices - the build and runtime becomes more
complex. This is part of the price to pay and if you've made the right decision in your
context, then benefits will exceed the costs.

Costs of microservices
• Managing multiple services tends to be more costly.
• New ways for network and servers to fail.

Conclusion
In the right circumstances, the benefits of microservices outweigh the extra cost of
management. https://www.nextplatform.com/2016/09/13/will-containers-total-package-hpc/
Docker vs AWS Lambda ”In General”
AWS Lambda will win - sort of.....  From a programming model
and a cost model, AWS Lambda is the future - despite so of the
tooling limitations. Docker in my opinion is an evolutionary step of
"virtualization" that we've been seeing for the last 10 years. AWS
Lambda is a step-function. In fact, I personally think it is innovations
like Amazon Elastic Beanstalk and CloudFormation that has pushed
the demand solutions like Docker.  In the near future, I predict that
open source will catch up and provide an AWS Lambda
experience on top of Docker containers. Iron.io is opensource
and appears to be going down this path.
Florian Walker Product Manager at Fujitsu
The future is now :) Funktion, part of Fabric8, aims to
provide a Lamda experience on-top of Kubernetes ->
https://github.com/fabric8io/funktion
Jason Daniels CTO - Fujitsu Hybrid Cloud EMEIA
project Kratos ..
https://www.iron.io/introducing-aws-lambda-support/
https://www.quora.com/Are-there-any-alternatives-to-Amazon-Lambda
Funktion is an open source event driven lambda style programming
model on top of Kubernetes. A funktion is a regular function in any
programming language bound to a trigger deployed into
Kubernetes. Then Kubernetes takes care of the rest (scaling, high
availability, load balancing, logging and metrics etc).
Funktion supports hundreds of different trigger endpoint URLs including
most network protocols, transports, databases, messaging systems,
social networks, cloud services and SaaS offerings. In a sense funktion is
a serverless approach to event driven microservices as you focus on
just writing funktions and Kubernetes takes care of the rest. Its not that
there's no servers; its more that you as the funktion developer don't have
to worry about managing them.
Announcing Project Kratos I’m happy to
announce that Project Kratos is now available
in beta. Iron.io is rolling out a set of tools that
allow you to convert AWS Lambda functions
into Docker images. Now, you can import
existing Lambda functions and run them via
any container orchestration system. You can
also create new Lambda functions and
quickly package them up in a container to run
on other platforms. All three of the AWS
runtimes are supported – Node.js, Python and
Java.
Docker Issues Size
Docker containers quickly grow in size as they need to contain everything required for deployment
http://blog.xebia.com/create-the-smallest-possible-docker-container/
https://ypereirareis.github.io/blog/2016/02/15/docker-image-size-optimization/
https://github.com/microscaling/imagelayers-graph
ImageLayers.io is a project maintained by Microscaling Systems since September 2016. The project
was developed by the team at CenturyLink Labs. This utility provides a browser-based visualization
of user-specified Docker Images and their layers. This visualization provides key information on the
composition of a Docker Image and any commonalities between them. ImageLayers.io allows
Docker users to easily discover best practices for image construction, and aid in determining which
images are most appropriate for their specific use cases.
Deploying in Kubernetes Please see deployment/README.md
https://www.ctl.io/developers/blog/post/optimizing-docker-images/
“Docker images can get really big. Many are over 1G in size. How do they get so big? Do they really
need to be this big? Can we make them smaller without sacrificing functionality?
Here at CenturyLink we've spent a lot of time recently building different docker images. As we began
experimenting with image creation one of the things we discovered was that our custom images
were ballooning in size pretty quickly (it wasn't uncommon to end up with images that weighed-in at
1GB or more). Now, it's not too big a deal to have a couple gigs worth of images sitting on your local
system, but it becomes a bit of pain as soon as you start pushing/pulling these images across the
network on a regular basis. “
https://blog.replicated.com/2016/02/05/refactoring-a-dockerfile-for-image-size/
“There’s been a welcome focus in the Docker community recently around image
size. Smaller image sizes are being championed by Docker and by the community.
When many images clock in at multi-100 MB and ship with a large ubuntu base, it’s
greatly needed.”
What is lambda architecture anyway?
https://www.oreilly.com/ideas/questioning-the-lambda-architecture
The Lambda Architecture is an approach to building stream processing applications
on top of MapReduce andStorm or similar systems. This has proven to be a
surprisingly popular idea, with a dedicated website and an upcoming book. 
kappa-architecture.com
Kappa Architecture is a simplification of Lambda Architecture. A
Kappa Architecture system is like a Lambda Architecture system with the batch
processing system removed. To replace batch processing, data is simply fed
through the streaming system quickly.
Kappa Architecture revolutionizes database migrations and reorganizations: just
delete your serving layer database and populate a new copy from the canonical
store! Since there is no batch processing layer, only one set of code needs to
be maintained.

CHALLENGING THE LAMBDA

The way this works is that an immutable sequence of records is captured and fed into a batch
system and a stream processing system in parallel. You implement your transformation
logic twice, once in the batch system and once in the stream processing system. You stitch
together the results from both systems at query time to produce a complete answer. There
are a lot of variations on this.
The Lambda Architecture is aimed at applications built around complex asynchronous
transformations that need to run with low latency (say, a few seconds to a few hours). A
good example would be a news recommendation system that needs to crawl various
news sources, process and normalize all the input, and then index, rank, and store it for
serving.
ARCHITECTURE: BUILDING APPS
FOR FAST DATA WITH VOLTDB V5.0
dataconomy.com
VoltDB is an ideal alternative to the Lambda Architecture’s speed layer. It offers horizontal scaling
and high per-machine throughput. It can easily ingest and process millions of tuples per second
with redundancy, while using fewer resources than alternative solutions. VoltDB requires an
order of magnitude fewer nodes to achieve the scale and speed of the Lambda speed layer. As a
benefit, substantially smaller clusters are cheaper to build and run, and easier to manage.

I like that the Lambda Architecture emphasizes retaining the input data unchanged. I
think the discipline of modeling data transformation as a series of materialized stages from an
original input has a lot of merit. I also like that this architecture highlights the problem of
reprocessing data (processing input data over again to re-derive output).

The problem with the Lambda Architecture is that maintaining code that needs to produce
the same result in two complex distributed systems is exactly as painful as it seems like it
would be. I don’t think this problem is fixable. Ultimately, even if you can avoid coding your
application twice, the operational burden of running and debugging two systems is
going to be very high. And any new abstraction can only provide the features supported by
the intersection of the two systems. Worse, committing to this new uber-framework walls off
the rich ecosystem of tools and languages that makes Hadoop so powerful (Hive, Pig,
Crunch, Cascading, Oozie, etc).
Managing containers
Enter Kubernetes (from Google)
DOCKER Management → enter Kubernetes

www.computerweekly.com/feature/Demystifying-Kubernete

Once every five years, the IT industry witnesses a major technology

shift. In the past two decades, we have seen server paradigm evolve
into web-based architecture that matured to service orientation before
finally moving to the cloud. Today it is containers.
Docker is much more than just the tools and API. It created a vibrant
ecosystem that started to contribute to a variety of tools to manage the
lifecycle of containers. 
One of the first tools that Google decided to make open source is
called Kubernetes, which means “pilot” or “helmsman” in Greek.
Kubernetes works in conjunction with Docker. While Docker provides the
lifecycle management of containers, Kubernetes takes it to the next
level by providing orchestration and managing clusters of
containers.
Traditionally, platform as a service (PaaS) offerings such as Azure,
App Engine, Cloud Foundry, OpenShift, Heroku and Engine Yard
exposed the capability of running the code by abstracting the
infrastructure.

Kubernetes and Docker deliver the promise of PaaS through a

simplified mechanism. Once the system administrators 
configure and deploy Kubernetes on a specific infrastructure,
developers can start pushing the code into the clusters. This hides the
complexity of dealing with the command line tools, APIs and
https://www.youtube.com/watch?v=PivpCKEiQOQ
dashboards of specific IaaS providers. 
Containers at scale

As has been demonstrated, it is relatively easy to launch tens of thousands of containers

nextplatform.com/2016/03/22
 on a single host. But how do you deploy thousands of containers? How do you
manage and keep track of them? How do you manage and recover from failure. While
these things sometimes might look easy, there are some hard problems to tackle. Let us
walk through what it makes it so difficult.

With a single command the Docker environment is set up and you can docker run until nextplatform.com/2016/03/03
you drop. But what if you have to run Docker containers across two hosts? How about
50 hosts? Or how about 10,000 hosts? Now, you may ask why one would want to do
 this. There are some good reasons why:

https://www.nextplatform.com/2015/09/29/why-containers-at-scale-is-hard/

Two founders of the Kubernetes project at Google, Craig McLuckie and Joe Beda, today
announced their new company, Heptio. The company has raised $8.5 million in a series A
investment round led by Accel, with participation from Madrona Venture Group.

Open source Kubernetes is a widely deployed technology for container orchestration. Now,

Heptio will bring a commercial version of the software to enterprises.

www.sdxcentral.com
Kubernetes
Kubernetes is an open-source system for automating
deployment, scaling, and management of containerized
applications.
It groups containers that make up an application into logical units for easy
management and discovery. Kubernetes builds upon 
15 years of experience of running production workloads at Google, combined with
best-of-breed ideas and practices from the community.

KubeWeekly — aggregating all interesting weekly news about Kubernetes

in the form of a newsletter. Manage a cluster of Linux containers as a single
system to accelerate Dev and simplify Ops.
https://kubeweekly.com/

https://www.youtube.com/watch?v=21hXNReWsUU

http://cloud9.nebula.fi/app.html http://nshani.blogspot.co.uk/2016/02/getting-started-with-kubernetes.html
Kubernetes concepts

http://www.slideshare.net/arungupta1/package-your-ja
va-ee-application-using-docker-and-kubernetes

http://www.slideshare.net/jawnsy/kubernetes-my-bff

Inference can be very resource intensive. Our server executes the following
TensorFlow graph to process every classification request it receives. The
Inception-v3 model has over 27 million parameters and runs 5.7 billion floating
point operations per inference.

Fortunately, this is where Kubernetes can help us. Kubernetes distributes

inference request processing across a cluster using its External Load Balancer.
Each pod in the cluster contains a TensorFlow Serving Docker image with the
TensorFlow Serving-based gRPC server and a trained Inception-v3 model.
The model is represented as a set of files describing the shape of the
TensorFlow graph, model weights, assets, and so on.

Since everything is neatly packaged together, we can dynamically scale the

number of replicated pods using the Kubernetes Replication Controller to keep
up with the service demands.

blog.kubernetes.io/2016/03
linkedin.com/pulse
Alternatives
medium.com/@mustwin
Bare Metal
Most schedulers with the notable exception of Cloud Foundry can be
installed on “bare metal” or physical machines inside your datacenter. This
can save you big on hypervisor licensing fees.
Volume Mounts
Volume mounts allow you to persist data across container
deployments. This is a key differentiator depending on your applications’
needs. Mesos is the leader here, and Kubernetes is slowly catching up.
https://news.ycombinator.com/item?id=10438273
https://www.oreilly.com/ideas/swarm-v-fleet-v-kubernetes-v-mesos
Conclusion
There are clearly a lot of choices for orchestrating, clustering, and managing
containers. That being said, the choices are generally well differentiated. In terms of
orchestration, we can say the following:
Swarm has the advantage (and disadvantage) of using the standard Docker interface.
Whilst this makes it very simple to use Swarm and to integrate it into existing
workflows, it may also make it more difficult to support the more complex scheduling
that may be defined in custom interfaces.
Fleet is a low-level and fairly simple orchestration layer that can be used as a base for
running higher level orchestration tools, such as Kubernetes or custom systems.
Kubernetes is an opinionated orchestration tool that comes with service discovery and
replication baked-in. It may require some re-designing of existing applications, but used
correctly will result in a fault-tolerant and scalable system.
Mesos is a low-level, battle-hardened scheduler that supports several frameworks for
container orchestration including Marathon, Kubernetes, and Swarm. At the time of
writing, Kubernetes and Mesos are more developed and stable than Swarm. In terms
of scale, only Mesos has been proven to support large-scale systems of hundreds or
thousands of nodes. However, when looking at small clusters of, say, less than a dozen
nodes, Mesos may be an overly complex solution.
Kubernetes Still on top?
https://news.ycombinator.com/item?id=12462261
http://www.infoworld.com/article/3118345/cloud-computing/why-kubernetes-is-winning-the-container-war.html
After all, Kubernetes is a mere two years old (as a public open source
project), whereas Apache Mesos has clocked seven years in market.
Docker Swarm is younger than Kubernetes, and it comes with the
backing of the center of the container universe, Docker Inc Yet the
orchestration rivals pale in comparison to Kubernetes' community,
which -- now under management by the
Cloud Native Computing Foundation -- is exceptionally large and
diverse.
•
Kubernetes is one of the top projects on GitHub: in the top 0.01
percent in stars and No. 1 in terms of activity.
•
While documentation is subpar, Kubernetes has a significant Slack
and Stack Overflow community that steps in to answer questions
and foster collaboration, with growth that dwarfs that of its rivals.
•
More professionals list Kubernetes in their LinkedIn profile than
any other comparable offering by a wide margin.
•
Perhaps most glaring, data from OpenHub shows Apache Mesos
dwindling since its initial release and Docker Swarm starting to
slow. In terms of raw community contributions, Kubernetes is
exploding, with 1,000-plus contributors and 34,000 commits --
more than four times those of nearest rival Mesos.
https://github.com/kubernetes/kubernetes
I would argue that general-purpose clusters like those managed by Google
Kubernetes are better for hosting Internet businesses depending on artificial
intelligence technologies than special-purpose clusters like NVIDIA DGX-1.
Consider the case that an experiment model training job is using all the 100 GPUs in the cluster. A
production job gets started and asks for 50 GPUs. If we use MPI, we'd have to kill the experiment job
so to release enough resource to run the production job. This tends to make the owner of the
experiment job get the impression that he is doing a "second-class" work.
Kubernetes is smarter than MPI as it can kill, or preempt, only 50 workers of the experiment job,
so to allow both jobs run at the same time. With Kubernetes, people have to build their programs
into Docker images that run as Docker containers. Each container has its own filesystem and
network port space. When A runs as a container, it removes only files in its own directory. This is to
some extent like that we define C++ classes in namespaces, which helps us removing class name
conflicts.
An Example A typical Kubernetes cluster runs an automatic speech recognition (ASR) business
might be running the following jobs:
1) The speech service, with as many instances so to serve many simultaneous user requests.
2) The Kafka system, whose each channel collects a certain log stream of the speech service.
3) Kafka channels are followed by Storm jobs for online data processing. For example, a Storm job joins the
utterance log stream and transcription stream.
4) The joined result, namely session log stream, is fed to an ASR model trainer that updates the model.
5) This trainer notifies ASR server when it writes updated models into Ceph.
6) Researchers might change the training algorithm, and run some experiment training jobs, which serve testing
ASR service jobs.
The famous 'classical big data' on Spark
Apache Spark has emerged as the de facto framework for big data
analytics with its advanced in-memory programming model and upper-level
libraries for scalable machine learning, graph analysis, streaming and
structured data processing. It is a general-purpose cluster computing
framework with language-integrated APIs in Scala, Java, Python and R. As
a rapidly evolving open source project, with an increasing number of
contributors from both academia and industry, it is difficult for researchers
to comprehend the full body of development and research behind Apache
Spark, especially those who are beginners in this area.

In this paper, we present a technical review on big data analytics using

http://dx.doi.org/10.1007/s41060-016-0027-9
Apache Spark. This review focuses on the key components, abstractions
and features of Apache Spark. More specifically, it shows what Apache
Spark has for designing and implementing big data algorithms and
pipelines for machine learning, graph analysis and stream processing. In
addition, we highlight some research and development directions on
Apache Spark for big data analytics.

In addition to the research highlights we presented in the

previous sections, there are other research works which have
been done using Apache Spark as a core engine for solving
data problems in machine learning and data mining [5,36],
graph processing [16], genomic analysis [60,65], time series
data [71], smart grid data [73], spatial data processing [87],
scientific computations of satellite data [67], large-scale
biological sequence alignment [97] and data discretization
[68]. There are also some recent works on using Apache
Spark for deep learning [46,64]. CaffeOnSpark is an open
source project [60] from Yahoo [61] for distributed deep
learning on big data with Apache Spark.
“BIG Data” Frameworks
Apache spark for example
 Tensorflow + Apache Spark
You might be wondering: what’s Apache Spark’s use here when most
high-performance deep learning implementations are single-node
only? To answer this question, we walk through two use cases and
explain how you can use Spark and a cluster of machines to improve
deep learning pipelines with TensorFlow:

Hyperparameter Tuning: use Spark to find the best

How does using Spark improve the
set of hyperparameters for neural network training,
accuracy? The accuracy with the
default set of hyperparameters is leading to 10X reduction in training time and 34%
99.2%. Our best result with
hyperparameter tuning has a
99.47% accuracy on the test set,
lower error rate.
which is a 34% reduction of the
test error. Distributing the
computations scaled linearly with
the number of nodes added to the
Deploying models at scale: use Spark to apply a
cluster: using a 13-node cluster, we
were able to train 13 models in trained neural network model on a large amount of
parallel, which translates into a 7x
speedup compared to training the
models one at a time on one
data.
machine.

http://advancedspark.com/:

https://www.youtube.c
om/watch?v=PFK6gsnlV5 https://www.meetup.com/Advanced-Spark-and-
E TensorFlow-Meetup/

The goal of this workshop is to build an end-to-end, streaming data analytics and recommendations pipeline on your local machine using Docker and the latest streaming analytics
tools. First, we create a data pipeline to interactively analyze, approximate, and visualize streaming data using modern tools such as Apache Spark, Kafka, Zeppelin, iPython, and
ElasticSearch.
Dask as an alternative to apache spark #1

https://youtu.be/1kkFZ4P-XHg

Matthew Rocklin's Blog

dask, the original project
dask.distributed, the distributed memory scheduler powering the cluster computing
dask.bag, the user API we’ve used in this post.

Amazon EC2 with Dask configured with Jupyter Notebooks, and

Anaconda. https://github.com/dask/dask-ec2
continuum.io/blog/developer-blog/high-performance-hadoop-anaconda-and-dask-your-cluster
Dask as an alternative to apache spark #2
https://news.ycombinator.com/item?id=10062076

http://dask.pydata.org/en/latest/spark.html https://www.quora.com/Is-https-github-com-blaze-dask-an-alternative-to-Spark

Dask seems to be aimed at parallelism of only certain operations (some parts of NumPy and
Spark is mature and all-inclusive. If you want a single project that does everything and you’re Pandas) on larger than memory data on a single machine. Spark is a general purpose
already on Big Data hardware then Spark is a safe bet, especially if your use cases are typical computing engine that can work across a cluster of machines and has many libraries
ETL + SQL and you’re already using Scala. optimized for distributed computing (machine learning, graph, etc.).

Dask is lighter weight and is easier to integrate into existing code and hardware. If The advantages of Dask seem to be that it is a drop in replacement for NumPy and Pandas.
your problems vary beyond typical ETL + SQL and you want to add flexible parallelism to Granted, given the prevalence of those two libraries that isn't a small advantage.
existing solutions then dask may be a good fit, especially if you are already using Python
and associated libraries like NumPy and Pandas.

If you are looking to manage a terabyte or less of tabular CSV or JSON data then you
should forget both Spark and Dask and use Postgres or MongoDB.

TensorFlow Basics
Weights Persistence. Save and Restore a model.
Fine-Tuning. Fine-Tune a pre-trained model on a new task.
Using HDF5. Use HDF5 to handle large datasets. GPU Computing with Apache Spark and Python
Using DASK. Use DASK to handle large datasets. by Continuum Analytics, .slideshare.net
Kubernetes + Dask
Running on kubernetes on google Dask Cluster Deployments
container engine http://matthewrocklin.com/blog/work/2016/09/22/cluster-deployments

This work is supported by Continuum Analytics and the 

This small repo gives an example Kubernetes XDATA Program as part of the Blaze Project
configuration for running dask.distributed on Google All code in this post is experimental. It should not be relied
Container Engine. upon. For people looking to deploy dask.distributed on a
cluster please refer instead to the documentation instead.
Dask is deployed today on the following systems in the wild:
• SGE
SciPy Tutorial Setup On Kubernetes • SLURM,
written by Benjamin Zaitlen on 2016-09-30 • Torque
http://quasiben.github.io/blog/2016/9/30/scipy-setup/ • Condor
Our goal was to give students access to a preconfigured • LSF
cluster with zero entry requirements: push a button get a • Mesos
cluster with all tools installed. To accomplish this we need • Marathon
a handful of docker images:
• Kubernetes
• Web application: button and info • SSH and custom scripts
• Jupyter notebook … there may be more. This is what I know of first-hand.
These systems provide users access to cluster resources
• proxy app (more on this later) and ensure that many distributed services / users play nicely
• cluster technologies: Spark, Dask, IPython Parallel together. They’re essential for any modern cluster
deployment.
And a handful of Kubernetes concepts:
For example, both Olivier Griesl (INRIA, scikit-learn) and 
• Pods: collection of containers (similar to docker-compose) Tim O’Donnell (Mount Sinai, Hammer lab) publish
instructions on how to deploy Dask.distributed on 
• namespaces: named and isolated clusters
Kubernetes.
• replication controller: a scalable Pod. • Olivier’s repository
• Tim’s repository
Managing Data
Storage layer
That is code What about data then?

recorditblog.com
Using the different software above, an application can be deployed, scaled
easily and accessed from the outside world in few seconds. But, what about
the data? Structured content would probably be stored in a distributed
database, like MongoDB, for example Unstructured content is traditionally
stored in either a local file system, a NAS share or in Object Storage. A local
file system doesn’t work as a container can be deployed on any node in the
http://www.slideshare.net/kubecon/kubecon-eu-2016-kubernetes-storage-101
cluster.

On the other side, Object Storage can be used by any application from any
container, is highly available due to the use of load balancers, doesn’t require
any provisioning and accelerate the development cycle of the applications.
Why ? Because a developer doesn’t have to think about the way data should
be stored, to manage a directory structure, and so on.

The Amazon S3 endpoint used to upload and download pictures is displayed

on the bottom left corner and shows that ViPR is used to store the data.

The fact that the picture is uploaded directly to the Object Storage platform
means that the web application is not in the data path. This allows the
application to scale without deploying hundreds of instances. This web
application can also be used to display all the pictures stored in the
corresponding Amazon S3 bucket.
The url displayed below each picture shows that the picture is downloaded
directly from the Object Storage platform, which again means that the web
application is not in the data path. This is another reason why Object
Storage is the de facto standard for web scale applications.
http://kubernetes.io/docs/user-guide/persistent-volumes/walkthrough/
Persistent Volumes Walkthrough
The purpose of this guide is to help you become familiar with Kubernetes Persistent Volumes.
By the end of the guide, we’ll have nginx serving content from your persistent volume.
You can view all the files for this example in the docs repo here.
This guide assumes knowledge of Kubernetes fundamentals and that you have a cluster up and
running.
See Persistent Storage design document for more information.
Data Lakes vs data warehouses #1
“A data lake is a storage repository that holds a vast amount of raw data in its
native format, including structured, semi-structured, and unstructured data. The
data structure and requirements are not defined until the data is needed.”

The table below helps flesh out this definition. It also highlights a few of the key www.kdnuggets.com/2015/09
differences between a data warehouse and a data lake. This is, by no means, an
exhaustive list, but it does get us past this “been there, done that” mentality:

http://www.smartdatacollective.com/all/13556

Data. A data warehouse only stores data that has been modeled/structured, while a data lake is no respecter of data. It stores it all—structured, semi-structured, and unstructured. [See my
big data is not new graphic. The data warehouse can only store the orange data, while the data lake can store all the orange and blue data.]

Processing. Before we can load data into a data warehouse, we first need to give it some shape and structure—i.e., we need to model it. That’s called schema-on-write. With a data lake, you just load
in the raw data, as-is, and then when you’re ready to use the data, that’s when you give it shape and structure. That’s called schema-on-read. Two very different approaches.

Storage. One of the primary features of big data technologies like Hadoop is that the cost of storing data is relatively low as compared to the data warehouse. There are two key reasons for this: First,
Hadoop is open source software, so the licensing and community support is free. And second, Hadoop is designed to be installed on low-cost commodity hardware.

Agility. A data warehouse is a highly-structured repository, by definition. It’s not technically hard to change the structure, but it can be very time-consuming given all the business processes that are tied
to it. A data lake, on the other hand, lacks the structure of a data warehouse—which gives developers and data scientists the ability to easily configure and reconfigure their models, queries, and
apps on-the-fly.

Security. Data warehouse technologies have been around for decades, while big data technologies (the underpinnings of a data lake) are relatively new. Thus, the ability to secure data in a data
warehouse is much more mature than securing data in a data lake. It should be noted, however, that there’s a significant effort being placed on security right now in the big data industry. It’s not a
question of if, but when.

Users. For a long time, the rally cry has been BI and analytics for everyone! We’ve built the data warehouse and invited “everyone” to come, but have they come? On average, 20-25% of them have. Is it
the same cry for the data lake? Will we build the data lake and invite everyone to come? Not if you’re smart. Trust me, a data lake, at this point in its maturity, is best suited for the data scientists.
Data Lakes Medical Examples

searchhealthit.techtarget.com

Unlike most relational databases' linear representation and analysis of

data, Franz's semantic graph database technology employs with which
users can graphically see data elements and their relationships.

Setting Up the Data Lake

http://www.slideshare.net/CasertaConcepts/setting-up-the-data-lake-55319460

Montefiore also recently started another program using the data lake to do cardio-
genetic predictive analytics to determine the degrees of possibility of patients having
sudden cardiac death based on their genetic background.
Deploying Deep learning models
Small scale interference
Data science pipeline development and deployment

https://www.continuum.io/blog/developer-blog/productionizing-and-deploying-data-science-projects

Note! Academia has been notoriously slow

to adopt new technologies in academic
research, and you should not be afraid of the
use of “business analyst”
academia.
Dockerizing analysis research
pipelines for end-to-end
reproducibility would be an
excellent way reducing “sloppy”
science.

Peer reviewers (or even the

journals automatically) could
potentially feed “standardized”
datasets via REST APIs at
various points of the pipeline to
ensure that the methodologies
used are of “good science”
Minimal r&D ready for production
Deep Learning with Keras on Google Compute Engine
by Cole Murray. Software Engineer. Mobile & Full-Stack Engineering. Machine Learning.
https://medium.com/google-cloud/keras-inception-v3-on-google-compute-engine-a54918b0058

“Keras Inception V3 image classification model Prediction with deployment on Compute Engine
w/ Docker & Google Container Registry (like Docker Hub), using Flask Front-end Webserver.”

PREDICT (Inference) Front-End

e.g. “Whether the image is of a cat or a dog?” e.g. “How to upload a photo from your phone,
and predict what it contains?”

When you have developed and trained your model, the

prediction part in development becomes rather easy

Note! If you are using TensorFlow (TF) backend, you can
use the TF computation graph (i.e. trained model)
trained “natively” on TensorFlow with the Keras
deployment for example when your data scientist(s)
keep on re-training the model, and you can then just
replace the computation graph (.ckpt, checkpoint) with
Keras, Tensorflow, Docker, Flask, Gunicorn, Nginx, the new weights
Supervisor tech stack in deployment
Flask web server allows fast creation of front-end
which you can use for example for quick demo of your
minimum viable product (MVP), progress report for
your boss, interactive front-end for showing scientific
results for your PI (principal investigator).
TensorFlow Serving Tensorflow for production
How to deploy Machine Learning models with TensorFlow. Part 1 —
make your model ready for serving.
https://medium.com/towards-data-science/how-to-deploy-machine-learning-models-with-tensorflow-part-1-make-your-model-ready-for-serving-776a14ec3198
By Vitaly Bezgachev
Streaming Machine learning Google
How to use Google Cloud Dataflow with TensorFlow for batch
predictive analysis
Justin Kestelyn, Google Cloud Platform, April 18, 2017
https://cloud.google.com/blog/big-data/2017/04/how-to-use-google-cloud-dataflow-with-tensorflow
-for-batch-predictive-analysis

This article shows you how to use Cloud Dataflow to run

batch processing for machine learning predictions. The article
uses the machine learning model trained with TensorFlow.
The trained model is exported into a Google Cloud Storage
bucket before batch processing starts. The model is
dynamically restored on the worker nodes of prediction jobs.
This approach enables you to make predictions against a
Using Cloud Storage as a data source.
large dataset, stored in a Cloud Storage bucket or Google
BigQuery tables, in a scalable manner, because Cloud
Dataflow automatically distributes the prediction tasks to
multiple worker nodes.
Cloud Dataflow is a unified programming model and a
managed service for developing and executing a wide range
of data processing patterns including ETL (Extract, Transform,
Load), batch computation, and continuous computation.
Cloud Dataflow frees you from operational tasks such as
resource management and performance optimization.

Using BigQuery as a data source.

Automated Kubernetes GPU Deployment #1 with tensorflow
How to automate deep learning training with Kubernetes GPU-cluster
https://github.com/Langhalsdino/Kubernetes-GPU-Guide
By Frederic Tausch

Why did I write this guide?

I have worked as in intern for the startup understand.ai and
noticed the hassle of firstly designing a machine learning
algorithm locally and then bringing it to the cloud for
training with different parameters and datasets.
The second part, bringing it to the cloud for extensive
training, takes always longer than thought, is frustrating and
involves usually a lot of pitfalls.
For this reason I decided to work on this problem and make
the second part effortless, simple and quick. The result of
this work is this handy guide, that describes how everyone
Automated Kubernetes GPU Deployment #2 with tensorflow
GPUs & Kubernetes for Deep Learning
By Samuel Cozannet

Deploy Kubernetes with GPUs

● Deploy k8s on AWS in a development mode (no HA,

colocating etcd, the control plane and PKI)

● Deploy 2x nodes with GPUs (p2.xlarge and p2.8xlarge

instances)

● Deploy 3x nodes with CPU only (m4.xlarge)

● Validate GPU availability

Add EFS storage to the cluster So what is a Deep Learning pipeline exactly? Well, my definition is a 4 step pipeline, with a
potential retro-action loop, that consists of :
● Programmatically add an EFS File System and Mount
Points into your nodes Data Ingest: This step is the accumulation and pre processing of the data that will be used to
train our model(s). This step is maybe the less fun, but it is one of the most important. Your model
● Verify that it works by adding a PV / PVC in k8s. will be as good as the data you train it on
Automating Tensorflow deployment Training: this is the part where you play God and create the Intelligence. From your data, you will
create a model that you hope will be representative of the reality and capable of describing it (and
● Data ingest code & package even why not generate it)
Training code & package
●
Evaluation + Monitoring: Unless you can prove your model is good, it is worth just about
Evaluation code & package nothing. The evaluation phase aims at measuring the distance between your model and reality.
This can then be fed back to a human being to adjust parameters. In advanced setups, this can
●

Serving code & package essentially be your test in CI/CD of the model, and help auto tune the model even without human
intervention
●

● Deployment process
Serving: there is a good chance you will want to update your model from time to time. If you want
to recognize threats on the network for example, it is clear that you will have to learn new malware
signatures and patterns of behavior, or you can close your business immediately. Serving is this
phase where you expose the model for consumption, and make sure your customers always enjoy
the latest model.
Jupyter notebooks vs IDE development
What is the practical difference?
Jupyter notebooks What are they all about
If you are a beginner in data science (and with Python) you might have noticed that there
are a lot of tutorial implemented as Jupyter notebooks (used to be referred as ipython).
The notebooks allow easy embedding of formatted text and images to executable code, which
make then very useful to provided with scientific papers and walkthrough example code.

Idea of a notebook
Both in academia (e.g. non-tech savvy old-school PI vs. new generation
data wizard) and in industry, one may want to communicate
concisely the rationale for the computing with the results.
Compare this to “Excel data science”, where a person manually
manipulates the data destroying end-to-end processing pipeline and
making reproducible research impossible.
In Jupyter notebook, one could include both data preparation and
data analysis parts, and the grad student could for example do the
dirty work and ask for insight from more senior researchers.
In theory, the collaborators would quickly realize that the notebook is
about, and allowing quick playing around that would easily usable by
the poor grad student as well eliminating the Excel →
Matlab/R/Python → Excel circus probably experienced in some
research labs.
Downside is that the notebooks are not ‘plug’n’play’ executables, and
can depend on a lot of standard libraries, are even worse for a non-
tech savvy PI, custom libraries written by the grad student
Jupyter notebook Features

Sep 10, 2016 • Alex Rogozhnikov

http://arogozhnikov.github.io/2016/09/10/jupyter-features.html
Jupyter notebooks Examples of capabilities
A gallery of interesting Jupyter Notebooks
David Haberthür
“Traditional” notebook example
Colour science computations with colour, a Python
package implementing a comprehensive number of colour
theory transformations and algorithms supported by a
dedicated collection of IPython Notebooks. More colour
science related IPython Notebooks are available on
colour-science.org.
Data-driven journalism
The Need for Openness in Data Journalism, by
Brian Keegan.
St. Louis County Segregation Analysis ,
analysis for the article
The Ferguson Area Is Even More Segregated Than
You Probably Guessed
by Jeremy Singer-Vine.
Reproducible academic publications
This section contains academic papers that have been published in the peer-reviewed literature or
pre-print sites such as the ArXiv that include one or more notebooks that enable (even if only
partially) readers to reproduce the results of the publication.
4) The Paper of the Future by Alyssa Goodman et al. (Authorea Preprint, 2017). This article
explains and shows with demonstrations how scholarly "papers" can morph into long-lasting
rich records of scientific discourse, enriched with deep data and code linkages, interactive
figures, audio, video, and commenting. It includes an interactive d3.js visualization and has an
astronomical data figure with an IPYthon Notebook "behind" it.
A 5-minute video demonstration of this paper is available at this YouTube link.
5 Better “Science Storytelling”
As we stated at the outset, communicating results by way of what cognitive scientists refer to
as "storytelling" has the deepest, most long-lasting, impact on a reader, viewer, or listener.
Until recently, journal articles only contained words, numbers, and pictures, but today we can
enhance journal articles' storytelling potential with audio, video, and enhanced figures that
offer interactivity and context. We consider each of these opportunities in turn, below.
Screen shot of the first 3D PDF published in Nature in
2009 (Goodman 2009). A video demonstration of how
users can interact with the figure is on YouTube, here.
Open the PDF here in Adobe Acrobat to interact.
Jupyter notebooks how do they compare to IDEs
“Real-life” development typically a lot easier with a
proper IDE (Integrated Development Environment) then

https://www.slant.co/versus/1240/15716/~pycharm_vs_jupyter

Compare to the Rstudio IDE and its notebooks You can use Jupyter notebooks also with
PyCharm
https://www.jetbrains.com/help/pycharm/using-ipython-jupyter-notebook-with-pycharm.html
IDE PyCharm
PyCharm the most commonly used IDE, and could be a safe bet to start with.

Jun 28, 2017

datacamp.com http://noeticforce.com/best-python-ide-for-programmers-windows-and-mac
By Paulo Vasconcellos
Top 5 Python IDEs For Data Science

https://www.kaggle.com/general/4308

Rodeo IDE has the feel of RStudio if you are coming from R

https://www.r-bloggers.com/
rstudio-clone-for-python-ro By Erik Marsja
deo/ http://www.marsja.se/rstudio-like-python
-ides-rodeo-spyder/
Deploying Deep learning models
Large-scale interference (+Training) with streaming
 Streaming Machine learning ”Heavier” tech stacks

“STREAMING”
i.e. continuous stream of data
for example from credit card
transactions, Uber cars, Internet
of Things devices such as
electricity meters or next-
generation vital monitoring at
hospitals.

End to End Streaming ML Recommendation Pipeline Spark 2 0, Kafka, TensorFlow Workshop

by Chris Fregly on 10/12/2016.

https://www.youtube.com/watch?v=UmCB9ycz55Q | https://github.com/fluxcapacitor/pipeline/wiki
Streaming Machine learning Confluent #1
R, Spark, Tensorflow, H20.ai Applied to Streaming Analytics
Kai Wähner, Technology Evangelist at Confluent, Published on Mar 24, 2017
https://www.slideshare.net/KaiWaehner/r-spark-tensorflow-spark-applied-to-streaming-analytics

Big Data vs. Fast Data SQL Server, Oracle, MySQL, Teradata, Netezza, JDBC/ODBC, Hadoop, SFDC, PostgreSQL, RapidMiner,
TIBCO Spotfire
Streaming Machine learning Confluent #2a
Machine Learning and Deep Learning Applied to Real Time with
Apache Kafka Streams
Kai Wähner, Technology Evangelist at Confluent, Published on May 23, 2017
https://www.slideshare.net/KaiWaehner/apache-kafka-streams-machine-learning-deep-learning

https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf

Kafka Cluster Kafka Cluster

Deployed anywhere:
Docker, Kubernetes,
Mesos, Java App, ...
Streaming Machine learning Confluent #2b

Continuously train and improve the model

with every new event

How to improve models?

1) Manual Update
2) Automated Batch
3) Real Time
Streaming Machine learning Databricks #1

https://www.slideshare.net/databricks/integrating-deep-learning-libraries-with-a
pache-spark

https://github.com/databricks/spark-deep-learning

One way to productionize a model is to deploy it as a

Spark SQL User Defined Function, which allows anyone who
knows SQL to use it. Deep Learning Pipelines provides
mechanisms to take a deep learning model and register a Spark
SQL User Defined Function (UDF).

The resulting UDF takes a column (formatted as a image struct

"SpImage") and produces the output of the given Keras model
(e.g. for Inception V3, it produces a real valued score vector over
the ImageNet object categories).
Streaming Machine learning Databricks #2
A Vision for Making Deep Learning Simple
From Machine Learning Practitioners to Business Analysts
by Sue Ann Hong, Tim Hunter and Reynold Xin Posted in ENGINEERING BLOGJune 6, 2017
Deploying Models in SQL
Once a data scientist builds the desired model, Deep Learning Pipelines makes
it simple to expose it as a function in SQL, so anyone in their organization can
use it – data engineers, data scientists, business analysts, anybody.
sparkdl.registerKerasUDF("img_classify",
"/mymodels/dogmodel.h5")
Next, any user in the organization can apply prediction in SQL:
SELECT image, img_classify(image) label FROM images
WHERE contains(label, “Chihuahua”)
Similar functionality is also available in the DataFrame programmatic API
across all supported languages (Python, Scala, Java, R). Similar to scalable
prediction, this feature works in both batch and structured streaming.
SIMPLIFY MACHINE LEARNING WITH APACHE SPARK
Read the White Paper
Download our Machine Learning Starter Kit
Getting the best results in deep learning requires
experimenting with different values for training
parameters, an important step called hyperparameter
tuning. Since Deep Learning Pipelines enables
exposing deep learning training as a step in Spark’s
machine learning pipelines, users can rely on the
hyperparameter tuning infrastructure already built into
Spark.
Conclusion
In this blog post, we introduced Deep Learning Pipelines, a new
library that makes deep learning drastically easier to use and scale.
While this is just the beginning, we believe Deep Learning Pipelines
has the potential to accomplish what Spark did to big data: make the
deep learning “superpower” approachable for everybody.
Future posts in the series will cover the various tools in the library in
more detail: image manipulation at scale, transfer learning,
prediction at scale, and making deep learning available in SQL.
To learn more about the library, check out the Databricks notebook
as well as the github repository. We encourage you to give us
feedback. Or even better, be a contributor and help bring the power
of scalable deep learning to everyone.
Streaming Machine learning StreamAnalytix
Impetus Technologies Unveils New, TensorFlow-Based Deep The combination of streaming analytics and deep learning enables a new
Learning Feature on Apache Spark for StreamAnalytix breed of applications and machine capabilities in industrial IoT, voice
LOS GATOS, Calif., June 15, 2017 /PRNewswire/ analytics and anomaly detection.
https://streamanalytix.com/industryBuzz/PR_15_06_2017
https://streamanalytix.com/architecture
Streaming Machine learning linagora
Making Image Classification Simple With Spark Deep Learning
Zied Sellami, Jun 28 2017
https://medium.com/linagora-engineering/making-image-classification-simple-with-spark-deep-l
earning-f654a8b876b8

Run pyspark with spark-deep-learning library Apache Spark is an open-source cluster-computing

framework. Apache Spark is a fast, in-memory data
spark-deep-learning library comes from Databricks and leverages Spark for its two processing engine with expressive development APIs to
strongest facets: allow data workers to efficiently execute streaming,
machine learning or SQL workloads that require fast
iterative access to datasets.
In the spirit of Spark and Spark MLlib, it provides easy-to-use APIs that enable deep
learning in very few lines of code.
Conclusion
It uses Spark’s powerful distributed engine to scale out deep learning on massive
datasets.
Apache Spark is a very powerful platform with elegant and
expressive APIs to allow Big Data processing.
The library run on pyspark. Let’s start pyspark:
We tried with success Spark Deep Learning, an API that
export SPARK_HOME=PATH/TO/spark-2.1.1-bin-hadoop2.7 combine Apache Spark and Tensorflow to train and deploy
export set JAVA_OPTS="-Xmx9G -XX:MaxPermSize=2G -XX:+UseCompressedOops an image classifier. It is extremely easier (less than 30 lines
-XX:MaxMetaspaceSize=512m" of code). Our next objective is to test if we can deploy facial
$SPARK_HOME/bin/pyspark --packages databricks:spark-deep-learning:0.1.0- recognition model with this API
spark2.1-s_2.11 --driver-memory 5g
While this support is only available on Python, we hope that
Below is a console output example.
integration will be done very soon on other programming
languages especially with Scala.

Deep Learning Pipelines on Apache Spark enables fast transfer learning with a
Featurizer (that transform images to numeric features).
from pyspark.ml.classification import LogisticRegression
from pyspark.ml import Pipeline
from sparkdl import DeepImageFeaturizer
featurizer = DeepImageFeaturizer(inputCol="image", outputCol="features",
modelName="InceptionV3")
lr = LogisticRegression(maxIter=20, regParam=0.05, elasticNetParam=0.3,
labelCol="label")
p = Pipeline(stages=[featurizer, lr])
p_model = p.fit(train_df)
view rawmodelCreator.py hosted with ❤ by GitHub
Industry
Use cases
Kubernetes in deep learning #1
https://openai.com/blog/infrastructure-for-deep-learning/

Deep learning is an empirical

science, and the quality of a
Once the model shows sufficient promise, you'll scale it up to larger datasets and more GPUs. This requires
group's infrastructure is a long jobs that consume many cycles and last for multiple days. You'll need careful experiment management,
and to be extremely thoughtful about your chosen range of hyperparameters.
multiplier on progress.
Like much of the deep learning community, we use Python 2.7. We generally use Anaconda, which has
Fortunately, today's open-source convenient packaging for otherwise difficult packages such as OpenCV and performance optimizations for
ecosystem makes it possible for some scientific libraries. We also run our own physical servers, primarily running Titan X GPUs. We expect to
have a hybrid cloud for the long haul: it's valuable to experiment with different GPUs, interconnects, and other
anyone to build great deep learning techniques which may become important for the future of deep learning.

infrastructure. Scalable infrastructure often ends up making the simple cases harder. We put equal effort into our
infrastructure for small- and large-scale jobs, and we're actively solidifying our toolkit for making distributed
use-cases as accessible as local ones.
In this post, we'll share how deep
Kubernetes requires each job to be a Docker container, which gives us dependency isolation and
learning research usually code snapshotting. However, building a new Docker container can add precious extra seconds to a

proceeds, describe the researcher's iteration cycle, so we also provide tooling to transparently ship code from a researcher's laptop
into a standard image. We expose Kubernetes's flannel network directly to researchers' laptops, allowing
infrastructure choices we've made to users seamless network access to their running jobs. This is especially useful for accessing monitoring
services such as TensorBoard. (Our initial approach — which is cleaner from a strict isolation perspective —
support it, and open-source required people to create a Kubernetes Service for each port they wanted to expose, but we found that it
added too much friction.)
kubernetes-ec2-autoscaler, a batch-
We're releasing kubernetes-ec2-autoscaler, a batch-optimized scaling manager for Kubernetes. It runs as a
optimized scaling manager for normal Pod on Kubernetes and requires only that your worker nodes are in Auto Scaling groups.
Kubernetes. We hope you find this
Our infrastructure aims to maximize the productivity of deep learning
post useful in building your own researchers, allowing them to focus on the science. We're building tools to
deep learning infrastructure. further improve our infrastructure and workflow, and will share these in upcoming
weeks and months. We welcome help to make this go even faster!
Kubernetes in deep learning #2
https://news.ycombinator.com/item?id=12391505

May 9 INFRA · DATA · RESEARCH · NEWS FEED · PYTHON

Introducing FBLearner Flow: Facebook's AI backbone

Jeffrey Dunn, https://code.facebook.com/posts/1072626246134461/introducing-fblearner-flow-facebook-s-ai-backbone/
Automated Kubernetes CPU Deployment Baidu Case
Baidu's deep learning framework adopts Kubernetes
Google's orchestration framework will provide smart resource allocation and cluster management to PaddlePaddle
http://www.infoworld.com/article/3167608/artificial-intelligence/baidus-deep-learning-framework-adopts-kubernetes.html
http://blog.kubernetes.io/2017/02/run-deep-learning-with-paddlepaddle-on-kubernetes.html
Why Run PaddlePaddle on Kubernetes
PaddlePaddle is designed to be slim and independent of computing
infrastructure. Users can run it on top of Hadoop, Spark, Mesos,
Kubernetes and others.. We have a strong interest with Kubernetes
because of its flexibility, efficiency and rich features.
A successful deep learning project includes both the research and
the data processing pipeline. There are many parameters to be tuned.
A lot of engineers work on the different parts of the project
simultaneously.
To ensure the project is easy to manage and utilize hardware resource
efficiently, we want to run all parts of the project on the same
infrastructure platform.
The platform should provide:
● fault-tolerance. It should abstract each stage of the pipeline as a
service, which consists of many processes that provide high
throughput and robustness through redundancy.
● auto-scaling. In the daytime, there are usually many active users,
the platform should scale out online services. While during nights, the
platform should free some resources for deep learning experiments.
● job packing and isolation. It should be able to assign a
PaddlePaddle trainer process requiring the GPU, a web backend
service requiring large memory, and a CephFS process requiring disk
IOs to the same node to fully utilize its hardware.
What we want is a platform which runs the deep learning system, the
Web server (e.g., Nginx), the log collector (e.g., fluentd), the distributed
queue service (e.g., Kafka), the log joiner and other data processors
written using Storm, Spark, and Hadoop MapReduce on the same
cluster.
We want to run all jobs -- online and offline, production and experiments
-- on the same cluster, so we could make full utilization of the cluster, as
different kinds of jobs require different hardware resource.
Docker customers

Published on Aug 11, 2016

In this video Ajay Dankar, Senior Director Product Management at

PayPal discusses why they selected Docker and Docker Trusted
Registry to help them containerize their legacy apps to more
efficiently utilize their infrastructure and secure workloads.

--

Docker is an open platform for developers and system administrators to build, ship
and run distributed applications. With Docker, IT organizations shrink application
delivery from months to minutes, frictionlessly move workloads between data
centers and the cloud and can achieve up to 20X greater efficiency in their use of
computing resources. Inspired by an active community and by transparent, open
source innovation, Docker containers have been downloaded more than 700
million times and Docker is used by millions of developers across thousands of the
world’s most innovative organizations, including eBay, Baidu, the BBC, Goldman
Sachs, Groupon, ING, Yelp, and Spotify. Docker’s rapid adoption has catalyzed an
active ecosystem, resulting in more than 180,000 “Dockerized” applications, over
40 Docker-related startups and integration partnerships with AWS, Cloud Foundry,
Google, IBM, Microsoft, OpenStack, Rackspace, Red Hat and VMware.

https://www.youtube.com/watch?v=wf4Jg-9gv9Q
Business data into value
8 ways to turn data into value with Apache Spark machine learning
OCTOBER 18, 2016 by Alex Liu Chief Data Scientist, Analytics Services, IBM
http://www.ibmbigdatahub.com/blog/8-ways-turn-data-value-apache-spark-machine-learning
1. Obtain a holistic view of business

In today's competitive world, many corporations work hard to gain a holistic view or a 360
degree view of customers, for many of the key benefits
as outlined by data analytics expert Mr. Abhishek Joshi. In many cases, a holistic view was not
obtained, partially due to the lack of capabilities to organize huge amount of data and then to
analyze them. But Apache Spark’s ability to compute quickly while using data frames to
organize huge amounts of data can help researchers quickly develop analytical models that https://thinkbiganalytics.com/big_data_solutions/data-science/
provide a holistic view of the business, adding value to related business operations. To realize
this value, however, an analytical process, from data cleaning to modeling, must
still be completed.

4. Avoid customer churn by rethinking churn modeling

Losing customers means losing revenue. Not surprisingly, then, companies strive to detect
potential customer churn through predictive modeling, allowing them to implement
interventions aimed at retaining customers. This might sound easy, but it can actually be very
complicated: Customers leave for reasons that are as divergent as the customers themselves http://dx.doi.org/10.1186/s40165-015-0014-6
are, and products and services can play an important, but hidden, role in all this. What’s more,
merely building models to predict churn for different customer segments—and with regard to
different products and services—isn’t enough; we must also design interventions, then
select the intervention judged most likely to prevent a particular customer from departing. Yet
even doing this requires the use of analytics to evaluate the results achieved—and,
eventually, to select interventions from an analytical standpoint. Amid this morass of choices,
Apache Spark’s distributed computing capabilities can help solve previously baffling problems.

5. Develop meaningful purchase recommendations

Recommendations for purchases of products and services can be very powerful when made
appropriately, and they have become expected features of e-commerce platforms, with
many customers relying on recommendations to guide their purchases. Yet developing
recommendations at all means developing recommendations for each customer—or, at the very ebaytechblog: Spark is helping eBay create value from its data, and so the future is bright
least, for small segments of customers. Apache Spark can make this possible by offering the for Spark at eBay. In the meantime, we will continue to see adoption of Spark increase at
distributed computing and streaming analytics capabilities that have become invaluable tools eBay. This adoption will be driven by chats in the hall, newsletter blurbs, product
for this purpose. announcements, industry chatter, and Spark’s own strengths and capabilities.
Open data and Apache Spark

-----Jump to Topic-----
00:00:06 - Workshop Intro & Environment Setup
00:13:06 - Brief Intro to Spark
00:17:32 - Analysis Overview: SF Fire Department Calls for Service
00:23:22 - Analysis with PySpark DataFrames API
00:29:32 - Doing Date/Time Analysis
00:47:53 - Memory, Caching and Writing to Parquet
01:00:40 - SQL Queries
01:21:11 - Convert a Spark DataFrame to a Pandas DataFrame
-----Q & A-----
01:24:43 - Spark DataFrames vs. SQL: Pros and Cons?
01:26:57 - Workflow for Chaining Databricks notebooks into Pipeline?
01:30:27 - Is Spark 2.0 ready to use in production?
https://www.youtube.com/watch?v=iiJq8fvSMPg
Internet of things (IoT)

To proof that our IoT platform is really independent on application environment, we took one IoT
gateway (RaspberryPi 2) from the city project and put into Austin Convention Center during
OpenStack Summit together with IQRF based mesh network connecting sensors that measure
humidity, temperature and CO2 levels. This demonstrates ability that IoT gateway can manage or
collect data from any technology like IQRF, Bluetooth, GPIO, and any other communication
standard supported on Linux based platforms.

We deployed 20 sensors and 20 routers on 3 conference floors with a single active IoT gateway
receiving data from entire IQRF mesh network and relaying it to dedicated time-series database, The following screenshot shows real time CO2 values from different rooms on 2
in this case Graphite. Collector is MQQT-Java bridge running inside docker container managed by floors. Historical graph shows values from Monday. You can easily recognize
Kubernetes. when the main keynote session started and when was the lunch period.
Healthcare
https://www.healthdirect.gov.au/
https://www.youtube.com/watch?v=ePp54ofRqRs
How open source container tech can impact healthcare
At Red Hat, we believe that creating open source platforms allows the tech community
to develop the best software possible. We recently launched a series of films
highlighting the open source movement’s impact on healthcare, including
initiatives that promote open patient data and provide 3D-printed prosthetics.
Health is a great context to start exploring OpenShift’s open source capabilities. We
designed OpenShift to allow developers to take full advantage of containers (Docker)
and orchestration (Kubernetes), without having to learn the internals of how to build
containers from scratch or understand sys admin enough to deploy production-quality
apps that can scale on demand.
OpenShift makes using containers and orchestration accessible by letting you focus
on code instead of writing Dockerfiles and running Docker builds all day. With the
integrated Source-to-Image open source project, the platform automatically creates
containers while requiring only the URL for your source code repository.
openshift.devpost.com
The logical resource boundaries established in Docker containers are almost as secure as
those established by the Linux operating system or by a virtual machine, according to a
report by Gartner analyst Joerg Fritsch. However, Docker and Linux containers in general fall
short when it comes to container management and administration, Fritsch said in his report, "
Security properties of containers managed by Docker."
Improving Container Security: Docker and More After 6 months and 15
successful beta deployments, Twistlock is announcing the general availability of our
container security suite. Twistlock came out of stealth in May 2015. Since then,
we have been working diligently with a select group of beta customers to validate the
value of our offerings. This diverse group of 15 beta testers, including Wix,AppsFlyer,
and HolidayCheck, spans financial services, hospitality, healthcare, Internet services,
and government. These customers confirmed that we are hitting the sweet spot of
their most pressing container security needs -- a majority of them already deployed
our product into their production environments, protecting live services and customer
data.
Academia
Use cases
 Neuroscience & bioinformatics

http://dx.doi.org/10.1016/j.conb.2015.04.002
“Cloud deployment also makes it easier to build tools that run identically for all users,
especially with virtual machine platforms like Docker. However, cloud deployment for
neuroscience does require transferring data to cloud storage, which may become a bottleneck.
Deploying on academic clusters requires at least some support from cluster administrators but
keeps the data closer to the computation. … There is also rapidly growing interest in the ‘‘data
analysis notebook’’. These notebooks – the Jupyter notebook being a particularly popular
example – combine executable code blocks, notes, and graphics in an interactive document that
runs in a web browser, and provides a seamless front-end to a computer, or a large https://www.docker.com/customers/docker-helps-varian-medical-systems-battle-cancer
cluster of computers if running against a framework like Spark. Notebooks are a particularly
appealing way to disseminate information; a recent neuroimaging paper, for example, provided all
of its analyses in a version-controlled repository hosted on GitHub with Jupyter notebooks that
generate all the figures in the paper [45]—a clear model for the future of reproducible
science

https://dx.doi.org/10.12688/f1000research.7536.1

http://dx.doi.org/10.1371/journal.pone.0152686

http://dx.doi.org/10.1186/s13742-015-0087-0

Most large-scale analytics, whether in industry or neuroscience, involve common patterns. Raw data
are massive in size. Often, they are processed so as to extract signals of interest, which are then used for
statistical analysis, exploration, and visualization. But raw data can be analyzed or visualized directly (top
arrow). And the results of each successive step informs how to perform the earlier ones (feedback loops). http://homolog.us/blogs/blog/2015/09/22/is-docker-for-suckers/
Icons below highlight some of the technologies, discussed in this essay, that are core to the modern large-
scale analysis workflow.
 Neuroscience streaming data with Spark

http://dx.doi.org/10.1016/j.conb.2015.04.002

Streaming analysis of two-photon calcium

imaging*
https://youtu.be/uUQTSPvD1mc?t=17m
* With these levels of analysis in mind, we probe the cortical
circuits underlying active tactile decision making. Whisker- Real-time visualization during the experiment rather than doing
based haptic tasks for head-fixed mice developed in our lab the experiment with no good idea of what is going on before post-
provide outstanding stimulus control and facilitate experiment offline analysis
applications of powerful biophysical methods, such as
whole cell recordings and two-photon microscopy. Real-time feedback on the experimental parameters, or for the
in collaboration with Karel Svodoba and Nicholas Sofroniew.
brain itself via optogenetic stimulation for example.
https://www.janelia.org/lab/svoboda-lab
Reproducible SCIENCE

http://t-redactyl.io/blog/2016/10/a-crash-course-in-reproducible-research-in-python.html

http://dx.doi.org/10.1038/nj7622-703a
http://conference.scipy.org/proceedings/scipy2016/pdfs/christian_oxvig.pdf

The use of 'custom MATLAB scripts'

http://dx.doi.org/10.1038/533452a
Reproducible SCIENCE with docker
ANACONDA AND DOCKER
BETTER TOGETHER FOR REPRODUCIBLE DATA SCIENCE
Monday, June 20, 2016, continuum.io/blog

Anaconda integrates with many different providers and platforms to give

you access to the data science libraries you love on the services you use,
including Amazon Web Services, Microsoft Azure, and Cloudera CDH. Today
we’re excited to announce our new partnership with Docker.

As part of the announcements at DockerCon this week, Anaconda images

will be featured in the new Docker Store, including Anaconda and
Miniconda images based on Python 2 and Python 3. These freely available
Anaconda images for Docker are now verified, will be featured in the Docker
Store when it launches, are being regularly scanned for security
vulnerabilities and are available from the
ContinuumIO organization on Docker Hub.

Anaconda and Docker are a great combination to empower

yourdevelopment, testing and deployment workflows with
Open Data Science tools, including Python and R. Our users often ask
whether they should be using Anaconda or Docker for data science
development and deployment workflows. We suggest using both - they’re
better together!
Reproducible SCIENCE Between Jupyter and Docker

Build own condo packages, and deploy

continuum.io/blog/developer-blog/whats-old-and-new-conda-build

Anaconda
Enterprise
Notebooks
continuum.io

Jupyter / JupyterLab does not come really as

'plug'n'play' and you still have to have all the
dependencies resolved