►
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
GPU Machine Learning From Laptop to Cloud - Mark Puddick, Pivotal
In this session Mark will talk about how to setup and build containers to run GPU accelerated Machine Learning workloads on your laptop and move these easily to cloud. The session will cover setting up a local GPU accelerated environment, which can be great for testing on smaller data, then moving this over to run on larger sets in the cloud. Although the concepts in this session will focus on Python and Kersas workloads this can easily be applied to other machine learning technologies.
https://sched.co/MPaW
A
Hi
so
I'm
saying
yank,
just
introduce
me,
I'm
Marc
product
from
pivotal
I've
come
here
today.
I've
actually
worked
for
pivotal
in
Australia
in
Sydney.
So
come
a
fair
way
to
do
this,
talk
I'm,
going
to
talk
about
GPU,
machine
learning
and
from
laptops.
Quite
so,
we're
gonna
have
a
few
demos
in
this
as
well,
and
we're
just
going
to
go
through
basically
a
workflow
from
getting
a
simple
model
and
getting
it
running
in
the
cloud
and
in
some
of
the
nuances
around
that.
A
Okay,
so
there's
a
few
things:
we're
gonna
just
cover
off,
as
I
said
we're
going
to
do
what
why
so,
this
talk
came
from
customer
that
I
was
dealing
with.
They
had
a
large
bare
metal
and
Linux
cluster
that
they
were
using
and
they
wanted
to
actually
who
then
scooped
in
Eddie's
and
also
some
stuff.
That
I
was
doing
some
machine
learning
stuff.
A
So
we'll
go
through
a
quick
bit
about
how
to
get
your
laptop
dev
up
top
up
and
running,
and
because
my
only
other
thing
around
this
is
it's
kind
of
based
as
well
about
workflow
as
well.
If
you
were
going
to
do
some
C
icd
around
this
as
well,
so
we're
quickly
goes
through
how
you
get
a
laptop
up
and
running.
We'll
do
a
couple
of
demos
around
that
and
then
we'll
talk
about
what
happens
when
move.
That's
cloud
and
what's
different
and
then
at
the
end
we've
got
we'll
get
the
conclusion.
A
We
might
have
some
time
for
questions
I'm,
not
sure
how
much
time
will
actually
have
and
when
I
ran
through
this.
It
did
take
about
the
time
of
this
session
is
we
might
not
have
much
time
for
questions,
but
I
will
be
on
the
pivotal
booth
tomorrow
morning
and
you're
welcome
to
come
and
talk
to
me
anytime,
the
province
after
here.
A
A
Hopefully,
people
are
the
things
like
cares,
so
it's
going
to
cover
two
carriers,
tensorflow
and
python-
we're
going
to
talk
about
how
we
docker
eyes
we're
gonna,
get
that
running,
mini
coop
locally,
with
GPU
acceleration
as
well,
and
then
we're
going
to
push
it
up
to
the
Clyde
and
run
it
inside
and
run
that
with
some
GPU
acceleration
and
I'm
not
going
to
cover
off
Jupiter
in
orbits.
You
can
also
use
Jupiter,
notebooks
and
I.
Think
the
tensorflow
GP
docker
image
has
that
built
in
I.
A
Didn't
want
to
do
that,
because
I
wanted
to
be
able
to
put
this
eventually
into
a
kind
of
CI
CD
pipeline
and
demonstrate
that
so
that
stuff
as
well.
Okay,
Nancy,
must
speak
your
nose.
Okay,
so
to
get
your
laptops,
I
was
originally
going
to
go
through
kind
of
a
laptop
install,
but
then
I
realized
it's
a
it's
gonna
be
different
for
everyone
and
B.
It's
going
to
be
kind
of
painful
and
taking
a
long
time.
A
So
it's
probably
not
super
valuable,
so
I'm
just
going
to
get
summarized
what
you
need
to
do
to
get
that
up
and
running.
So
you
need
to
get
the
Nvidia
drivers
installed.
This
is
pretty
obvious
stuff
right,
so
those
are
video
drivers,
there's
proprietary
and
open
source
ones,
I
think
I'm,
currently
using
the
proprietary
ones,
but
you
can
either
you
can
actually
use
either
them
and
they
will
both
work.
There
are
some
challenges
around
that
getting
them
running
mainly
secure
boot
seems
to
cause
a
lot
of
problems
as
well.
A
When
you
start
with
the
CUDA
toolkit
and
the
cDNA
and
stuff
in
which
you'll
also
need
to
do,
I,
probably
just
go
through
the
tents
of
flow
axle
instructions
for
installing
those,
because
it's
the
simplest
way
to
do
it
and
again
going
through
it
in
this
session,
would
probably
be
a
bit
waste
of
it
ones.
Time
and
then
eventually,
once
you've
got
all
that
up
and
running.
A
A
Ok,
so
that's
basically
how
you
laptop
set
up
so
I
said:
I
wanted
to
do
that
quite
quickly,
and
all
that
is
just
to
run
normally
without
doc,
rhizome,
with
opening
mini
to
bring
like
that.
So
so
one
of
the
things
that
you
probably
find
when
you
do
this
is
you
have
a
lot
of
version
conflicts
and
things
like
that,
and
there's
not
really
much
I
can
advice.
I
can
give
you,
apart
from
you'll,
just
have
to
kind
of
work
through
it.
It
can
be
a
little
bit
laborious
and
they
say.
A
Secure
boot
generally
causes
a
few
problems
as
well,
so
personally,
for
this
laptop
I
have
secured
but
off
for
the
moment,
because
I
only
use
it
for
this
so
to
fast
about
having
on
okay.
So
we'll
do
a
quick
example
and
we'll
go
to
you
a
little
bit
cold
quickly,
just
to
show
you
and
what
I'm
going
to
do
this
an
example
today
and
we'll
run
through
how
we
move
that
across
the
different
environments.
Okay,
so
minute.
A
A
A
We've
got
a
it's
a
cut-down
kind
of
each
18
16,
it's
just
in
some
image
recognition
and
what
it's
actually
going
to
do
is
load
up.
A
bunch
of
pictures
of
LEGO
bricks
is
eight
different,
Lego
bricks
and
it's
trying
to
detect
which
one
so
I
thought
it's
made
quite
good
demo.
This
is
something
I
did
ages
ago,
because
we
have,
we
were
trying
to
categorize
some
Lego
for
my
kids,
which
I
won't
go
into
in
this.
A
So
you
can
see
it's
really
simple:
there's
just
a
train
model,
Scott
the
optimizer
there,
and
then
we
just
fit
it
and
it's
going
to
output
it
so
we'll
run
it
out
quickly
and
you
can
just
see
what
it
does.
I've
only
going
to
run
this
with
like
three
epochs
needs
to
do
much
just
so
that
we
can
run
it
quickly
in
the
demo.
A
A
A
Here
we
go
it's
a
little
bit
bad,
so
it's
just
gonna
run
through
it
will
run.
It's
gonna
use
the
GPU
for
this.
So
it's
going
to
run
a
lot
quicker.
I
generally
you'll
find
it's
about
10
to
20
times
quicker,
I
reckon
when
you'd
run
it
from
just
perfectly
a
pure
CPU
on
this
machine.
So
that's
going
to
run
through
it's
only
gonna
run
through
epochs
and
you'll,
see
that
I'll
come
back
to
the
slides.
A
A
Ok,
so
we've
run
that.
So
what
we're
actually
going
to
do
now
is
we're
just
going
to
door.
I
start
I
have
to
docker
ways
to
be
authorized.
This
I'm
not
going
to
show
your
daughter
eyes
in
it,
because
that
would
be
kind
of
boring.
We
use
I'm
using
the
Nvidia
route
rather
than
the
intensive
floor
route.
You
can
call.
A
A
So
we
can
quickly
show
up
what
so
once
we've
depth
or
eyes
that
it's,
then
what
we
can
do
is
we
need
to
run
up
docker
with
the
Nvidia
runtime.
So
you'll
have
to
firstly
install
that
in
video
runtime
for
docker
to
get
it
to
run
and
once
you've
done
that,
then
you
can
run
it
up.
That's
just
an
example
run
there
for
running
there,
NVDA
SMI,
so
you
can
check
that
it's
working
and
I
also
would
recommend
and
you'll
need
this
for
mini
coop.
To
put
that
issue
default.
A
A
Okay,
so
once
we're
also
going
to
mount
a
local
directory,
because
we
want
to
read
we're
going
to
read
our
data
from
disk
and
I'll
explain
a
little
bit
more
about
that.
How
we're
going
to
run
that
as
we
go
so
we'll
build
on
image
and
then
we'll
run
it
so
we'll
run
it
quickly,
I'll
see
if
we
can
get
it
run,
doc
and
now
and
there's
just
the
reason
to
just
do.
This
is
just
make
sure
you've
got
everything
set
up
beforehand.
A
A
A
Okay,
so
it's
running
there
within
the
docker
container
right,
so
we've
dock
rights.
There
we've
got
that
running,
so
that's
cool
move
that
one
out
the
way.
Well,
that's
one,
and
that's
like
that,
one,
true,
okay,
so
the
next
thing
we
kind
of
want
to
do
with
this
is
we
want
to
run
it
and
we
want
to
run
it
in
mini
chip.
A
There's
KVM
to
driver!
You
come
up
with
KVM
to
driver,
you
need
to
have
I,
think
I
Oh
you
enabled
on
wherever
you're
running
it.
So
for
that
to
work
and,
however,
you
can
I
think
can
dedicate
more
of
the
GPU
to
it.
So
you
can
have
less
memory
issues
around
the
GPU,
if
you're
going
to
do
that,
but
what's
actually
a
lot
simpler
is
is
to
run
it
with
VM
driver
of
none
and
then
I'll
basically
fall
through
the
underlying
talker
and
thing.
A
A
Well,
here
we
go
straightaway
so
there's
a
couple
of
things
here:
we've
got
a
persistent
volume,
we'll
have
a
persistent
volume
of
business
and
volume
claim
for
the
data,
so
we're
gonna
bring
the
data
and
just
from
desk
effectively,
as
we
had
done
before,
and
there's
some
pros
and
cons
of
doing
that
which
we'll
talk
about
in
a
minute
and
then
we
are
going
to
run
through
a
job.
So
but
I
didn't
write
up
just
a
deployment.
We're
gonna
run
a
Cuban
Eddie's
job
because
we
want
to
run
this.
It's
gonna
run
and
exercise.
A
So
we
want
it
to
run
and
finish.
Try
it's
not
going
to
continually
run
forever
a
couple
of
things
in
there
that
we
have
to
have
in
there.
We
have
to
have
resource
limits,
GPU,
one
to
say
that
we
need
to
use
the
GPU,
otherwise
it
won't
schedule.
Potentially
what
schedule
and
support
the
GPU
enabled
couple
of
things
there
and
mini
cube.
The
resource
limits
don't
seem
to
get.
Actually
they
don't
seem
to
work,
particularly
well
for
GPU.
A
So
if
you
run
up
more
than
one
still,
you'll
probably
have
issues
because
Charis,
why
actually
and
intensively
will
actually
just
try
and
use
all
of
it?
You
could
potentially
cut
down,
say,
say
only
use
half
the
GPU
for
caris,
but
probably
not
really
a
huge
amount
going
and
doing
that.
So
you
have
a
look
so
well
gone.
Have
a
quick
demo
of
that
one
as
well.
So
let
me
just
get
to
the
right
directory.
A
A
I
look
over
here,
okay,
so
we're
gonna
use
our
mini
cube
context.
My
mini
Cuba
at
the
moment
has
some
issues
with
its
certificates,
so
we
are
gonna
have
to
run
it
insecure,
which
is
kind
of
no
ideal.
But
this
morning,
when
I
reinstalled
everything
it
seemed
to
lose
this
certificate.
Somehow
and
I
wasn't
able
to
fix
in
time,
for
the
demo
should
be
okay
for
the
demo.
So
we'll
just
have
a
look,
see
things
then
running
just
dull
check,
I
didn't
leave
anything.
A
So
we've
got
nothing
running
from
the
jobs
right.
So
if
I
have
a
look,
I
have
a
jobs
m'kay
file
here,
which
is
they're
all
job
spec,
full
so
on
again,
I'm
going
to
share
all
this,
so
you're
welcome
to
think
you
can
be
able
to
download
afterwards.
So
this
has
got
all
the
stuff
in
here.
It's
got
and
labels
it
Scott's
alien
video.
We
also
have
a
bunch
of
environment
variables
which
I'll
talk
about
a
little
bit
and
we
have
a
bunch
of
other
stuff
in
the
bottom.
They've
got
the
volume
claim
for
these.
A
A
A
A
Ok,
so
you
can
see
it
running
in
the
ice.
Well,
so
that's
cool,
so
we've
now
got
two
containerized.
We've
got
a
job
definition
for
it
and
we've
got
it
all
up
and
running.
So
that's
great.
So
what
we
want
to
do
now
is
move
it
to
the
cloud.
So
let's
just
talk
about
that
quickly,
so
we
can
just
move
this
drop
our
container
onto
the
kubernetes
cluster
and
run
it
now.
Well,
we
kind
of
can
and
can't
right
so
there's
a
few
things.
That's
going
to
cause
some
intentional
few
issues
here.
A
Firstly,
all
of
the
previous
demos
they're
loading
from
the
local
disk,
so
we're
gonna
have
to
mount
the
disk.
Somehow,
if
we're
going
to
use
this,
but
we
may
not
want
to
use
disk,
we
send
her
in
to
stand
out
sounded
out
well
I
poop,
which
is
kind
of
fine,
not
too
unhappy
with
that
and
the
graphs
that
we
would
display
locally.
Often
we
use
in
PI
plot,
for
which
I
know.
Obviously
we
can't
use
everyone
in
a
cloud
remote.
So
so
we
need
to
want
to
make
a
little
bit
more
cloud.
A
Native
and
it'd
be
nice
to
be
able
to
spell
me
right
in
cloud
to
be
able
to
spin
up
multiple
things
that
we
could
do
some
high
parameter
tuning
and
that
kind
of
stuff.
Okay,
so
we
can
get
our
data
I
had
a
look
at
various
ways
of
bringing
data
in.
We
can
use
object,
stores,
databases
this
for
this
demo,
we're
just
going
to
use
disk
and
what
we're
going
to
actually
have
a
master
or
Orchestrator.
A
So
we
are
going
to
do
a
persistent
volume.
There's
going
to
be
one
portal
effectively
read
that
data
from
a
remote
location
and
store
it,
and
then
the
other
Pods
will
have
a
read-only
claim
on
that.
So
they
can
read
it
and
run
it,
so
the
other
option
would
be
to
and
other
other
kind
of
options
you
could
do
is
peel
the
data.
Every
time
you
run
up,
that's
probably
not
the
best
of
ideas.
It's
probably
going
to
be
slow.
A
You
know
every
time
you
start
up,
you're,
actually
looking
download
the
data
is
probably
more
and
you
could
pull
it
from
s3
or
something
like
that.
Obviously,
if
you're
going
to
distribute
it
and
do
just
more
distributed
machine
learning
and
becomes
kind
of
a
different
game-
and
you
want
it
probably,
then
you
probably
going
to
read
from
different
places
again
and
maybe
more
distributed
file
system
or
distributed
and
cluster
systems.
A
A
So
this
one
here
we're
just
going
to
go
through
it's
going
to
do
a
bunch
of
stuff
this,
so
this
nice
week
or
three,
so
it's
gonna
run
through
and
we're
just
going
to
download
the
check
to
see
if
the
files
exist
then
download
it
right.
So
it's
really
simple.
So
we
run
that
we're
going
to
run
that
out
and
say
a
pod
on
our
and
it's
on
pod.
So.
A
We'll
show
this
in
a
minute,
I
think
because
the
times
a
bit
tight
okay,
so
we
basically
deploy
that
just
says,
deployed
acid
deployment
because
it's
going
to
be
continually
running,
it
just
checks
every
time.
You
start
something.
It
checks
to
see
whether
I
need
to
download
the
data,
will
download
the
data
and
then
store
it
into
the
storage,
and
the
reason
it
can
see
only
runs
is
also
running
a
web
server
on
there
and
because
what
we
also
want
to
do
with
that
or
trades
me
want
to
get
our
graph
recite
from.
A
So
what
we're
going
to
use
carrots
callbacks?
So
we
need
to
have
a
endpoint,
so
I
cannot
go
out
so
going
to
have
an
endpoint
to
take
those
callbacks
and
then
it's
going
to
store
those
locally
and
then
display
the
graphs.
So
you're
gonna
see
this
like
that.
So,
let's
see,
if
we
can
get
this,
what
we'll
show
you
I
think
I'll
show
you
that
in
a
minute
we'll
do
one
more
thing.
This
is
one
more
thing.
That's
going
to
do.
A
A
A
A
Yeah,
look
at
that,
so
what
I
run
in
there,
and
so
if
we
go
and
have
a
look,
I
think
I'm
already
running
the
orchestrator
on
there.
So
if
we
do
a
control
god,
you
should
be
able
to
see
that
we've
got
that
running.
Hopefully
yep
cool,
so
that's
running,
and
so,
if
we
now
go
to
get
another
browser
window
up,
we
can
have
a
look
to
see
what
that
looks
like
so
just
escape.
A
A
A
A
And
you
can
see
it's
just
gonna
pull
what
the
only
difference
is.
It's
gonna
pull
the
image
from
there
rather
than
the
local
one
and
it's
gonna
always
pull
yeah,
because
we
always
want
to
pull
up
when
we've
got
something
again,
there's
some
missing
pieces
in
there.
Okay,
so
let's
go
back
to
our
terminal
and
see
if
we
can
see
if
that's
running.
A
A
A
A
Okay,
so
the
last
thing
we
want
to
do
is
we
want
to
chain
some
hyper
parameters
so,
instead
of
using
the
batch,
the
actual
batch,
the
amal
here,
we're
actually
going
to
codify
this.
I've
got
this
and
some
go
codes
and
we
can
pass
in
and
we've
just
got
a
little
api
endpoint
there.
We
can
sense
and
some
stuff
too,
and
get
it
to
run.
So
it's
going
to
run
up
a
bunch
of
jobs.
So
I'll
show
you
that
quickly.
Well,
that
looks
like
and.
A
A
A
A
A
Let
me
all
data
refresh
this
you'll
see
there's
a
couple
of
things
here,
so
we're
starting
to
get
our
second
trace
on
there.
So
we're
the
other
model.
That's
running.
The
other
thing
here
is:
we've
got
their
weights
from
the
previous
model
that
we've
trained,
so
we
can
download
that
now
and
we
can
run
it
as
a
test
okay.
So
this
is
going
to
take
a
little
bit
to
run
because
I
didn't
spend
like
another
node,
which
is
you
know
before
we
started
that
will
let
it
run
for
a
little
bit.
A
A
A
A
A
A
A
A
A
Okay,
all
right,
so
we
do
as
I
say
we
can
just
post
that
JSON
and
I
see
Mayan
script.
To
do
that
and
that's
ran
out.
So
you
can
see
that's
all
running,
so
we
run
the
multiple
versions
we
say
we
did.
There
was
a
little
few
tricks.
I
had
to
do
when
I
was
running
this
locally
in
mini
cube
to
stop
over
and
consuming
the
GPUs.
A
So
we
got
a
results,
though,
and
we've
got
a
weight
stored
on
disk.
So
it's
pretty
much
the
end,
because
I
think
I've
got
about
three
minutes
left
so
couple
of
things
we
can
do
that
which
we
didn't
really
show
which
is
simple
enough
to
do.
We
can
actually
increase
the
data
size
now
we're
running
on
Clyde
pretty
easily.
A
We
could
also
paralyze
the
attention
for
jobs
if
we
wanted,
so
we
could
actually
run
it
in
parallel,
so
we
could
have
one
big
job
that
would
run
in
parallel.
The
good
question
is:
is
it
worth
containerized?
And
someone
was
talking
to
me
about
this,
and
they
said:
can
we
just
run
all
this
in
bare
metal
and
it
kinda
depends
on
the
workload
you're
running
and
I
think
in
some
cases,
if
it's,
if
there's
plenty
of
other
ways
to
do
machine
learning,
there's
obviously
Clyde
machine
learning
tools
that
are
there.
A
This
is
kind
of
a
nice
way
of
things
to
do
it.
Just
if
you
want
to
put
in
a
CICE
pipeline,
you
don't
have
massive
jobs.
You
can
get
the
downloads
from
my
ipod,
my
github
I
think
it
needs
updated,
so
I
updated
this
afternoon,
so
I
think
I
got
about
two
minutes
left.
So
if
you
any
questions
is
probably
not
time
knife,
but
you
can
crypt
grab
me
afterwards
or
feel
free
to
combine
the
pivotal
booth
tomorrow
morning
from
which
I'll
be
there.
Thank
you
very
much.