►
From YouTube: OpenShift Commons ML Briefing Operationalizing Object Detection on OpenShiftwith Will Benton Red Hat
Description
Will Benton (Red Hat OpenShift) demoing how to Operationalize Image Detection on OpenShift with Tensorflow to the Machine Learning on OpenShift SIG of OpenShift Commons
A
A
You
want
to.
You,
know,
pass
this
image
to
a
program
and
tell
you
what
objects
are
in
it.
This
is
a
sort
of
important
problem
in
actually
processing
images
at
scale
today,
and
it's
one
that
Intel
recently
had
been
out
of
out
of
scope
for
computational
methods
and
something
you'd
have
to
sort
of
devote
a
person
to
the
thing
I'm
going
to
talk
about
today
is
how
I
took
an
off-the-shelf
model
for
object.
Detection
ran
it
in
tensorflow
and
operationalized
it
on
OpenShift.
A
So
the
end
result
of
what
we
have
here
is
that
we
want
to
be
able
to
take
one
of
those
images,
pass
it
to
a
service
and
get
back
a
list
of
all
the
objects
that
we
think
are
in
that
image.
So,
if
I
want
to
pass
it
to
this
web
service
here
that
I
have
running
in
openshift
on
I
have
three
components
here:
I
have
a
script
that
encodes
an
image
as
a
JSON
payload
that
I
can
pass
to
a
rest
service.
A
I
have
just
in
connecting
to
that
rest
service
and
posting
that
payload
and
then
we'll
just
see
what
the
output
for
that
looks
like
for
this
kid
and
Terrier
picture.
A
So
if
we
actually
want
to
see
where
these
things
are
getting
identified,
I
have
another
script
that
will
take
this
and
put
bounding
boxes
around
the
objects
that
it
found
and
it'll
render
that
as
a
new
image
which
we
can
look
at
and
I
guess,
I
see
how
you
might
think
that
these
are
these
kids
shoes
or
our
snowboards.
But
so
we
have.
We
have
a
kid
and
a
dog,
then,
whatever
this
bit
of
hand
is
identified
it
as
well.
So
we
can
repeat
this
with
these.
A
With
these
other
images,
we
looked
at
and
I'll
just
use
a
little
script
to
make
it
easier-
and
this
is
just
sort
of
a
shell
function
that
I
made
so
that
I
don't
have
to
type
as
many
things
and
mess
them
up
here.
But
if
I
want
to
look
at
the
object
classes
in
this
image
of
Katie
Compton
I
see
if
there's
a
bicycle
and
a
person
well
very,
very
good.
We
we
should
hope,
there's
a
bicycle
and
a
person
in
there
and
then
I'll
annotate.
That
image.
A
And
see
that
these
bounding
boxes
are
a
little
harder
to
see
because
the
image
is
is
larger,
but
we
can
see
that
there's
a
bounding
box
around
miss
Compton,
a
bounding
box
around
her
bicycle
and
actually
it
I
dent
affine
you,
this
sort
of
blurry
people
in
the
stands
as
well,
we're
spectating
the
race
and
then.
Finally,
let's
look
at
this
for
the
macaw.
A
So
we
have
actually
some
spurious
results
here.
It
would
be
interesting
to
see
what
sort
of
confidence
levels
we
had
for
those,
but
we
can
annotate
that
image
with
bounding
boxes
and
then
see
what
that
looks
like
as
well,
and
we
see
that
we've
identified
the
bird,
but
we've
also
identified
some
parts
of
the
bird
as
potentially
other
things.
So
I
don't
want
to
talk
too
much
about
how
the
actual
technique
Yolo.
A
A
I
want
to
operationalize
this
in
openshift
I
can
just
create
an
application
template
which
is
what
I'm
running
this
as
and
then
deploy
it
to
OpenShift
very
easily
and
I
have
an
automatic,
build
setup
for
this
image.
I
can
rerun
it
as
well,
so
some
of
the
so
I
think
I
think
the
main
interesting
thing
about
this
is,
as
you
saw
it's
it's
not
lightning
fast,
but
it's
pretty
fast
by
web
service
standards
and
the
interesting
thing
about
operationalizing
this
was
that
I
used
a
CPU
optimized
to
build
the
tensorflow.
A
So
a
lot
of
times
you
think
about
using
GPUs,
for
doing
machine
learning
and
and
certainly
for
scaling
up.
You
can
use
GPUs
and
in
more
recent
versions
of
openshift,
you
can
schedule
jobs
on
GPUs
as
well,
but
you
don't
always
have
a
GPU.
My
personal
laptop
doesn't
have
a
GPU
that
I
can
run
tensorflow
on
pretty
painful,
and
sometimes
you
want
to
get
as
much
performance
as
the
CPU.
A
So,
even
with
all
the
serialization,
that's
sort
of
an
acceptable
latency
for
a
occasional
use
service.
So
in
conclusion,
these
resources
will
be
online.
I
have
a
link
to
a
talk
explaining
Yolo,
as
well
as
to
the
author's
website,
with
more
implementations
and
details
of
refinements.
Here's,
the
link
of
the
dark
flow
sensor
flow
implementation
of
Yolo
and
the
dark
net,
a
system
that
I
use
and
I
have
a
link
to
my
github
repository
with
a
an
open
shift
template.