►
From YouTube: ONNX Edge WG meeting 20190612
Description
Recording of the ONNX edge meeting on June 12, 2019
B
B
C
C
C
B
C
Yeah
I
can
scroll
down
good,
so
they're
just
some
changes.
The
first
one
I
sing
yeah
the
definition
of
an
edge
device.
So
in
the
last
meeting
we
talked
about
the
connectivity
of
an
edge
device.
Actually
we
agreed
that
connectivity
is
state
rather
than
capability
for
for
an
edge
device.
It
can
be
connected
or
can
be
not
connected.
It's
a
state,
so
I
changed
my
statement
of
the
definition
of
edge
devices
so.
A
C
A
C
Great,
so
one
change
is
that
the
definition
of
an
edge
device,
so
we
talked
about
the
connectivity
as
a
state
in
the
last
meeting,
so
everybody
agrees
that
connectivity
is
a
state
rather
than
and
capability.
So
there
are
two
states
for
an
edge
device
connected
or
not
connected,
so
I
elaborate
a
little
bit,
give
some
examples
of
the
connected
device
and
are
not
connected
device,
so
that's
fulfills
the
definition
of
an
edge
demise.
So
this
is
a
first
first
major
changes.
Oh
yeah,
an
idea,
I.
C
I
think
because
each
device
is
the
difference
between
cloud
computing
as
computing,
that
at
computer
is
closer
to
the
source
of
the
data,
so
the
work
flow
is
done
within
the
H
device,
but
at
last
you
must
retrieve
the
data.
You
must
deal
with
the
data
flow,
so
even
if
it
cannot
connect
to
the
internet,
it
still
has
some
other
ways
of
connectivity
and
may
be
produce,
maybe
MB
IOT
or
maybe
through
USB
port
portal
or
others,
so
that
that's,
why
I
think
every
edge
device
at
least
has
some
ways
of
connectivity.
A
D
C
A
A
C
C
B
I
can
state
my
thinking
that
it
affects
the
use
cases
in
case
there
is
constant
connectivity.
You
may
have
scenario
of
hybrid
processing
that
the
edge
device
is
processing.
Some
part
of
the
neural
network
and
the
rest
of
it
is
being
processed
in
the
cloud.
Well
says
a
scenario
that
there
is
no
connectivity
and
then
the
entire
network
is
possessed
on
the
edge
device.
For
example,
let's
say
that
we
look
at
the
scenario
fame
not
to
natural
language
processing.
Okay,
all
of
the
ASL
use
cases,
so.
B
Today,
all
of
them
all
happening
in
the
cloud.
Sorry,
alright.
Now
they
are
moving
into
a
mixed
scenario
that
the
device
would
be
doing
some
of
the
voice,
a
pause,
a
sink,
and
it
will
still
not
do
the
complete
enemy
they.
The
rest
of
it
would
be
done
in
the
cloud,
but
you
can
think
of
a
case
where
the
device
is
completely
disconnected
and
then
someone
sting
wants
is
Google
assistant
to
be
able
to
look
into
his
calendar.
D
B
C
D
Know
my
current
good:
do
you
think
we
need
to
elaborate
on
such
scenarios
a
little
bit
more?
So
you
know
the
folks
who
are
reading
it
can
understand
where
we're
getting
you
know
what
we
want
to
get
to
close
with
this
I
could
be
scenario
that
you
described.
The
hybrid
infant
scenario
is
quite
interesting,
I
would
say
some
others,
I
would
think
of
this.
D
Maybe
we
are
just
encrypting
on
the
edge
and
the
input
tensor,
and
then
we,
you
know
the
graph
that
we
have
the
modify
we
have
is
just
going
to
do
some
basic
pre-processing
and
encryption,
and
then
it's
an
address
traffic
to
the
cloud.
Maybe
we
want
to
describe
that,
but
that
such
things
would
have
require
some
more
arcs
to
be
defined.
So
those
are
things
ready.
We
would
like
to
bring
it
to
be
for
a
ever
making
sense.
B
B
B
C
D
B
A
B
So
I
think
this
document
is
trying
to
be
generic
in
describing
what
each
profile
describes
and
not
trying
to
point
down
each
pole
file.
We
can
have
kind
of
a
separated
document
for
each
profile
and
in
that
profile
we
can
community
say
which
scenarios
or
use
cases
are
clickable
to
this
profile
and
so
on.
But
I
need
to
agree
that
we
need
to
add
some
clarification,
maybe
under
the
device,
maybe
under
the
profile.
That
explains
how
connectivity
is
affecting
the
potential
use
cases.
D
D
B
We
may
be
talking
about
profiles,
not
just
one
profile,
but
if
you
profiles
like
it
kind
of
the
way
that
this
is
described,
it
kind
of
implies
that
there
is
a
single
edge
profile
and
what
we
may
want
to
do
is
to
say
there
is
an
edge
device
that
was
defined,
there's
a
family
of
edge
devices
that
we
defined
in
the
previous
section,
and
now
we
want
to
within
that
larger
family
of
edge
devices.
We
want
to
cut
it
into
different
profiles
based
on
the
finger
tardis
card
here.
B
B
C
C
Right
now,
this
document
is
talking
about
onyx
edge
profile
because,
in
my
opinion,
from
the
onyx
point
of
view,
the
scenes
this
community
can
do
a
are
three
things.
Some
quantized
opieop
opiez
operator
sets
some
smaller,
smaller
smaller
size
of
the
models
and
some
edge
runtimes.
This
is
these
are
three
six
eyes
think
about
that.
The
onyx
community
can
do
so.
C
This
document
I
wrote
this
document
from
the
onyx
point
of
is
not
from
the
edge
device
point
over
because
if
we
come
from
the
start
from
the
edge
point
of
view,
edge
device
point
of
view
from
the
edge
Computing's
view,
we
are
not
talking
about
those
operators
subsets.
We
are
not
talking
about
these
model,
size,
runtimes
accuracy
or
anything,
because
if
we
ask
we
start
from
the
edge
devices
we
can-
and
we
can
only
start
from
the
different
scenarios
based
on
a
different
scenario,
elaborates
or
those
different
attributes
or
characteristics,
but
from
the
onyx
eyes.
B
Think
one
of
the
things
that
we
described
I
think
it
also
appears
in
the
statement
of
work-
is
that
we
want
to
talk
about
at
some
point.
We
want
to
talk
about
compliance
yeah,
so
compliance
implies
that
there
is
a
set
of
models
that
you
need
to.
Let's
say
that
we
don't
want,
and
we
kind
of
said
we
don't
want
to
hand
them
performance
performance
is
for
other
organizations
to
handle
at
this
point.
B
Now
that
means
we
have
a
set
of
models
that
we
want
to
test
the
edge
device
on
and
I
think
this
is
well
different.
Profiles
may
become
applicable
because
the
set
of
models
that
I
want
to
test
on
one
type
of
edge
device
may
not
be
the
same
set
of
models
that
I
would
like
to
test
on
another
type
of
edge
device.
B
B
C
C
C
C
So
there
are
some
the
last
major
changes
is
a
detailed
locality
as
well
as
a
power
consumption
cuz
in
last
meeting
I
said
the
hallways
membership
wheezy
mff
was
yes
suspended
for
now,
so
I
cannot
be
the
recruited
him
to
cover
it
with
a
ml,
perv
or
other
bench,
benchmarking
community,
so
I
changed.
My
statement
is
a
power.
Consumptions
I
provided
several
possible
ways
to
lower
down
the
power
consumption
to
optimize.
That,
and
the
last
part
is
data.
Locality
I
talked
about
both
the
data
locality
and
the
some
security.
C
So
yes
just
so
so,
depending
on
the
applications.
So
it
is
up
to
the
developer
just
to
select
the
best
transmission
knology,
whether
through
cellular
technology,
maybe
through
internet
or
or
Bluetooth
or
any
other
transmission
technology.
So
it's
up
to
the
developer,
so
it
is
still
possible
to
create
a
local
network
and
to
send
the
data
to
the
cloud
or
further
processing
or
being
stories
so.
C
Through
these
communications
attacks
can
involves
and
like
something
like:
sensors
collect
those
privacy
data
from
users
or
likes
involving
in
some
auto
pilot
cars
or
some
healthcare
devices,
or
some
literary
every
electronic
items
that
would
potentially
have
an
access
to
the
network,
so
I
think
of
edge
devices
should
be
are
equipped
with
some
cryptographic,
some
accelerators
for
optimizing
power,
consumption
and
the
latency
and
the
security
of
the
data
transmission,
and
this
hardware's
can
also
be
used
for
blocking
some
basic
IP
attacks,
encrypting
the
data
for
a
better
security.
So
that's
a
last
part.
C
B
C
D
B
A
B
C
C
That's
what
I
think.
So
that's
what
what
I
didn't
add
the
compute
capability
here,
but
if
we
want
to
add
this
section
by
C
or
C,
so
in
my
device
we
can
process
like
ten
thousand
pictures
per
second
or
in
your
device.
You
can
only
process
one
so
in
pictures
per
second,
if
we
want
to
do
a
comparison
between
these
two
devices.
C
Yes,
we
can
add
this
attribute,
but
in
my
opinion,
because
we
already
have
latency
speed
and
the
accuracy
to
do
to
describe
because
you
can
see
in
my
devices
to
process
this
kind
of
this
amount
of
data.
I
only
need
one
second,
but
in
your
T
value
you
need
ten
seconds
so
that
that
is
another
way
of
comparison
between
these
two
devices.
B
B
B
B
When
you
go
up
and
you
look
at
the
power
fine,
it's
not
obvious
in
a
sense
that,
if
you
combine
the
attributes
together
generate
a
specific
model
in
a
specific
influences
per
second
power
consumption
accuracy.
This
is
basically
characterizes
the
modem
the
device
compute
capabilities
in
a
sense.
Yes,.