►
From YouTube: ONNX Edge WG meeting 20190605
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
B
A
A
Okay,
so
I
know
that
I
just
saw
email
from
offer.
He
he
told
me
that,
unfortunately,
he
has
some
things
to
attend
for
for
his
daughters.
This
morning
is
cool,
so
I
left
home.
You
won't
be
able
to
join
okay.
Well,
it's
okay,
I!
Guess
we
can.
We
have
I
guess
relative
from
from
each
company
nope
the
three
active
companies,
I
guess
so
that's
good!
So.
A
A
C
A
A
A
A
Okay,
so
you
could
you
don't,
do
you
mind
like
walking
us
through
it
and
how
it
was
you're
taking,
and
then
we
can
maybe
spend
some
time
as
I
said,
I
didn't
have
chance
to
read
it
through
and
obviously
we
all
can't
comment
directly
to
the
full
request
and
that
is
going
to
be
default
action
after
afterwards
to
do
that.
Okay,.
B
Okay,
so
I
was
super
busy
in
the
last
few
weeks.
So
sorry
for
the
delay
and
yeah,
this
talk
is
not
a
finalized
version.
I
see
we
our
deadline
is
by
the
end
of
q2.
S
always
do
have
I,
guess
one
more
meeting,
so
I
guess
we
have
to.
We
must
finalize
this
in
the
next
meeting
and
we
can
this
a
Maitre
D'
atsc,
so
we
still
has
some
times.
A
B
Problem
no
problem
yeah,
okay,
so
so
they
star
consists
of
mainly
two
parts:
one
is
a
scope
of
edge
and
one
is
the
profiling
of
the
edge
devices,
so
actually
I
added
some
scenarios,
because
I
scenario
helped
us
to
understand
the
scope
of
edge.
If
you
only
talks
about
scope
of
edge,
sometimes
people
don't
have
they're
a
very
concrete
image
of
what
the
edge
devices
is.
What
is
this
crimp
of
how
to
interact
with
the
real
world
or
okay?
What
does
he
represent
in
a
whole
network?
So
the
first
part
is
a
scope
of
X.
B
So
by
definition
we
have
to
define
what
is
edge
computing
so
as
computing,
so
is
down
at
or
near
the
source
of
the
data.
So
where
the
data
is
generated,
we
collect
analyze
and
process
the
data
at
the
exact
site
and
then
respond
to
the
response,
generate
the
output
and
respond
to
the
real
world.
So
that
is
the
edge
computing,
so
as
computing
is
large
or
completely
performed
distributed
edge
devices.
B
So
the
target
of
edge
computing
is
any
application
or
general
functionality
needing
to
be
closer
to
the
source
of
the
action
or
source
of
the
data
where
it
generated
and
rail
system
can
process
and
interact
with
the
real
world
instantly.
So
that
is
the
information
and
the
edge.
So
the
scope
of
edge
is
by
definition,
is
any
systems
or
devices
that
can
process
collect
and
process
the
data,
the
exact
site
so
I
think
that
is
it
all.
The
devices
that
can
perform
such
functionality
is
within
the
scope
of
H.
A
C
A
B
A
Yeah
so
yeah,
you
could
say
fork
right.
This
is
the
the
edge
of
the
of
the
of
the
cloud
right
access
edge
of
the
clock.
Basically,
maybe
that
that
is
maybe
we
could
use
that
term
fork
right,
but
yes,
it's
edge
of
the
cloud,
so
maybe
that,
maybe
should
be
spelled
spelled
out
kind
of
directly,
but
but
it's
okay.
Let's
keep.
B
Keep
going
did
this
make
sense,
I
mean
okay,
so
the
second
part
is
talking
about
some
edge
scenarios,
so
I
get
the
inspiration
from
the
COCOMs
slice
in
the
last
workshop,
so
the
computation
for
real-time
applications
such
as
a
facial
recognition,
smart
city,
intelligence,
traffic
control.
These
are
or
edge
scenario,
but
they
are
still
some
differences
between
each
scenarios,
like
some
scenarios
request
a
super-fast
interaction
and
very
very
low
latency
like
auto
pilots
and
other
scenarios
requires
very
high
accuracy.
B
Z
can
tolerate
some
high
power
consumption,
but
they
need
very
accurate
response.
Like
facial
recognition,
you
don't
want
to
recognize
a
run
person,
so
here
we
just
discussed
some
scenarios
or
which
scenarios
is
included
in
the
edge
scenario
and
which
will
be
except
I,
didn't
talk
about
which
will
not
be
included.
B
So
I
only
talked
about
which
what
scenarios
are
included
in
their
scenarios
and
some
examples
of
all
those
included
scenarios
I
mentioned
about
unlike
gaming,
so
game
servers
running
in
the
cloud
and
the
rendered
video
will
be
transferred
to
the
lightweight
clients,
like
some
PR
devices,
like
a
mobile
phone
like
VR
glasses.
So
this
part
is
something
I.
B
Think
should
also
be
included
in
the
edge
computing,
although
there
are
still
some
debates
whether
a
VR
classes
should
be
considered
mobile
as
devices
or
not
I.
Think
because
the
VR
glasses
generate
and
the
process
the
data
and
generate
output
right
at
the
side,
so
by
definition
is
still
and
at
scenario.
So
that
is
my
yeah
yeah.
That.
A
Makes
sense
I
mean
it
is
it
there
is
a.
There
is
a
processing
that
happens
on
device
on
the
sides
which
you
say
so
yeah
I
agree
I
mean
and
some
of
it,
some
of
it.
Actually,
this
might
be
hybrid
scenario
where,
yes,
yes,
exactly
that,
where,
where
you
have
part
of
processing
done
on
device
on-site
and
some
part
of
it
done
in
on
fork,
right,
yeah
yeah,
exactly
so
with
you
know,
assuming
that
there
is
always
you
see,
link
in
between.
A
Think
the
scope
goal
for
this
document
is
really
to
define,
attributes
and
definition
of
what
is
the
edge
profile
right.
Yes,
yes,
OOP
is
won
by
the
edge
profile
is
like
what
defines
the
edge
profile.
What
are
the
attributes
that
define
edge
profile
so
from
that
perspective,
I'm
just
reading
looking
at
the
title
of
the
document,
so
that
might
be
a
little
bit
confusing
because
it
says
profile
link
which
is
which,
which
is
different.
So
maybe
we
should
rename
that,
but.
C
A
But
anyway,
the
profile
the
finishes,
define
the
edge
profile,
attributes
right
and
what
the
course
is
used
at
right
and
and
then
you
know
the
the
the
device
like
a
phone
or
a
camera
disconnected
connected
camera
in
or
the
infrastructure
edge.
Those
are
two
different
profiles.
I
would
I
would
think
about
them
as
two
different
profiles
of
the
edge
profile
right
yeah.
A
A
We
are
profile
which
is
maybe
maybe
will
be
in
category
of
connected
or
hybrid
hybrid
profile,
which
which
involves
the
on
device
processing,
end
on
the
edge
processing,
and
then
there
will
be
some
pure
profiles,
like
you
know,
like
camera
profile,
or
things
like
that
right.
This
is
where
I
am
when
I'm
trying
to
go
right.
So
if
that
makes
sense,
so
I
think
both
of
them
will
be.
C
I
had
a
comment
here
on
this:
how
do
we,
where
do
we
draw
the
line
like
let's
say
we
have
a
heavy
edge
kind
of
a
device
and
would
it
its
peripherals.
C
C
C
A
A
So
it's
not
it's.
Not
a
data
server
is
not
not
cloud
computing.
So
if
you
define
that
as
a
scope
of
the
edge
and
and
it's
like-
you
know,
that
means
it
could
be
fog
plus
4
game
below
you
know,
then
then
you
know
PC
in
that
context.
Yes,
it
is.
It
is
a
profile
that
is
maybe
it
looks
like
a
composite
profile.
B
A
C
A
Just
my
thinking
of
the
overhead
I,
don't
know
I
mean
yeah,
that's
that's
what
I
think
right
now
yeah,
it
makes
sense.
Actually
you
know
it
could
be
composite
profiles
right.
You
know
it
could
be.
You
know,
device
that
has
multiple
if
it
so
happens
that
we
have
separate
profile
for
camera.
That
is
a
smart
camera,
for
example,
or
video
processing
device
right
profile,
I
mean,
obviously
that
is
a
part
of
the
PC
right
yeah
and
there
is
a
audio
processing
or
a
smart
speaker
thing.
A
A
A
Edge
scenario
it
could
be,
we
could
cover,
have
a
paragraph
I
think
it
covering
the
possibility
of
oh
well.
They
even
let
me
say
this
I
mean
currently
even
the
exact
fairest
scenarios
like
you
could
have
that
PC
to
this
or
laptop
or
whatever,
but
say,
for
example,
the
car
is,
is
brought
up
as
a
as
a
use
case.
Well,
car
is
for
sure
composite
composite
edge
edge
device
right.
So
maybe
we
should.
We
could
have
a
paragraph
talking
about
that,
the
the
edge.
No
it
does
not
necessarily
you
know
it
could
have.
A
Smaller
scenarios
so
to
speak
right
because
you
know
car,
for
example,
to
have
the
video
pressing
could
have
the
approach
think
we
are
processing.
All
of
that
is
is
part
of
a
car
as
a
scenario
so
I
know
if
you
need
specifically
to
spell
it
out,
maybe
it's
obvious.
Maybe
it's
not
do
you?
Would
your
wife
to
spell
this
out
in
history
explicitly
yeah.
C
B
B
B
A
A
There
could
be
smart
as
devices
that
have
a
local
you
know,
processing
using
using
like
sensors
and
using
the
net.
No,
not
for
processing,
I
mean
very,
very
local,
but
it's
not
necessarily
connected
and
then
only
maybe
it's
something
that
is
deployed
and
then
collected
and
and
and
you
know
some
data
collection
is
done
offline
now.
C
A
A
And
you
know
so
both
both
of
the
statements,
but
then
the
I'm
thinking
of
the
device
that
does
the
no
not
for
processing.
You
know
and
is
not
necessarily
maybe
it
has
clinic
okay,
so
say
it
doesn't
have
connected
connectivity
capability,
Wireless
or
in
network
connectivity.
There
is
no
connectivity.
You
have
to
physically
the
device
to
get
results.
A
For
example,
you
know
it
means
you
have
to
collect
device,
you
have
to
fight
it
in
USB,
port
or
something,
and
so
it's
not
part
of
the
network.
You
know
infrastructure,
you
know
there
is
no
way
to
communicate
through
through
internet
or
other
means
to
the
device
right.
You
have
to
physically
collect
the
device
and
get
the
output
right,
so
maybe
some
spunk
no
sense
or
sprinklered
around.
So
it.
B
But
I
still
think
or
edge
devices
has
ability
to
connected
into
the
service
network,
because
if
the
it's
huge
device
is
not
so
provide
connections
in
secure
internet
search
drivers,
so
your
state
rather
than
capability.
So
it's
just
the
state
that
you
are
right
now
connected
or
not.
It's
not
a
capability
that
you
can
connect
or
not,
because
I
think
all
the
edge
devices
should
have
the
capability
to
connect
early.
B
A
A
Internet
connectivity,
that
or
do
you
assume,
Internet
connectivity
or
you
know
wireless
or
wired
or
whatever,
but
because
what
I'm
trying
to
say
is
connectivity
and
capability
needs
to
be
clear,
clearly
specified
as
a
part
of
edge
device.
I
think
is
it
required
or
not,
and
what
connectivity
capability
means?
Maybe
connectivity
capability
means.
A
You
know
USB
connectivity,
oh
yeah,
I,
see
you
know
it
can
be
something
that
is
not
necessarily
infrastructure
network
network
infrastructure.
So
anyway,
that's
that's
my
thinking.
Maybe
we
should
not
try
to
tie
it
tight
strictly
to
B
to
D
to
the
Internet
network
connectivity
or
maybe
we
should
I
mean
I'm,
not
saying
we
I
mean
this
is
just
the
question
for.
C
D
A
C
A
B
B
Talking
topics
already
discussing
the
last
meeting
about
the
metrics
about
that
multi,
dimension
radar
charge
and
so
actually
right
here,
I'll
this
trade
I,
remember,
I,
remember
six
metrics
I
didn't
include
a
connectivity
or
not,
but
I
will
added
on
this
time.
So
right
now
we
have
accuracy,
size,
speed
and
power
consumption.
There's
one
thing
because
in
the
original
dog
the
importance
of
consumption
I
propose
to
collaborate
with
the
MFF
community.
But
right
now
we
have
some
issue.
The
hall
we
have
some
issue
is
I'm
a
perv
because
I'm
a
pro
feels
more.
B
Like
a
standardization
organization.
We
have
to
comply
with
a
year,
our
so
according
to
to
the
feedback
of
my
colleagues.
If
always,
logo
has
already
been
removed
from
the
membership
list.
So
why
now
we
can
I
cannot
communicate
with
a
mapper
frightening
or
if
we
want
to
the
onyx
edge.
Won't
you
still
collaborate
with
ml
perf.
Maybe
I
cannot
be
the
representative,
so
there
is
one
a
problem
about
this
and
there
are
some
other
metrics
like
security
and
bandwidth.
B
A
C
D
A
A
A
B
Is
a
problem
because
actually
I
myself
tested
the
first
version
of
the
Atlas
series,
sensors,
the
monitor,
which
can
run
the
facial
recognition
program
and
actually
the
first
version
is
terrible,
I
and
all
by
all
of
my
colleagues
tested
that
and
the
accuracy
is
almost
close
to
zero.
It
can
not
tell
who
is,
who
is
it's
completely
run
so
and
so
I
think
for
many
image.
B
Processing
application
like
image
classification,
facial
recognition
is
still
very
important
and
accuracy
is
still
very
important
metrics,
and
there
is
another
concern
that
if
we
move
an
application
front
cloud
to
add,
sometimes
we
have
to
compress
the
model.
We
have
to
change
the
opiez
into
the
lower
bit,
but
we
still
need
to
maintain
somewhat
at
very
decent
level,
something
like
about
90
percent
above
85
percent
level
of
accuracy.
So
there
is
what
we
are
talking
about:
how
to
maintain
this
kind
of
accuracy
in
a
pretty
decent.
B
A
I
was
this
is
tied
to
the
actual
use
case
right
and
whether
the
specific
use
case,
you
know
it
seems
like.
If
you
introduce
accuracy
as
attribute
of
the
profile
of
the
edge
profile,
then
it
seems
like
you
could
have
profiles
that
are
sort
of
high
accuracy,
low
accuracy
I
mean
in
a
sense
like,
but
then,
if
he
and
I.
A
A
May
be
different,
it
is
interesting.
Question
I
think
that
it's
very
much
tied
to
this
specific
use
case,
and
perhaps
when
we
get
to
the
point
of
defining
the
you
know
edge
profiles,
we
will
get
to
the
point
where
we
were.
We
will
we
I
think
we
want
to
for
each
of
the
profiles
we
define.
We
would
want
to
specify
the
sort
of
compliance,
compliance,
suite,
right,
yeah
and
and
and
define
some
soft
characteristic
use
we
use
for
the
compliance
test.
Alright.
A
B
C
D
C
C
C
A
What
we
want
to
have
as
a
proto
profile
right
yeah,
you
know,
but
we
think
someone
through
I
mean
I,
guess
we
have
some
time
until
they
have
another
meeting
next
week.
I'm
not
gonna,
cancel
that
so
I'll
remove
it.
So
we
can
continue
next
week
as
well,
but
you
know
in
the
meantime,
I
obviously
I
need
to
put
some
more
total
I.
Don't
know.
B
A
A
A
A
D
A
B
A
A
A
B
A
B
B
A
So
we
don't
need
to
introduce
any
proof
in
the
context
of
this
document
honestly,
because
why
we
are
defining
the
the
attributes
of
the
profile.
Yeah
yeah
I,
see
I,
think
the
doll
will
demo
first,
so
I
aspire
I'm
concerned
in
this
document.
This
this
whole
paragraph
could
be
removed
completely
no
problem
and
then
for
security.
This
is
my
service.
Well,
we
should
talk
about
the
italo
quality
and
one
thing
that
that
is
I'm,
just
thinking
of
is
it
shouldn't
be
or
not.
The
attribute
that
we
should
consider
consider
is.
A
C
Stuff
describing
or
providing
some
guidance
to
the
runtime
on
how
to
handle
these
also
like
I,
mean
it's
a
good
example
like
either
accuracy
or
power.
Right.
Let's
say
power
is
the
metric
and
if
the
device
says
I
am
of
this
profile
and
it
expresses
that
this
is
my
min
power
or
power
average
power
or
max
power.
Something
like
that,
then
runtime
can
decide
what
implementation
to
pick
like.
Similarly
for
accuracy
also,
it
could
pick.
You
know
this
is
very
low
tolerance,
one
and
I
need
to
have
the
most
accurate
maths
going
on.
A
A
This
is
like
for
us
deciding
these
are
the
dimensions
or
vectors.
We
want
to
use
to
define
the
profiles
right,
so
we
can
basically
classes
or
categories
of
the
edge
devices
right
right.
You
know,
and
so
this
profile
is
right
now
we're
defining
just
the
attributes
that
we
should
use
to
delineate
the
the
to
use
to
define
the
edge
profile.
A
Now,
when
you
just
go
ahead
and
start
defining
edge
profiles,
for
example,
I
think
like,
for
example,
phone
or
a
car
or
or
you
know,
infrastructure
edge
or
you
know,
then
you
will
put
boundaries
on
these
metrics
or
value
to
these
to
these
metrics
or
we
can
call
it.
There
are
metrics
attributes
KPIs
whatever,
and
then
this
is
going
to
be
the
bounding
space
for
the
given
edge
profile.
That's
the
way
how
I
see
make
sense?
A
A
A
C
Thinking
maybe
we
can
express
that
you
know
like,
of
course
speed
and
size
and
power
would,
you
know
basically
reflect
the
compute
capability
to
the
other
aspect
is
just
the
maths
like
you
know
whether
if
something
is
capable
to
run,
you
know
fp6.
You
know
like
64-bit
or
just
end
date,
or
you
know
less
number
of
ops.
If
it's
an
LST,
you
know
audio
DSP
device,
it.
A
C
A
Capabilities
I
think
saying:
I
mean
I
mean
power.
Power
consumption
is
one
thing.
Obviously
it
could
be
I
mean
even
though
you
could
argue
the
power,
consumption
and
compute
capabilities
or
someone
connected,
but
still
computing
capabilities
is
like.
Oh
maybe
it's
device
that
can
only
do
fixed
point
drops.
C
C
So,
okay,
let
me
so
I
think
we're
on
the
same
page
I-I've.
What
I
understand
from
your
definition
of
compute
capability
is
you
know
things
like
if
a
device
can
only
do
integer
mats
or
can
only
do
you
know
five
offs
in
X?
We
want
to
express
that
right.
So
what
I
want
to
say
is
that
that
term
compute
power
is
probably
too
broad
and
white.
You
know
it
can
be
misinterpreted.
As
other
things
you.
C
C
Line
compute
those
new
strains,
correct
I,
am
seeing
that
we
of
all
these
points
here
right
other
than
accuracy.
All
of
them
are
reflected
in
the
document
that
we
were
seeing
earlier
yeah.
So
what
we
are
not
seeing
is
compute
resource
constraint.
Yes,
and
your
definition
from
what
I
understand
of
compute
resource
constraint
is
data
types
and
ops,
right,
yeah,.
A
You
know,
pure,
you
know,
mix,
mix,
capabilities
of
an
overview
of
the
device
right
like
how
many
maps
or
gulps
or
whatever
we
you
know.
Can
it
crunch
right
and
should
that
be
right?
So
there
is
an
aspect
of
the
ops
and
data
types.
There's
aspect
of
pure
compute
power
like
when
I
say:
compute
power
I,
don't
mean
what
I
mean
compute
capabilities.
I
didn't
sounds
like
really
really
how
many
instructions
per
second
can
run
right.
C
A
C
D
A
D
A
A
Late
here
so
I
can
only
work
tomorrow.
You
know
no
I'm,
not
saying
now,
I
mean
next
week.
Next
Wednesday
we
have
another
meet
by
next
week.
You
know
you
also
could
update
some
of
some
of
the
some
earthquake
earthquake
and
then
we
also
will
comment
directly
into
the
the
document
here
in
the
pr
debugger.