►
From YouTube: ONNX Edge WG meeting 20191030
Description
Description
A
B
B
Okay,
so
for
the
agenda,
we
say
that
I
ordered
that
we
want
to
continue
what
we
talked
last
week,
which
was
to
to
review
the
updated
statement
of
work
and
start
talking
about
it,
laughs
that
won't
equate
to
to
describe
the
different
types
of
profile
settings,
and
then
we
could
give
it
a
write
it
to
a
vote.
Oh
I
do
actually
by
the
way
there's
another
thing
that
I
just
thought
about
is
any
of
you
going
to
join
the
the
workshop
of
Android
I.
A
We
need
to
plan
for
the
the
presentation
and
update
in
the
in
the
work
group
on
the
work
group
activities
and
so
I'm.
You
know,
by
default
a
su
dong.
You
will
be
the
one
presenting
them.
Yes,
yeah
I
will
too
our
work.
Our
work
includes
uptake
yeah,
so
we
need
also
maybe
for
the
next
next
meeting,
I
guess
to
to
Natalie
to
plan.
You
know,
review
of
the
slides
or
update
okay.
B
B
B
B
B
Talking
about
collaboration
so
kept
to
the
modern
Zoo
and
Camilla
raised
I
modified
it
to
have
number
three.
So
what
used
to
be
number
three
becomes
four
and
so
on,
and
it
actually
extends
because
now
we're
saying
that
the
definition
of
the
method
of
topology
is
what
we're
targeting
to
the
video.
Now
during
q3
I
will
start
me
late
I
think
they
went
to
change
it
to
q4
I,
make
it
2020
q1,
but
that's
awfully
the
changes
to
the
s
Oh
W,
any
additional
changes
that
people
would
like
to
do.
At
this
point.
I.
B
B
B
B
B
What
are
the
different
ways
to
define
families
of
poor
funds,
and
last
week
we
mentioned
that
it
might
be
a
good
idea
to
put
it
to
a
vote
just
like
they've
Foundation,
walking
up
deed
and
have
the
other
members
vote
on.
What
they
see
is
a
reasonable
way
to
map
the
different
profiles
in
topology
I,
don't
mean
any
kind
of
network
topology
or
something
like
that.
I
just
mean
a
mapping.
B
B
C
Think
I'm
probably
a
little
confused
here,
so
there
are
two
ways
one
is:
there
is
a
compliance
feed
somebody
runs
updates
the
github
and
the
NAP
vendor
can
go
and
get
the
profile
table
lookup
table
and
know
what
are
the
capabilities
of
the
underlying
device
correct
and
the
second
option
that
you
are
is
that
the
runtime
could
query
the
hardware
in
some
way
right,
which
I
run
time.
We're
talking
about
not
just
one
extra
time.
A
Having
said
that,
I
mean,
if
I
mean
as
a
part
of
the
our
I
guess
scope.
They
could
make
recommendations
or
give
options,
but
not
necessarily
you
know,
demand
certain
way
of
doing
it
right.
This
is
what
the
way
I
said
it
is
something
is
we
could
give
put
on
table
and
say
these
are
the
possible
options,
maybe
not
all
of
them,
but
you
know
this
could
be
achieved
this
way
or
that
way
and
then
leave
it
I
think
to
six
I
to
to
to
really
kind
of
decide
what
to
do
right.
C
B
C
So
one
thing
is
the
rent,
runtime
right
and
the
runtime
is
going
to
query,
and
eventually
the
hardware
is
going
to
say:
okay,
the
driver
will
query
the
hardware
and
the
hardware
will
be
able
to
address
the
runtime
that
these
are
the
list
of
hops.
That
I
support
right,
and
the
second
thing
is
just
you
on
the
hardware
itself,
you
it's
a
smart
speaker.
I
can
only
support
this
thing
and
it's
a
black
box
and
it's
a
sticker
that
goes
on
the
hardware
itself.
I,
don't
know
what
that
even
means
was.
B
B
People
who
create
the
hardware
you
know
this
is
can
be
part
of
the
discussions.
How
it's
been
very
marketed.
Is
there
some
kind
of,
as
you
said,
is
there
a
sticker
that
says
onyx
compliant?
But
it's
thing
doesn't
give
you
a
lot
of
details
on
components
to
what
I'd
say
that
you
end
up
with
a
24
files,
so
you're
compliant,
but
to
each
one
of
them.
So
you
probably
need
more
information.
A
A
C
Think
so,
because
I
probably
am
a
little
struggling
on
what
is
the
end
picture
looking
like,
and
maybe
it's
clear
more
to
the
others,
but
having
that
end
picture
in
front
of
us-
and
this
is
these-
are
the
steps
that
we
are
taking
in?
That
direction
would
help
and
I
believe
this
would
be
a
good
clarity.
Then,
if
the
six
are
aligned
that,
then
we
know
that
this
is
the
direction
we
are
headed
to
right.
Also,
I
would
say
that
it
would
help
us
Clara
get
guidance
from
the
product
community.
C
A
A
B
B
B
We
would
have
a
good
definition
of
compliance
testing
for
edge
devices
by
saying
that
some
device
support
some
list
of
offerings
any
past
these
compliance
tests.
To
me,
the
purpose
of
the
working
group
was
around
interoperability
between
different
devices.
To
assured
it
to
the
application,
an
edge
device
can
run
some
models
and
its
own.
You
know
in
a
sense,
it's
all
about
compliance
in
a
way
I'm.
C
With
you
but
offer
I
think
I've
expressed
that
in
the
past
to
that
in
more
scalable
configurable
devices,
which
model
I
am
run,
is
a
very
dynamic
thing.
It's.
It
can
happen
that
I
am
able
to
run
a
model
right
now
and
then
something
else
happens
in
the
platform
and
I'm
no
longer
able
to
run
it.
So
that's
where
probably
I
probably
need
your
help
to
accommodate
how
we
fit
in
devices
which
have
a
lot
of
things
going
on
internal.
Also.
B
So
I
think
the
goal
here
was
to
the
two
to
examine
the
static
IQ
I
think
you
defined
it
here
properly.
The
way
maybe
I
would
like
I'm
looking
for
a
different
term,
so
I
want
to
check
to
check
to
come,
validate
the
device
capability
to
one,
a
specific
set
of
models
when
running
only
this
specific
set
of
products,
so
each
profile
is
tested
on
the
device.
B
When
the
device
is
running,
only
this
finger
can
say
exclusively.
The
purpose
here
is
not
to
try
to
assess
all
to
hang
people
to
understand
what
is
the
dynamic
state
of
the
device?
I,
don't
see
it
as
a
goal
of
the
working
group.
This
is
by
the
way,
a
much
harder
problem
to
kind
of
solve,
and
it's
I
don't
think.
This
is
the
scope
of
the
working
group
or
I'm,
not
even
sure,
it's
the
scope
of
onyx
to
try
to
Hank
the
application
understand
what
the
dynamic
state
of
the
device.
C
B
B
B
It
doesn't
say
that
it
is
up
to
the
application
to
understand
what's
the
situation,
but
the
working
group
of
con
is
not
trying
to
say
that
it
will
take
care
of
the
dynamic
node
on
the
GPU.
This
is
not
part
of
the
Vulcan
certification
process.
The
Vulcan
certification
is
all
about
functionality.
You
saw.
C
C
So
I
think
what
we
can
do
is
what
Milan
had
said
earlier
right.
Maybe
we
can
add
this
that
what
we
cannot
do
is
should
be
the
statement
that
you
just
made
that
we're
not
expecting
compliance
to
be
something
that
expects
an
edge
device
to
be
able
to
run
that
for
file
all
the
time
in
any
condition
right.
So
then
it
becomes
not
something
that
we
have
to
guarantee
all
the
time.
I.
B
A
I
mean
I,
guess
you
know,
this
is
the
you
know,
putting
the
bounds
and
the
whole
thing,
but
what
you
just
said
is
also
kind
of
saying
like
what
is
the
content
compliance
I
mean
if
you're
saying
it's,
not
the
you
know,
it's
not
required
to
be
compliant
all
the
time
and
then
it's
like
what
are
the
bounds
of
compliance?
You
know
what
does
it
make
sense
here,
but
I
mean
we
are
just
saying.
This
is
the
compliance
kind
of
suggestions
and
here's,
how
we
define
compliance
parameters
and
then
we
also
say
well.
A
C
A
Mean
we
have
to
put
some
parameters
around
this?
He
you
know
when,
when
you
say
compliance
we
are
compliant.
You
know
usually
describe
the
environment,
you
are
making
this
device
component
right,
meaning
like
okay.
This
is
like
with
you
know,
with
only
these
processes
running
on
device-
and
you
know
on
average
loading
or
something
as
you
can
describe,
this
device
is
expected
to
be
compliant
on
the
nominal
nominal
sort
of
situation.
You
know
and
not
under
heavy
loading
or
something
like
that,
so
I
mean
otherwise.
Compliance
doesn't
make
sense
and.
A
C
It
would
be
simpler
if
our
the
attributes
of
other
edge
profile
had
only
a
list
of
ops.
That
would
be
a
static
thing,
the
definition
that
we
have
as
part
of
our
profile
attributes.
We
have
latency
there.
We
have
power
consumption
there,
we
have
size
there,
we
have
accuracy
there.
These
are
all
dynamic
factors,
so
by
very
you
know
the
thing
where
we
started
from
right,
like
data
locality,
was
there
another
one
right
network
bandwidth
is
available.
C
B
Actual
name
I
disagree
with
this
statement:
okay,
accuracy:
okay,
that's
latency,
the
cue!
We
see
our
things
that
if
a
device
and
I'm
again
I'm,
saying
assume
a
profile
is
being
tested
on
a
device.
B
These
profiles
are
only
for
compliance.
Okay,
maybe
we
should
put
it
on
the
table.
The
profiles
that
we
describe
our
only
for
the
compliance
purpose
of
the
device
they
should
be
running
separately
on
the
device
where
nothing
else
runs,
and
then
the
device
is
expected
to
meet
this
accuracy
and
latency
as
described
in
the
buffer.
B
B
If
you
want
to
use
this
profile
also
to
describe
a
runtime
capability,
then
this
description
has
to
be
strict.
If
you
cannot
meet
the
accuracy
and
agency,
you
cannot
run
or
we
do
this
for
fun.
At
this
time
we
shouldn't
start
to
relax
for
fangs,
because
we
try
to
use
them
for
two
separated
purposes.
If
we
do
that,
it
loses
the
meaning
of
what
we
just
defined
okay.
B
So
if
the
device
I'm
getting
back
to
my
point,
if
what
you're
describing
there's
a
PC
I,
somehow
I
know
that
these
species
compliant
to
to
the
image
advanced
profile,
running
mobile
net
and
I'm
loading
mobile
net
and
I'm
telling
the
device
you
should
learn
it
in
a
2d
advanced
mode
either
he
can
run
it
or
he
cannot
I'm
not
starting
to
relax
day.
What
does
2d
advance
means
because
the
device
cannot
on
it,
so
the
device
in
his
green
state
is
capable
of
running
the
2d
advanced
for
fine.
B
B
Think,
maybe
maybe
we're
trying
to
stretch
it
too
much,
maybe
we're
trying
to
define
things
for
the
edge
devices
which
we
didn't
we
didn't
plan
to
you
know.
Maybe
we
should
focus
our
that
was
by
the
way,
our
purpose
myself
and
England's
purpose
all
the
way
from
the
start,
to
focus
only
on
compliance,
and
that
should
be
the
thing
I.
C
Think
we
have
some
ambiguity
here:
right,
we're
talking
edge
devices,
we're
talking
compliance,
strict
compliance,
also
and
so
first
edge
devices.
Our
definition
is
covering
scalable
computing
devices.
It's
not
fixed
speakers,
which
only
do
one
thing
with
pre-allocated
resources
right,
we're
talking
about
scalable
computing
devices
which
are
sharing
resources
and
the
memory
footprint
working
set
memory
changes
all
the
time
in
the
scalable
devices,
so
our
definition
of
edge
as
it
stands
right
now,
it's
covering
those
kind
of
devices
agreed
now.
C
C
How
can
you
guarantee
that
a
model
will
run
at
a
fixed
latency
on
a
device
which
has
so
many
other
things
running
all
the
time,
then
we
need
to
come
up
with
something
which
says:
okay,
these
profiles
are
either
optional
or
you
know,
an
edge
device
doesn't
need
to
go
and
say
that
I
am
a
static
profile.
You
know,
let's
make
these
the
definition
as
a
static
profile.
Let's
call.
B
B
Not
going
okay,
I
think
again,
the
purpose
was
components
of
the
device
in
a
test
environment.
In
the
scenario
of
testing
that
we
described
well,
the
only
thing
that
runs
on
the
device
is
the
tested
profile.
That
was
the
original
purpose.
If
we
try
to
extend
the
then
say
something
else,
then
I
think
we
should
react
salmon.
A
So
is
it?
Is
it
fair
to
say,
okay,
I
mean
to
exactly
that,
but
I
mean
maybe
saying
the
same
thing
that
offer
just
said,
but
the
thing
is
there
is
when
you
have
a
profile
that
you're
saying
you're
compliant
with
you
or
vendor,
as
a
vendor
is
free
to
describe
the
environment.
This
compliance
is
achieved.
A
If
my
system,
you
know
if
my
system
is
like
loaded
up
to
fifty
percent,
if
it
goes
beyond
50
percent,
I
cannot
guarantee
compliance.
I
mean
those
are.
Those
are
reasonable
things
to
say.
You
should
probably
be
free
to
explain
the
test,
conditions
or
environment
the
device.
Compliance
is
achieved
right
and
so
then
you
know
some
of
the
okay.
You
know
if
I
have
ten
different
things
running
on
my
device
and
other
things.
Well,
you
know
the
compliance
is
not
guaranteed,
but
if
you
are
with
these
within
these
parameters,
I
guarantee
you
will
have
compiled.
A
Yes,
similar
is
a
mobile
phone.
Mobile
phone
is
multi
hundred
process
device.
There
are
many
things
happening
on
these
devices
right
these
days.
There
are
many
computers
right
and
it's
like
you
know,
it's
the
same
thing.
These
dynamic
bright
environment,
but
I
would
imagine
you
know,
welcome.
As
a
former
you
know,
supplier.
A
Argue
that
if
we
will
have
a
device,
I
would
say:
ok,
yes,
we
are
compliant
with
this
device
with
this
profile,
assuming
nominal
use
case
of
device,
meaning
like
just
default
processes
running
on
device
after
buta
right
and
that's
it.
There
is
no
video
playing
in
the
background.
There
is
no
whatever
things
like
the
touch
sucking
the
power
off
and
all
that
so
I
mean
I.
Think
if
I
say
that
I
would
be
okay
in
my
mind,
so.
C
What
is
your
definition
of
device
here?
Also
in
a
phone,
let's
say
there
is
a
AI
accelerator
like
a
DSP
or
something
right
or
would
would
you
expect
this
profile
to
be
something
for
the
whole
phone
as
a
platform,
whether
it's
a
platform
that
would
be
compliant
or
are
we
talking
about
a
particular
device.
A
B
B
C
Actually,
coming
from
a
mindset
where
I
you
know,
I
don't
want
these
profiles
to
foster
restrictions
and
folks
are
not
able
to
innovate
right.
They're
not
able
to
hey
I,
cannot
go
and
do
some
power
saving
tricks
to
reduce
the
latency
or
to
save
power
or
to
compromise
on
accuracy,
because
it's
didn't
my
reach
and
and
the
profiles
become
that
restriction
that
people
are
not
able
to
do
that.
They
are
not
able
to
run
all
these
accelerators
together.
C
B
We
see
the
Gong's
difference,
but
the
condo
geo
describing
is
Sam
gone,
but
it's
not
the
gone
that
I
saw
of
this
walking
group.
A
gong
that
I
saw
was
around
ensuring
compliance
which
part
of
the
problem
is
that
people
got
today
in
the
world
of
machine
learning.
I
am
loading
my
application
on
some
system.
B
It
can
be
an
edge
device,
it
can
even
be
a
cloud
or
whatever
I'm
loading
it
on
a
system.
I,
don't
know
what
I'm
going
to
get
from
the
system.
No
Manoj
described
this
point
that
date
Microsoft.
They
have
this
problem
when
they
send
something
to
any
of
their
targets.
We
don't
know
if
this
target
supports
the
modern.
What
is
going
to
be
the
accuracy
in?
What's
going
to
be
the
latency
right?
Now
we
don't
have
this
mechanism.
B
A
You
know
thinking
about
the
device
as
a
device
device
and
how
you
run
this
particular
model
or
use
case
on
device,
because
on
the
same
device
you
can
have
different
or
multiple
execution
environments
or
frameworks.
For
example,
we
are
talking
about
variety
of
neural
network
runtimes,
you
don't
even
cover
that
I
mean,
but
actually
our
thinking
we
we
may
need
because
the
end,
you
know
you
have
n
graph.
You
have
welcome
solutions.
You
have
you
know,
you
know
you
have
all
these
different
frameworks
that
the
old
a
can
run
on
the
same
device.
A
C
B
B
A
This
this
far
in
into
the
working
group
actually
I,
want
to
raise
one
on
the
other
topic.
Maybe
I
have
only
a
minute
or
two
but
I'm,
not
sure
you
know
at
some
level,
you
know
the
way
we
started
with
the
edge
and
how
we
define
the
edge
and
how
much
it
includes
a
lot
of
things.
But
then
also
the
concept
of
profiles
and
concept
of
compliance
overall
seems
to
be
also
applicable
to
I.
Don't
know
any
any
kind
of
compliance,
not
just
edge
compliance
right,
I
mean
in
in
in
in
the
wider
sense.
A
A
C
C
B
Think
that
you
know
in
a
way
what
Megan
is
saying
is
corrected
it
just
by
saying
that
we
it's
an
edge,
walking
group
and
we
try
to
cover
all
of
the
aspects
of
edge
devices.
We
are
trying
to
cover
too
much
like
that.
When
someone
is
looking
at
the
walking
group,
he
says.
Oh
it
it's
going
to
cover
all
of
the
aspects
of
edge
devices
and
in
in
a
way,
that's
that
can
be
the
source
of
the
issue
that
we
need
to
break
it
down
to
smaller
components.
B
B
Just
like
mean
I
mentioned
compliance.
You
know,
you
always
talked
about
edge
devices
and
edging
the
infrastructure
edge,
and
so
what's
the
difference
sense
between
if
PC
the
turns
at
the
edge
of
the
infrastructure
and
a
PC
that
runs
inside
the
cluster
they'll
pretty
much
have
the
same
different
capabilities
of
performance,
but
canasa
fiying
what
they
can
do
in
describing
their
environment,
it's
pretty
much
the
same.
They
run
OSS,
they
run
applications.
So
it
looks
like
concept
of
compliance.
I
don't
have
to
be
on
the
edge.