►
From YouTube: ONNX Edge WG Meeting 20191113
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).
A
B
A
A
A
A
The
key
challenge,
let
me
start
stating
maybe
from
from
last
last
week's
discussion
and
that
we
had
and
I
think
my
mash
wasn't
there
though
it
was
it
was
about.
It
seems
like
we
came
to
the
a
bit
of
a
crossroads
in
terms
of
this
working
group,
to
me,
at
least
in
a
way
that
we
couldn't
totally
agree
on
updated
statement
of
work
and
what
is
the
scope
of
our
efforts?
A
I
think
it
was
a
discussion
between
like
stated
using
profiles
as
a
compliance
criteria
and
defining
compliance
around
the
profiles
versus,
like
you
know,
having
that
having
and
how
that
applies
to
the
to
the
devices
like
dynamic,
dynamic
execution
environment.
Like
you
know,
for
example,
pcs
right
arm
up
tops
and
what
does
what
does
compliance
means?
Furthermore,
it
felt
like
there's
also
desire
to
you
know,
to
engage
into
defining.
A
Using
profiles
as
execution
control
mechanism
or
a
hint
to
the
for
the
for
whoever
is
running
the
divide,
the
model
on
device
that
they
can
be
hinted
to,
okay
to
to
execute
the
given
model
and
a
certain
profile,
and
so
those
are
two
kind
of
I
felt
known.
As
you
know,
kind
of
conflicting
arguments-
and
there
is
I
would
say
there
is
no
agreement
right
now
in
terms
of
the
in
between
the
contributors
of
this
workgroup
in
terms
of
what
should
be
should
not
be
in
a
state
in
the
work.
A
C
B
Right
now,
onyx
does
not
describe
how
to
run
a
model.
It
describes
what
to
run
in
a
model
right
unless
onyx
expands
the
scope
and
starts
making.
You
know
contributions
in
the
spec,
where
we
can
describe
that.
This
model
needs
to
be
run
like
this.
Only
then
I
think
as
a
working
group.
We
can
take
that
off.
B
The
challenge
I
was
having
with
different
approach.
We
have
with
edges
edge
profile
this
as
part
of
some
of
the
attributes
in
edge
profile.
We
are
describing
things
like
accuracy,
things
like
you
know,
through
port
and
thing
you
know
other
performance
kpi's,
which
gold
ventures
are
sent
to
a
place
where
we
are
actually
saying
how
to
run
a
model
which
makes
it
somewhat
ambiguous
for
a
multi
function
or
integrated
device,
which
has
many
ways
to
run
the
same
platform
and
where
we
left
off
last
week
was
like
in
an
integrated
device
like
a
PC.
B
What
gets
certified?
Is
it
just
the
IP
or
we
are
certifying
the
whole
platform,
because
I
could
run
the
same
model
either
on
a
GPU
which
would
be
very
different
in
terms
of
accuracy,
then
some
other
as
accelerator
or
whether
you
know
like
CPU
would
be
different.
So
that's
where
probably
we
need
a
little
bit
more
clarification.
What
we
mean
by
profiles
and
how
do
they?
D
I
think,
let
me
try
and
provide
my
point
of
view
on
that
and
I
think
in
the
way.
There
is
some
reference
to
that
because
listen
to
me,
I've
been
looking
into
the
MMF
published
results
and
I'm
thinking
about
applying
some
of
the
principles
in
their
description
of
how
to
do
testing
into
what
we're
trying
to
do
so.
D
Doesn't
help
I
think
what
getting
for
a
second
to
day
basis
of
what
we
want
to
do.
We
want
to
create
a
profile
that
would
map
into
some
kind
of
tests,
so
this
profile
is
for
certification.
It
would
had
a
mapping
into
a
test
package.
The
way
that
I
see
it-
and
maybe
we
should
probably
better
clarified
in
the
statement
of
work
profile-
is
to
certify
something
and
it
is
connected
to
a
test
package.
The
way
that
you
learn
this
test
package,
I
think
is
where
we
need
to
better
clarify.
D
D
On
ideal
state
would
be
able
to
add
these
tests
package.
That's
one
thing
and
the
other
thing
part
of
the
test
they
vendo
that
is
going
to
pass.
The
test
would
need
to
provide
details
on
what
were
the
test
conditions,
so
if
he
has,
for
example,
take
Snapdragon
Qualcomm
Snapdragon,
we
have
four
ways
right
now
to
our
stuff.
Okay,
we
can
run
it
in
four
different
ways.
Basically,
each
one
of
these
ways
can
have
different
characteristics.
D
Maybe
part
of
the
recommendation
would
be.
How
can
we
extend
it
beyond
just
certification,
validation
into
other
usages
right
and
then
meeting
what
you
said
so
that
may
require
some
additional
changes
into
the
modern,
a
photograph
it
will
require
extending
the
scope
of
entire
Onix.
Maybe
that
can
be
part
of
next
steps
or
commendation
that
shouldn't.
In
my
view,
it
shouldn't
gate
the
activity
of
this
working.
B
C
So
I
offer
I
agree
with
you.
I
wanted
to
add
that
you
know
why.
We
all
agree
that
we
need
to
define
the
profiles
right.
There
is
no
disagreement
on
the
need
to
define
the
profiles.
I
think
where
we
are
going
back
and
forth
is
how
that
information
is
going
to
be
used
from
a
device
perspective.
I
think
what
you're
suggesting
is
that
the
profile
information
could
be
used
to
validate
capability
of
the
device
about
executing
model.
I
would
also
like
to
add
that
that
profile
information
could
be
useful
for
the
model
creator.
C
A
C
A
Why
I
mean
in
a
sense
like
if
we
are
this
sorry,
if
I,
if
we
are
defining
the
test
package
to
continue
on
offers?
Oh,
if
I
understood
it
correctly,
if
we
define
we
have
a
profile,
and
there
is
a
test
package
that
describes
clearly
on
giving
device
and
how
to
execute
and
actually
completely
describes
how
to
execute
the
given
model
on
a
given
device
with
what
environment?
It's
like
full,
full
description
of
the
setup
right
and
how
to
well
proposal
either.
A
Did
we
say
that
the
test
package
that
is
provided
is
is
providing
the
bounds
on
how
this
profiling
is
executed?
And
hence,
if
you
do
it
that
way,
it
will.
You
know,
meet
meet
the
given
profile
characteristics,
and
that
is
for
us
to
say
if
this
is
satisfied.
That
means
the
device
under
these
circumstances
when
tested
like
this
is
meeting
the
profile
requirement,
and
now
it's
not
having
something
more
formal
in
the
on
expect.
That
is
kind
of
embedding.
This
information
into
the
profile,
perhaps
is
not
necessary.
B
B
B
Very
much
possible
in
a
IOT,
tiny
device
there,
not
a
lot
of
things
are
going
on,
but
in
a
general-purpose
computing
device
you
have
sometimes
thousands
of
processes
running
all
the
time
running.
An
IP
does
not
mean
you
don't
need
a
CPU
it
might
you
know
it
just
depends
on
the
CPU
you
have
and
just
because
we
are
trying
to
be
very
comprehensive
and
we
want
this
to
be
a
set
fiying
like
a
compliance
kind
of
a
thing.
That's
what
is
worrying
me
that
this
ambiguity
might
hurt
us.
B
You
know
compliance
around
it.
It
will
be
safe,
at
least
in
my
experience.
I
don't
have
any
other
instances
where
such
a
thing
has
been
done
before,
because
it
being
a
general-purpose,
very
chaotic
environment
in
case
of
you
know,
like
a
general-purpose
device
like
the
PC,
which
is
configurable,
but.
D
This
is
exactly
again.
If
you
look
at
em
and
pills.
This
is
exactly
what's
happened:
engineering
pills,
there
is
a
test,
and
companies
submit
the
results
of
the
tests
that
include
how
many
influences
per
second
and
what
was
the
accuracy
of
what
they
managed
to
achieve,
and,
of
course,
when
they
do
that
and
by
the
way
there
is
a
test
of
intense
emitting
CPU
tests.
I
would
imagine
in
10
submitted
a
bunch
of
different
tests.
Present.
One
of
them
was
CPU
only
the
other
one
was
CPU
with
the
new
n
MPI
accelerator.
D
So
I
would
imagine
that
when
you're
running
mm
12
tests
on
the
CPU
only
you're
trying
to
make
sure
that
your
system
is
as
clean
and
as
basic
as
possible,
because
this
is
what
the
test
defines
that
you
should
strive
to
get
to
a
situation
that
your
system
or
the
tested
IP
is
in
maximum
possible
capacity.
Whatever.
However,
you
do
that,
and
but.
B
D
D
What
we
want
to
do
is
to
cover
a
wider
range
of
models
and
tell
you
that
in
I
think
the
maximum
best
scenario
that
you
can
do
you
are
able
to
reach
this
latency
and
this
accuracy
it's
not
given
by
the
way.
It's
not
it's,
not
there
100%
sure
that,
even
in
ideal
connection
irrigations
the
device
would
be
able
to
achieve
these
fins.
Okay,
I
understand
your
point
that
you're
saying
the
device
was
able
to
achieve
it
in
an
ideal
congregations.
D
D
B
A
Things
more,
if
we,
our
term
compliance,
is,
is
a
strong
term.
I
I
mean
in
essence
what
we
are
trying
to
achieve,
and
we
started
a
long
time
ago,
mentioning
potentially
you
know,
I'm
aligning
with
Mr
perv
as
a
kind
of
effort
or
body
that
we
could
could
leverage
and
the
approach
for
a
sort
of
certification.
In
essence,
the
Emer
perf
is
providing
two
things:
performance
and
accuracy
they
are
now
they
are
not.
They
are
not.
What
is
not.
There
is,
but
it's
not
saying
they're
not
going
step
above.
They
are
not
saying.
A
They're
just
reporting
results
right,
and
this
is
what
we
you
can
achieve.
If
you
do
this,
you
know
you
run
the
test
like
this.
You've
got
this
right
and
we
are
trying
to
say
that
we,
we
are
tying
these
test
results
with
with
meeting
criteria
that
belongs
to
given
profile,
and
this
is
where
we
are
struggling.
This
is
the
step
where
we
are
going.
I
feel
saying
like
this
is
a
strong
thing
that
you're
saying
you
need
this
profile
characteristics
and
then
you
say
you're
compliant
with
this
profile.
That
is
our
goal
right.
A
That
was
our
general
guide
right
now,
so
so
so
clearly,
test
results.
If
you
say,
if
you
use,
am
a
perfect
use
case
and
they
are
reporting
the
results
with
a
well-defined
infrastructure
or
rules
around
it,
submitting
the
actually
full
test
package
and
test
execution,
environment
and
all
that
stuff.
So
once
you
get
this
set
up,
their
report
results
accuracy
and
performance.
We
have
a
few
others,
but
maybe
that's
like
now
going
beyond
it,
but
let's,
let's
stay,
let's
name
it
performance
and
accuracy.
A
With
this
setup
that
is
very
well
defined,
set
up
right
for
any
device,
you
know
it
is.
There
is
like
anybody
could
report
with
a
description
that
is,
that
was
provided
once
you
have
results.
It
is
I
mean
it's
technically
saying.
Okay,
now,
I
have
these
results
now,
given
these
results,
I
am
compliant
with
these
Onix
profiles
that
we
define.
That
is
all
you
are
trying
to
say
is
that
ambiguous.
B
A
Point
I
guess
point
being
is
like
even
on
the
laptop
or
anywhere
right
you
could.
You
could
describe
a
test
environment,
meaning
like
I,
am
booting
the
stock
windows.
Whatever
and
running
you
know
you
could
describe
exactly
how
you
boot.
Yes,
there
are
many
processes
behind
the
scenes
you
could
say:
okay,
y5
is
turned
off.
Screen
is
turned
off.
You
could
describe
whole
set
of
things
that
will
constrain
environment
right
and
you
say
and
I
run
test.
A
I
mean
this
is
as
good
as
you
can
get
them
for
chrome,
Snapdragon
or
you
know,
or
or
we
you
know
to
us.
That
is,
you
know,
acceptable
in
a
way
you
describe
your
environment
and
you
say:
I
ran,
50
runs
in
this
setup
and
I
get
these
results
under
these
conditions.
I'm
making
these
profiles,
that's
basically
the
workflow.
A
Now
if
we
are
saying
that
that's
not
sufficient
to
describe
the
the
you
know,
the
test,
setup
or
constrained
constrained
the
execution
around
the
test
and
constrain
other
pcs
in
a
more
complex
devices
like
PC,
then
okay,
then
you
know,
then
we
have
V.
Don't
we
have
this
agreement?
Basically,
right,
that's
that's
where
we
are
and
so.
C
C
A
D
D
B
D
Saying
once
you
meet
big
old
ten
milliseconds,
you
are
compliant,
so
it's
a
different
state
of
mind
and
we
can
actually
talk
about
when
publishing
the
actual
zone
off
the
test.
How
do
we
make
sure
that
someone
is
maybe
there's
a
way
for
someone
to
punish
the
results
without
actually
publishing
the
actual
number?
D
That's
a
different
discussion,
but
to
your
point,
I'm
Manoj
about
what
do
we
do
with
these
results?
It's
a
very
good
point
and
it's
kind
of
a
wider
question.
I
forgot
if
it
was
you
away
so
as
that
raised
it
a
couple
of
minutes
ago,
the
question
is
actually
what's
the
scope
of
onyx
in
general,
the
onyx
project,
so
in
the
limitations
of
the
existing
onyx
project,
the
best
we
can
do
now
is
to
create
a
location
in
github
that
people
can
put
results
into
and
we
can
publish
the
results.
D
This
is
what
we
can
do
with
the
current
limitations
of
onyx
activities.
We
can
come
up
with
a
recommendation
that
says
there
needs
to
be
a
way
to
query
that,
and
that
way
is
probably
some
API
or
something
like
that,
but
that
would
require
onyx
steering
group
to
come
and
say
that
they
wants
to
create
a
working
group
that
defines
an
API
because
so
far
there
isn't
such
an
effort.
D
So
in
order
to
have
our
working
groups
successful
and
meet
whatever
it
needs
to
do,
we
need
to
to
bound
it
to
bound
it
so
that
we
can
achieve
something
within
the
limitations
of
the
existing
onyx
project
and
provide
recommendations
for
the
next
steps
to
the
still
and
go
either
way.
That's
the
way
that
I
feel
I
think
that
if
we
won't
come
up
and
start
spending
time
on
defining
an
API,
we
would
come
back
to
the
SIG's
or
the
steering
groups,
and
they
said
well.
D
That
was
not
the
mandate,
and
this
is
actually
what
something
beyond
the
scope
of
onyx.
So,
instead
of
spending
time
defining
the
API,
we
should
just
come
with
commendation.
That
says,
as
a
next
step,
we
suggest
that
the
onyx
steering
group
would
extend
the
scope
of
onyx
project
to
define
an
API
once
he
does
it.
We
kick
off
a
working
group
to
describe
exactly
how
the
API
looks
like
that
queries
or
finds,
and
maybe
by
the
way
this
API
is
also
going
to
do
what
Josh
wants
to
do,
which
is
to
Kweli
I.
B
B
The
IP
is
itself
balanced,
like
the
throttle
thermal
because
of
thermal
reasons
or
frequency
reasons,
and
we
have
power
as
an
attribute
theorem,
struggling
is
I,
don't
see
a
way
to
create
an
isolated
environment
without
having
an
IP
level
control,
and
you
know
actually
be
able
to
describe
the
working
environment
all
the
way
to
the
ID,
because
you
know
without
it.
The
results
would
not
make
a
lot
of
sense.
B
What
what's
the
number
I
should
be
coating
at
that
time,
because
I
ended
up
picking,
something
which
impacted
my
accuracy
also,
so
it's
quite
the
environment
that
we're
running
in
needs
to
be
also
described
when
we
are
coming
there
I
to
me.
They
are
just
hand
and
glove
kind
of
a
thing,
otherwise
that
we
have
ambiguity
so
in.
C
D
Okay,
well
so
I
think
we
have
to
log
it
into
decides
that
we're
going
to
present
in
the
workshop
to
kind
of
show
what
are
the
questions
that
we
have
been
struggling
with
and
to
get
to
some
quality
saying
this
is
what
we're
going
to
focus,
and
we
know
that
this
is
there's
some
extra
step
that
needed,
but
we
think
that
it
requires
a
change
to
the
other
Onix
mandate
in
that
change.
Before
doing
that,
there
is
no
before
agreeing
on
this
change.
There
is
no
point
doing
further
into
defining
the
second
block.
D
Group,
we
should
provide
recommendations
in
general
lines
how
these
funds
can
be
further
used
to
query
what
the
device
can
do
and
to
query
the
dynamic
state
of
the
system,
so
that
someone
who
do
what
to
do
it
describe
it
in
a
general
and
give
it
to
the
steering
group
and
say
that
requires
defining
an
API
which
is
right
now
outside
the
scope
of
Onix
project.
Do
you
approve
this
activity?
Do
you
approve
extending
the
Onix
project
scope
if
they
do
there's
a
point
in
going
into
the
details?
D
B
A
D
A
A
So,
let's
agree:
what
what
do
we
need
to
put
up
in
front
of
the
audience
in
in
Shanghai
right?
So
we
are
saying
that
iris
I
mean
I'm,
saying
like
based
on
this
discussion.
We
are
saying
we're
in
the
Crossroads
instance
like
there
is
a
how
and
there
is
what
right
may
be,
the
compliance
part.
There
is
a
car
part
how
to
how
to
basically
define
how
part
is
a
critical
here,
because
without
that
you
know
how
to
run
certain
tests,
and
you
know
you
need
certain.
A
B
Nice
to
know
it
is
so
I
think
by
the
way,
I.
You
know
we've
also
pondered
on
the
security
aspect.
Do
you
guys
know
what
is
the
official
position
of
onyx?
What
didn't
mean,
but
security
here,
like
the
other
attributes,
I
make
sense
they're
generic
enough,
but
security
could
mean
different
things
to
different
people.
I
think
we
stayed.
A
C
B
A
D
I
think
that
if
we
try
to
describe
the
test
environment
into
more
details,
that
can
also
sell
in
to
qualify
the
point
that
the
idea
is
that
all
of
your
processing
stays
on
the
device
or
not
in
a
sense,
and
once
you
qualify
the
test
environment,
that's
for
the
profile.
You
actually
verify
this
point
as
well.
A
So
so
I
think
this
statement
here
you
just
sort
of
question
I
mean
I.
Don't
think
these
are
the
final
attributes
we
came
up
with
right.
This
was
like
female
work
about
definition
of
edge
profile
and
what
what
attributes
constitute
edge
profile?
We
have
a
document
stating
what
are
those
attributes
right?
So
no,
not
necessarily
these
listed
here.
You
know
we
have
different.
We
have.
We
have
power,
accuracy,
speed,
data
locality
and.
A
Sighs
right,
sighs,
yeah
power,
accuracy,
speed,
sighs
and
data
locality.
Those
are
the
five
attributes.
Did
we
define?
We
have
a
basically
concluded
on
the
document
now
whether
that
security
I
don't
know,
but
this
as
I
said
I
just
want
the
point
here
that
these
are
listed
here
as
examples,
not
necessarily
what
we
decided,
and
so
we
finish
step
two
and
we
have
a
edge
profile
definition
document
defining
the
attributes.
C
A
C
A
A
A
Sort
of
you
know,
I,
don't
say
block,
but
in
some
way
we
are
blocked
because
we
are,
we
didn't,
have
much
progress
in
defining
any
profiles
since
last
workshop.
So-
and
this
is
why
right
this
has
to
be
all
up
in
front.
Otherwise,
it's
it's
just
not
accurate.
Do
you
agree
to
do
this
kind
of
you
know
presentation
saying
that
the
working
group
is
basically.
C
E
D
So,
in
a
way
we
are
returning
to
redefine
part
of
the
statement
of
walk
and
to
define
a
core
fine,
because
we
also
know
it's
clear
that
we
have
to
define
out
of
the
perfect
definition,
needs
to
be
defined,
test,
environment
and
defined
which
device
was
tested
and
that
that's
probably
going
to
be
about
us
running
the
profile
on
some
platform
for
certification
purposes.
So
I
think
all
of
this
we
can
actually
come
into
describing
in
the
next
Walkinshaw
workshop
session.