►
From YouTube: ONNX Edge WG meeting 20191204
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
All
right
so
onyx
heads
working
group,
meeting
December
4th
present
are
myself:
Milano
Jack
offer
Rosenberg
syrup
degree
you
don't
do
and
and
Manoj
Manoj
and
I
guess
there
is
Manoj,
okay
and
so
I
sent
I
sent
you
the
screen
actually,
first.
So.
B
Okay,
I'll
do
a
brief
recap
of
the
workshops,
so
only
two
of
our
working
groups
are
present
here
on
me
and
Manish.
So
I
think
we
will
receive
the
email
which
contains
a
whole
slice
shortly
from
Microsoft.
So
maybe
this
email
welcome
within
pace.
So
the.
B
Email
so
actually
is
a
goals.
For
this
workshop
is
something
like
most
of
the
links
between
hardware,
vendors
developers,
users
and
Harry
King
I,
remember
from
Intel
introduced
the
steering
committee,
the
current
situation
of
the
steering
committee,
the
rows
of
the
elections
and
the
cigs
and
working
group
chairs,
and
then
we
have
some
partner
updates,
update
everyone
on
the
current
state
of
onyx,
how
many
developers
we
have
how
many
companies
joining
and
how
many
operators
or
we
are
having
on
how
many
runtimes
we
have
right
now
and
yeah.
B
One
thing:
the
Foundation's,
yes,
the
next
foundation,
I
ever
figure.
So
actually
we
formed
a
working
group
called
foundation
WG
to
do
a
survey,
and
then
we
voted
to
join
the
FBI's.
There
is
a
woman,
I
think
she's
called
Jessica
Jessica
Jessica
Kim,
and
she
is
the
8a
aya
officer,
the
Outreach
Committee
chairperson,
and
she
she
speaks
on
behalf
of
every
eyes
and
onyx
is
part
of
the
FBI
right
now.
So
onyx,
together
with
Akuma's
I,
think,
are
the
only
few
projects
which
were
graduate
projects
under
FBI
guys.
B
So
this
is
one
thing
foundation
and
the
foundation
working
groups,
then
I
think
post
Microsoft
and
the
Intel
dude
did
some
presentations.
One
thing
is
the
steering
committee,
so
five
steering
committees,
Microsoft
Intel,
media,
Amazon
and
Facebook
I-
see
maybe
two
or
three
prison
in
China,
Shanghai,
Microsoft
and
Intel
are
definitely
were.
Definitely
there.
Amazon
and
Facebook
were
absent
and
not
sure
whether
a
media
present
or
known
I
think
might
be
the
Machine
some
immediate
China,
a
local
guy
to
attend
this
workgroup.
So
okay,
yeah
yeah
for
Microsoft
Z
just
do
some
basic
updates.
B
Think
Harry
introduced
the
working
group
and
seek
situations
for
models
o,
which
is
one
of
the
four
six,
and
what
also
is
not
established
right
now
yet,
and
actually
the
onyx
committee
is
welcome
if
there
is
a
volunteer
or
our
company
volunteered
to
pick
up
this
role
as
the
models
o
leaders
and
another
one
is
a
quantization
quantization
is
close
and
foundation.
Working
group
is
still
alive,
I
think,
even
though
we
already
voted
to
join
the,
even
though
the
presentation.
B
B
So
yes,
I,
remember:
Intel
introduced
something
about
Asia
and
Mail
Plus
Intel,
open,
V,
note,
plus
onyx
long
time
actually
is
a
very
interesting
combination
to
do
gesture
recognition
and
the
introduced
some
of
their
products
and
openly
no
based
ap,
covering
from
cloud
to
edge
and
onyx
for
entire
integrated
ways.
Open,
we
know,
will
be
used
in
the
verification
of
onyx
models,
also
display
some
speed
up
spy,
open
Thanos
and
that's
about
as
I
saw
about
the
student
committees.
I'm.
Sorry.
B
Repeat
what
open,
what
open
V?
No,
yes,
Arab!
You
should
know
that
yep
I
know
yeah
open,
V!
No,
and
there
are
also
some
interesting
topics
in
some
Chinese
companies
like
JD
the
deploy,
onyx
models
on
Android,
with
onyx
runtime
and
Tencent,
which
Z
couturier
NCAA
mobile
phones
and
for
working
groups
and
the
six
updates
I.
Remember
a
guy
from
Microsoft
he's
called
Kirchen
and
he
is
in
charge
of
the
infra
6.
He
spends
about
10
minutes
around
the
6,
which
is
a
pretty
stable
because
on
all
the
basics,
things
had
already
done.
A
A
B
B
C
You
covered
most
of
that.
Most
of
the
points.
The
only
thing
I
wanted
to
bring
up
is
that
we
put
up
a
question
about
the
sig
taking
on
whether
you
know
the
they
will
be
fine,
a
method
for
how
onyx
models
will
be
executed,
which
would
then
be
used
for
any
profile
definition.
If
that
is
going
to
be
a
priority
for
for
the
cig,
but
there
wasn't
any
feedback.
We
put
the
question
on
the
table,
but
we
didn't
get
any
like
feedback
during
the
session.
C
A
So
so
what
I
want
to
you
know
why?
My
you
know
second
topic
on
agenda
here
was
like
you
know:
I
wanted
to
basically
to
hear
that
and
then
you
know
we
have
been
like
him
as
I
stated
in
this
police
report.
Really
we
were
basically
bit.
You
know
certain
thing
for
you
know
last
three
months
and
no
significant
progress
was
made,
and
so
you
know
discuss
discussed
the
pot
forward
right.
A
A
A
A
We
we
seem
to
have
the
disagreement
in
terms
of
how
do
we
kind
of
defines
the
little
work
and
what
it
should
be.
I
think
that
that
is,
you
know
clear
to
everybody,
and
so
what
do
we?
What
do
we
do
about?
It
is.
Do
we
continue
or
do
we
think
there
is
a
there?
Is
a
value
in
to
in
continuing
or
not
I
would
like
to
kind
of
survey,
thoughts
and
opinions,
because
clearly
we
are
not
making
big
progress.
Last
three
months.
C
From
an
ecosystem
perspective,
I
believe
that
there
is
value
in
the
working
group
to
go
ahead
and
define
the
set
of
parameters
required
to
execute
onyx
models
across
different
different
hardware
platforms.
The
fragmentation
in
the
ecosystem
is
going
to
increase
challenges
in
terms
of
how
consistent
the
onyx
model
execution
is
going
to
be.
So
there
is
a
problem
and
if
the
state
of
work
goes
out
and
addresses
or
tries
to
solve
that,
that
would
be
value.
A
C
C
To
be
executed,
maybe
there
are
performance
boundaries.
Maybe
there
are
you
know
accuracy
boundaries
quantization
all
of
that
right.
So
if
the
spec
and
onyx
can
define
how
a
model
model
has
a
preferred
of
execution,
that
would
you
know,
put
the
responsibility
on
the
hardware
to
ensure
that
it
can
meet
that.
D
D
D
That
we
are
targeting
the
ring
phase,
we
can
define
stuff
and
then
we
can
have
other
onyx
members
coming
and
saying.
This
is
actually
not
the
thing
that
she
should
have
defined,
and
that
would
happen
after
many
months
of
hope,
so
I'm
concerned
from
the
lack
of
involvement
of
other
onyx
members
in
the
world.
Multiple
attempts
throughout
the
three
last
workshops
to
increase
the
number
of
medals
and
I
didn't
see
too
many
hardware,
vendors
addressing
it.
A
That
is
one
way
to
think
about
it,
sure
what
posing
means
I
mean
if
he
posed,
for
you
know
without
activity
for
for
a
couple
of
months
by
rules
of
regulations
or
whatever
it's
called
governance
rules
right.
It
is
automatically
working
group
which
is
kind
of
disband
I
mean
in
a
sense
like
you,
there's
no
activity,
then
it's
it
becomes
like
cancel
right.
So
if
he
posed
for
more
than
three
months,
that
will
automatically
mean
this.
You
know
breakage
or
working
group
would
be.
A
Disbarred,
basically
of
what
is
the
proper
term,
it
would
be,
can
stop
and
so
I
mean
yeah
I
mean.
That
is
the
way
that
think
about
it.
Is
it
like
okay?
So
if,
if
we
are
trying
to
impose
and
raise
2
to
to
kind
of
specify
to
the
hardware
that
how
models
should
be
executed
and
what
is
the
expectation
of
execution
without
wider
hardware,
vendor
participation,
it's
hard
to
kind
of
say,
oh,
this
is
the
way
to
go.
I
mean
we
don't
know
who
is
gonna
technically
barf
on
this
going
forward
right
and.
A
A
You
know
that
often
open
Meno
will
be
used
as
a
framework
to
validate
onyx
operators
right
is
that
what
we
are
onyx
models
over
onyx
models
right
each
model
Zoe's
coming
there
but
model
models.
The
working
group
was
in
making
right
or
something
the
models
will
see.
Actually
models,
ooh,
sick,
okay,.
A
E
A
C
I
I
don't
think
we
want
to
say
that
this
is
a
hardware
qualification
program
or
hardware
qualification.
Working
group,
I
I
would
say
that
what
is
missing
is
a
vendor
or
as
a
hardware
infraspecific
inference
engine
if
there
is
no
guidance
on
how
to
locate
eyes
across
the
different
computer
locks
and.
C
I
mean
NL
off
is
trying
to
standardize
how
hardware
itself
performs
across
standard
models,
but
in
the
real
world
it
is
not
going
to
be
about
benchmarking
or
getting
the
you
know,
showcasing
a
hardware,
its
value.
It
will
be
about
how
you
can
maximize
peb
power
and
compute
on
that
end,
point
to
to
execute
the
specific
model
that
you
have
so
I
know
I'm
saying
a
lot,
but
there
are
two
things
here:
one
is
from
a
hardware
perspective,
vendors
perspective.
Perhaps
it
is
all
about.
C
You
know
benchmarking
and
mixes
showcasing
the
power
of
the
hardware
that
is
covered
by
NL
poor,
but
from
a
developer's
perspective,
if
I
have
an
onyx
model
and
I
want
to
allocate
it
or
execute
it
across
this,
you
know
Universal
edge
hardware.
The
model
should
be
able
to
give
some
guidance
on
you
know
what
kind
of
computer
or
private
computer
it
is
if
it
should
be
used
with.
A
C
Cloud
the
cloud
is
a
cloud
is
not
that
big,
a
problem,
because
the
ecosystem
in
cloud
is
relatively
controllable.
The
cloud
is
all
about
data
center
hardware,
which
gets
qualified
and
and
assorted
through
a
process,
so
there
only
exists
some
boundaries
on
what
a
a
cloud-based
you
know,
shape
or
an
inference
engine
can
and
cannot
do
on.
The
edge
is
where
we
really
don't
know
what
hardware
is
gonna
get
deployed
where
I.
A
Through
the
whatever
API,
this
is
generic.
That's
what
I'm
saying
it's
it's!
It's
general
framework
to
prioritize
what
you
call
it
execution.
So
in
that
sense,
it's
like
beyond
edge
right.
Do
we
agree
on
that?
It's
it's
just!
It's
it's
beyond
edge
it,
because
it's
it's
really
generic
request
for
for
model
execution,
prioritization,
basically
and
giving
hints
through
hardware-
and
you
know
and
I
get
the
point
that
edge
is
is
more
importantly
from
that
perspective,
but
it's
more
relevant
to
the
edge,
but
as
a
concept,
its
its
generic
I
agree.
A
A
A
Really,
you
know
what
is
the
set
set
of
models
or
set
of
operators
that
the
given
profile
or
given
Hardware
can
do
now?
If
you
say
just
that,
then
technically,
you
know
anybody
who
is
supporting
onyx.
You
know
there
is
a
model
zoo
and
they
could
say
yeah.
You
can
run
these
models
in
the
model
Zoo,
therefore,
operators
that
are
part
of
this
model
Zoo
or
these
models.
A
Technically
we
support
right
the
aspect
of
benchmarking
and
all
that
like
how
how
accurate
and
all
that
is
perhaps
secondary
from
the
perspective
of
somebody
who
wants
to
use
a
certain
solution,
they
would
like
to
know
first
can
I,
do
the
model
can
I
do
operators
on
the
given
platform,
yes
check
and
then
take
in
benchmark
and
providing
benchmarking
numbers
could
come
from
the
vendor
itself.
They
could
say
well,
this
is
what
we
can
achieve
for
these
models
and
so
on,
and
perhaps
this
is
part
of
the
whole
model.
Zoo
working
group
or
cig.
A
That
is,
you
know,
can
say
that
you
know
vendors
should
advertise,
which
models
from
models
do
they
can
do
and
be
protectors.
So
it's
kind
of
benchmarking.
I
mean
we
don't
want
to
get
into
benchmark
right
technically,
but
you
know-
and
so,
if
you
put
this
bounds
like
that,
then
it
becomes
like.
A
Maybe
maybe
the
problem
statements
shifts
a
little
bit,
you
know
in
in
the
context
of
really
going
back
to
the
you
know,
maybe
I'm
trying
to
decode.
Why
is
the
lack
of
interest
in
what
you're
doing
right
and
from
the
general
hardware
vendors?
They
don't
care,
apparently
right
and
and
either
inner,
we
don't
advertise
enough
or
we
don't.
A
We
did
bring
it
up
in
workshops,
but
so
far
with
little
kind
of
you
know,
engagement
level,
and
so
that
tells
me
like
really
nobody's
like
super
excited
about
this
now
the
context
that
was
brought
up
recently
to
the
discussions
about
prioritization
and
how
do
we
prioritize
or
control
execution
is
different,
and
that
is
that
is
something
that
we
as
a
group
could
say.
Okay,
this
is
not
part
of
this
group.
This
is
technically
important,
but
it's
a
general
concept.
A
Give
it
give
notice
to
foundations,
sake
and
say:
missus
we
think
is
important
and
could
please
consider
it
right
and
apparently
you
did
brought
it
up.
You
did
bring
it
up
in
the
workshop,
but
we
didn't
I
guess
get
any
immediate
feedback
on
that
from
either
and
so
so
and
I
circle.
All
this,
it's
just
like,
okay,
what
is
the
purpose
of
this
group
right
and
and
that's
the
question
that
I
am
posing
like?
Do
we?
What
is
the
path
forward?
Do
we
do
we
continue?
Do
we
pose?
Do
we
decide
to
say
okay,
minashi.
E
Last
time
you
mentioned
that
we
can
post
the
steering
committee
to
give
us
some
indication
on
how
to
describe
execution
and
all
that.
Yes,
they
didn't
answer
it
at
the
workshop.
But
do
we
have
a
line
of
sight
like
you
know?
Maybe
you
did
need
more
time
and
more
data
to
ponder
on
and
we
need
to
formally
approach
them
in
another
way
to
get
a
formal
answer
from
them.
Or
do
you
think
just
they
are
not
interested
I.
C
Got
the
sense
that
we
will
need
to
formally
approach
them
because
the
work
subsection
and
the
and
the
the
pace
at
which
we
went
through
topics?
This
may
not
have
this.
You
know
this
may
not
have.
You
know
had
the
time
to
go
to
bait
and
deliver
it
over.
So
I
actually
would
think
that
if
there
is
another
forum
where
we
could
approach
and
so
that
this
can
be
debated
and
then
decided
that
okay
Ike
make
sense
that
we
we
cause
on.
C
And
pick
up
the
matter
elsewhere,
that
can
be
I,
I.
Think
this
group,
although
you
know
we
don't
have
more
than
three
hardware
vendors
here
there
is
definitely
representation
and
and
but
there
are
more
hardware
partners
that
we
should
bring
in.
That
may
be
the
question
that
we
give
the
steering
committee.
D
A
A
A
A
You
know
the
directly
approach
steering
community
at
this
point
without
waiting
for
them
to
come
to
us,
even
though
we
are
bringing
it
up
to
workshop
and
all
that,
but
she
basically
does
just
have
maybe
dedicated
approach
and
suggest
and
propose
what
we
you
know.
Basically
things
that
we
discussed
in
this
meeting
and.
A
So,
let's,
let's
okay
I'll,
take
it
myself
to
draft
the
email
or
letter
or
something
other
email
which
I
would
I
would
then
share
with
you
guys
for
feedback.