►
From YouTube: ONNX Edge WG meeting 20190822
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
The
first
slide
is
just
the
same
as
Megan
showed
talking
about
the
working
group
and
some
logistics,
and
everything
second
slide
is
deny
that
I
mentioned.
We
should
probably
add
so
that
people
would
understand
what's
the
scope
of
the
working
group,
so
this
guy
just
takes
the
drawing
and
shows
that
whatever
favors
edge
is
both
edge
infrastructure
and
edge
devices.
That
can
be
anything
from
mobile
phones,
drones,
cows,
virtual
reality,
so
just
to
give
people
the
right
scope
to
explain
what
edge
is
about
and
then
the
next
three
slides.
B
B
A
A
B
So
I
would
really
like
more
feedback
from
both
of
you,
because
this
is
still
very
early,
and
you
know
it
still,
even
though
there
are
things
written
up
here.
You
know
there's
we're
just
in
very
early
discussion
phase
here
right
and
it's
based
on
my
understanding
of
where
we
were
going
with
edge
profiles
and
feel
free
to
steer
me
either
way,
right
so
I
say,
and
the
write
up
is
basically
split
into
these
sections,
namely
background
context
of
the
compliance
proposal.
What
is
the
objective?
Some
key
definitions,
elaborating
on
the
scope
of
compliance.
B
What
is
going
to
be
the
compliance
policy
and
then
more
about
the
procedure
and
then
I
had
some
questions
around
you
know.
Of
course
these
can
be
removed,
and
these
are
just
questions
for
discussion.
I
had
which
I
wasn't
clear
about.
You
know
where
we
going
with
the
profiles:
okay,
okay,
so
go
ahead.
Yeah.
A
I
just
wanted
to
say
that
I
think
the
questions
are
good
and
maybe
tomorrow
the
workshop
is
the
breakout
session.
Is
the
right
mechanism
to
get
feedback.
I
do
hope
that
in
the
breakout
session,
because
it's
the
workshop,
we
will
have
much
more
attendees,
so
we
can
heal
the
opinion
of
other
companies.
Not
just
you
know
four
of
us,
but
we
can
get
feedback
from
additional
companies
makes
complete
sense
yep,
so
that
that's
why
I
believe
we
shouldn't
try
to
close
these
questions
now.
We
should
probably
just
show
them
make.
B
Sense,
I
would
still
get
some
advice
from
you,
but
if
the
questions
are
good
for
that
workshop,
then
we
can
bring
it
and
bring
them
over.
But
if
you
think
that
these
are
already
addressed
in
the
material
and
we
ever
going
to
already
have
that,
then
we
don't
need
to
bring
it.
Maybe
I
have
lack
of
understanding
of
the
proposals,
also,
okay,
okay,
all
right.
So
let
me
start
off
by
some
of
the
background.
I
wrote
up
here
says:
each
profile
is
a
mechanism
to
express
a
configuration
needed
to
run
a
particular
machine
learning
model.
B
A
B
A
A
So
the
way
that
think
me
and
mingun
in
perceived
it
originally
was
that
when
we
describe
a
profile,
this
profile
has
a
set
of
attributes.
It's.
It
is
guaranteed
to
pass
a
set
of
onyx
models.
Those
holdings
also
it's
guaranteed
to
pass
a
set
of
onyx
modems
successfully,
and
then
the
vendor
of
this
device
can
claim
that
he
is
compliant
to
some
offer.
A
The
way
that
I
saw
it
Onix
doesn't
right
now.
Standardized
is
an
API
okay,
but
if
there
would
be
a
standard
API
often
on
time,
then
the
runtime
would
declare
a
set
of
profiles
that
it
supports.
In
based
on
this
set
of
profiles,
the
application
could
be
able
to
adjust
a
correct
modern
to
one
on
this
device,
but
since
we
don't
have
an
API
right
now,
onyx
is
more
about
standardizing.
The
container
I
think
the
idea
here
is
to
create
an
infrastructure
or
a
framework
of
description
of
tests
that
this
is
more
of
an
agreement.
A
A
You
are
sharing
the
results,
wait
the
working
group
or
with
the
onyx
entirely
community
and
based
on
that
you're,
getting
compliance
to
your
device
for
that
poor
it.
Let's
take
a
small
speaker,
for
example,
so
we
have
a
small
speaker.
You
one
test
of
boffin
a
on
the
small
speaker,
get
back
the
results
and
hand
them
over
to
the
walking
up
and
then
you're
smart
speaker
is
compliant
to
profile
a
so.
A
The
working
group
is
giving
you
the
proven
for
it
and
then
based
on
this
approval,
you
know
you
can
publish
it
on
your
documentation
and
everything
related
to
it.
Okay,
if
we,
if
we
would
have
a
standardized
API
I,
would
say
that
there
is
a
standardized
way
to
query
which
profiles
specifically
are
supported,
but
we
don't
have
a
standardized
API.
So
the
mechanism,
the
only
mechanism
that
I
say
now-
is
kind
of
an
offline
compliance
process
that
you're
getting
certification
and
appreciate
in
your
documentation
make
sense
make
sense.
So.
B
B
B
B
B
B
Then
now
it
is
certified
to
run
as
a
smart
speaker
now
that
model
comes
on
to
a
platform
for
execution.
Now
the
platform
needs
to
say
that
okay
I'm
ready
for
this
profile,
also
right.
So
it
needs
to
know
what
profile
it
needs
to
run
on,
and
it
says.
Ok
here
is
a
model
file.
This
was
the
intention
of
the
producer.
He
wanted
this
model
to
be
a
smart
speaker
and
Here
I
am
going
to
allocate
and
make
its
execution
environment
as
a
smart
speaker.
I'm
gonna
go.
Do
that
so.
A
B
A
B
B
A
I
think
it's
airing
and
I
say
it's
a
really
nice
idea
that
it
extends
the
meaning
and
I
think
maybe
we
should
start
by
describing,
but
the
original
intent
of
the
porphyrin
was
to
describe
a
set
of
moderns
and
attributes,
and
now
we
want
to
extend
it
into
using
the
same
mechanism
to
define
the
operation
mode
of
the
specific
model
that
now
you're
loading,
okay
and
what
you're
saying
is
because
you
don't
have
an
API
to
it
and
what
we
have
is
the
container.
So
we
can
add
a
fiend
in
the
container.
A
B
You
put
it
right
over
and
here
I've
described
some,
you
know
I,
so
just
ignore
this.
This
is
a
connectivity
is
just
the
same
thing
as
what
we
had
data
locality
right
accuracy,
so
at
runtime
the
model
producer
had
already
published
that
my
model
has
this
accuracy
post
training
right.
He
put
it
in
there,
then
at
runtime.
Now
you
know
when
we
do
the
compliance
we
certified.
We
make
sure
that
these
this
platform
is
ready
to
run
for
this
accuracy.
We
make
sure
this
much
of
memory
is
available.
This
much
of
latency
throughput.
A
Now
I
understand
this
document
in
a
different
way
and
I
think
that
using
the
term
compliance
here
is
kind
of
confusing,
because
we
would
like
to
resolve
the
concept
of
compliance
to
the
original
concept
of.
There
is
a
test
package
and
okay,
you
passed
the
test
package.
There
was
a
procedure,
there's
a
compliance
procedure
which
is
take
these
package
of
models,
run
it
on
your
device
and
give
us
the
results,
and
then
your
compliant,
so
I
would
resolve
this
terminology
today,
original
meaning
and
just
here
I
would
try
to
I.
B
Now
see,
power
is
another
interesting
one
like
you
could
run
it
on
a
low
power
DSP
or
you
could
run
it
on
on
your
course,
which
take
more
power.
So
a
lot
of
those
considerations
and
you
can
only
respect
those
attributes
if,
during
runtime,
all
those
things
are
available
to
you
so
I
guess,
I
need
probably
guidance
of
where
this
should
go
and
if
you
guys
agree,
which
is
the
right
place
to
put
it
in.
B
A
So
now
I
understand
better
the
direction
and
I
think
that
doing
today,
maybe
you
can
send
me
the
updated
version.
Oh
yes,
send
me
the
updated
version
I
will
upload
it
today
to
the
discussion,
material
and
then
I
can
comment
directly.
Maybe
you
don't
and
yourself
we
can
comment
directly
on.
This
sounds
good.
Okay,.
A
B
A
Yeah,
let's
let's
go
over
them,
but
as
I
said,
okay,
so
now
question
one
makes
sense
and
I
would
actually
there
was
going
to
be
perform
is
going
to
be
used
for
compliance
and
it's
going
to
be
used
for
execution
directive
inside,
so
we're
not
going
to
in
a
sense
rename
each,
but
the
same
offer
is
going
to
be
used
for
two
different
meanings.
I
think.
A
B
A
B
Can
bring
that
question
there?
We
can
change
this
to
data
locality
and
whether
we
want
two
fields,
one,
a
binary
and
another
chord
quality
of
service
or
just
binary
yep.
So
you
agree.
We
should
bring
up
that
question
yes
good
so,
and
the
other
question
was
how
many
profiles
can
be
added
to
a
more
particular
model
file.
A
A
So
I
created
this
mom
going
to
describe
what
Megan's
idea
was,
and
the
idea
is
that,
instead
of
having
one
profile
tail,
let's
call
it
a
system.
For
now.
Let's
say
we
have
a
small
speaker
profile
other
than
that
change.
The
way
profiles
are
targeting
and
talk
about
type
of
input
or
sensor
or
whatever
we
want
to
do
it
and
inside
it
has
different
levels
and
then
a
specific
system
or
a
specific
configuration
would
have
a
collection
of
effects.
A
If
I
want
to
map
it
into
what
we
just
discussed,
then
your
operation,
mode
or
execution
mode
will
describe
something
for
the
2d
and
something
for
the
audio
and
something
for
the
any
other
type
of
sensor
that
you
have
in
your
make
sense.
So
that
actually
interacts
with
your
question
and
the
answer
would
be.
There
is
a
connection
because
each
one
of
them
describes
for
a
specific
type
or
kkatalk
classification
of
models
to
some
input
type.
There
is
a
dedicated
set
of
attributes
that
your
sister
need
to
be
adopted.
A
B
Okay,
so
the
next
question
was:
how
do
we
express
subset
of
upsets
in
a
particular
profile,
which
means
you
know
we
have
more
than
120
ops,
mm-hmm
for
a
profile.
Are
we
thinking
that
the
things
like
2d
and
audio
these
will
be
like
published
profiles
on
github
bionics,
just
like
USB
spec
has
USB
profiles.
Similarly,
we
would
predefined
that
this.
A
A
It
yes,
yes,
the
idea
is
that
a
profile
is
excuse
me.
A
profile
is
published
on
onyx
website
and
the
poor
firm
has
fixed
set
of
models
and
the
fixed
set
of
attributes,
so
in
operations
that
each
supports,
so
so,
let's
say
moderns
and
operations
together,
and
this
is
the
list
of
operations
supported
by
this
profile,
and
if
we
want
to
update
it,
we
just
update
the
porphyrin
version.
I
think
we
excuse
me.
A
We
already
discussed
it
on
the
previous
workshop
that
the
profiles
are
kind
of
a
in
in-process
walking
boss's
finger,
so
you
describe
a
poor
farm,
it's
a
2d
intermediate,
you
describe
it,
you
freeze
it
in
time
for
version
1.0,
but
then
half
a
year
afterwards
over
here
afterwards,
you
can
advance
to
version
1.1
and
maybe
increase
the
set
of
operations.
Okay,
so
mapping
it's
what
you
just
described.
B
Perfect,
so
operation
mode
is
something
that
we
don't
need
to.
Publish
though
right
produce
would
be
able
to
say
that
you
know
okay,
this
model
I
say
it
can
run
in
this
memory
or
or
we
can
run
it
in
more
memory.
Also
like
now,
if
you
can,
you
can
spin
out
more
threads
and
you
can
burn
more
power
also
and
have
more
memory
footprint
but
run
faster.
Its
operation
won't
be
able
to
tackle
that,
and
so
we
can
keep
the
two
things
slightly
separate.
Then
right.