►
From YouTube: ONNX Steering Committee Update
Description
ONNX Steering Committee Update from: Prasanth Pulavarthi, Alexandre Eichenberger, Rajeev Nalawadi, Mayank Kaushik, Andreas Fehlner
A
Hey
everyone
really
great
to
see
all
of
you
here
today.
My
name
is
prasanth.
I
work
at
microsoft
on
the
onyx
and
onyx
runtime
team,
I'm
also
on
the
onyx
steering
committee,
and
I
have
four
other
co-members
of
my
steering
of
the
steering
committee
with
me
and
you'll
be
hearing
from
some
of
them
today.
A
A
And
that's
where
the
screen
grab
is
for
from,
but
basically
onyx
was
designed
to
solve
a
number
of
machine
learning
problems
that
different
companies
were
having.
We
were
having
it
at
microsoft
and
then,
when
we
talked
to
folks
in
the
industry,
they
were
all
having
similar
problems
as
well
and
at
the
core
of
it.
It's
about
interoperability
in
multiple
dimensions.
A
So
when
you
boil
it
down,
onyx
really
kind
of
consists
of
two
key
things:
there's
a
common
set
of
operators
that
define
the
building
blocks
for
machine
learning
models
and
then
there's
a
common
file
format
for
exchanging
these
models
between
different
components
like
the
training
systems
and
the
inferencing
systems,
and
this
was
super
useful
for
a
lot
of
different
people
and
for
a
couple
of
reasons,
one
if
you're
a
software.
If
you're
on
the
software
side,
you're
creating
applications
and
services
and
ml
models,
then
onyx
lets
you
kind
of
use
whichever
framework
or
tool.
A
A
So,
okay,
thank
you.
You
know.
Onyx
has
been
pretty
useful,
so
there's
a
lot
of
people
in
the
community
who
are
using
it.
You
know
these
are
just
some
of
the
logos
of
the
companies
that
are
using
it
and
it's
great
to
see
this
awesome
community
that
has
evolved
around
onyx
over
the
past
few
years
by
the
way.
If
your
logo
is
not
on
here
and
you
want
it
to
be,
please
reach
out
and
we'll
get
that
taken
care
of.
A
And
onyx
there's
a
community
around
onyx,
but
the
onyx
development
is
also
done
by
the
community.
Onyx
is
an
open
source
project.
It's
part
of
the
linux
lf
ai
and
data
foundation,
the
linux
foundation,
and
it
has
open
governance.
That
means
there's
no
one
company
that
that
controls
it
right,
and
this
comes
out
in
kind
of
two
ways:
one
there's
a
steering
committee,
that's
elected
every
year.
A
A
Can
you
go
back
there?
We
go.
I'm
really
happy
to
welcome
andreas
felder,
so
he's
a
new
member
on
the
steering
committee
and
I
want
to
thank
wenming
who's
in
the
audience
who
was
on
the
steering
committee
last
year
for
his
service
for
the
all
the
last
whole
year.
So
while
well,
there's
a
steering
committee
that
kind
of
looks
over
broad
things
for
the
for
the
project.
The
real
work
for
onyx
actually
happens
in
the
special
interest
groups
and
the
working
groups.
A
This
is
where
all
the
work,
the
actual
technical
work
gets
done,
and
we
have
a
number
of
sigs
as
we
call
them
and
working
groups
here.
There's
architecture
and
infra
operators,
converters
the
model,
zoom
tutorials
and
then
there's
a
new
pre-processing
working
group.
That's
happening
as
well,
and
this
these
are
led
by
community
members
as
well
and
we've
had
some
new
leaders
join
us
in
this
past
nine
months
or
so
so
you
know
the
ones
with
the
stars:
lacun
thiago,
kevin
jackie.
A
Thank
you
for
stepping
up
and
helping
lead
these
sigs
and
working
groups.
And,
of
course,
thank
you
to
the
folks
who
have
been
already
leading
the
working
groups
for
for
some
time
and
you're
going
to
hear
from
each
of
these
working
groups
on
what
they've
been
doing
over
the
last
few
months
and
the
latest
technical
updates.
But
for
now
I'm
going
to
hand
it
off
to
my
aunt
who's
going
to
give
you
an
update
on
the
latest
progress
with
onyx.
B
B
So
this
chart
shows
some
companies
that
contribute
tools
to
different
categories
of
onyx.
We've
divided
this
into
three
categories,
and
you
know
we
couldn't
fit
everyone
in
there.
So
apologies.
If
we
left
anyone
out
the
three
categories
we
defined
three
categories,
the
first
one
is
tools
that
help
in
the
creation
of
onyx
models
or
manipulation
of
existing
onyx
models.
B
B
Onyx
runtime,
obviously,
is
very
well
known,
supports
several
device
back-ends
and
then
a
lot
of
companies
produce
tools
that
support
their
particular
architectures
and
the
last
one.
The
last
category
is
tools
used
for
visualization
of
onyx
models
and
netron
is
familiar
to
most
people,
but
there
are
other
options
as
well.
So,
okay,
now
I'm
going
to
talk
about
some
statistics
from
engagement
about
engagement
and
usage
of
onyx.
These
are
over
the
last
six
months
since
the
last
community
update
was
made.
B
We
see
good
metrics
across
the
board.
The
number
of
pr's
has
gone
up
by
15
10,
more
contributors,
10,
more
stars,
63,
more
dependent,
repos,
28,
more
forks,
68,
more
mentions
in
research
papers,
which
is
great.
The
number
of
models
in
the
model
zoo
has
gone
up
by
50
over
50,
which
is
awesome
and
a
very
healthy
increase
in
the
number
of
monthly
downloads.
So
these
stats
are
quite
encouraging
to
see.
B
B
The
onyx
model
hub
was
introduced,
which
is
a
tool
that
lets
you
pull
pre-trained
models
from
any
model
zoo,
basically
but
the
onyx
model
zoo
by
default,
and
you
can
point
it
to
a
different
model
zoo
that
follows
the
prescribed
structure.
There
were
a
few
utilities,
other
utilities
introduced
as
well.
The
compose
utility
allows
folks
to
create
combined
models
with
pre-processing
graphs
and
the
pre-processing
networks
and
inference
networks.
B
There
is
a
function
builder
utility
introduced
that
helps
create
in
creating
function.
Ops,
there
were
bug,
fixes,
infra
improvements
and
doc
updates
the.
If
you
go
to
the
release
page
by
following
the
link
here,
that
will
give
you
all
details
of
which
ops
were
introduced,
and
it
will
also
let
you
find
out
more
about
these
utilities
if
you're
interested
and
then
a
few
days
ago,
onyx
1.12
was
introduced
as
well.
That
came
with
onyx
offset
17
with
a
lot
more
new
operators
and
a
few
updated
operators
as
well.
C
So
we
held
like
a
full
full-fledged,
like
request
for
the
features
by
requesting
the
community
to
submit
what
they
would
like
to
see,
and
these
are
long
long
term
intercepts
right
and
some
are
ongoing.
Some
are
finished,
so
you'll
see
in
the
status
that
I'll
go
through.
So
the
request
for
features
happened
in
july
2021
of
last
year
and
then
there
were
different
presenters
with
selected
topics.
C
Our
next
slide,
so
this
is
a
quick
snapshot
of
what
those
requests
were,
for
example,
I'll,
not
go
in
detail,
you
know
to
each
one
of
them,
but
new
operators
for
data
pre-processing.
There
is
a
pre-processing
working
group
and
that
that
is
being
handled
in
the
in
that
particular
working
group.
So
essentially,
there
were
different
operators
that
were
needed,
and
these
are
also
going
over
to
the
operator
seg
for
approval,
so
a
lot
of
them
got
implemented
and
still
the
work
continues
on.
C
Currently,
there
is
an
individual
framework
for
java,
as
well
as
c
sharp
and
then
of
course,
c
plus
plus.
Is
there
then
the
meta
information
and
tensors
like
there
is
more
structured,
quantized
information
that
is
needed
to
flow
through
the
model
and
the
tensors,
and
that
is
also
trying
to
get
a
path
identified
of
how
that
can
be
done
and
on
the
on
the
end-to-end
pipeline
with
onyx
operators.
C
This
is
identified
as
a
long-term
intercept
and
further
refining
of
the
proposal
is
happening
at
this
moment
and
you
can
see
like
the
proposed
things
that
the
requesters
have
been
asked
to
go
and
work
with.
So
if,
in
the
sake
as
chairs,
if
you
see
some
of
these
come
come
as
requests,
these
are
sort
of
being
identified
as
long
term,
not
essentially
like,
as
is
accepted,
but
going
through
the
regular
cycle
of
of
understanding
and
what
needs
to
be
done.
C
For
the
converters,
there
is
some
progress,
some
of
the
higher
functioning
ops
where
have
gotten
support
in
the
converter.
So
thanks
for
that,
and
the
effort
still
continues
as
the
opportunities
rise
next
slide,
please
so
with
different
offsets
and
with
quantization.
Essentially,
there
was
a
gap
like
offset
five
six
seven.
How
do
we
go
and
convert
them
to
offset
13.?
So
that
was
a
request.
C
So
the
past
two
onyx
releases
have
been
handling
that
pretty
well
to
have
support
for
the
transition
of
these
old
operators
into
the
new
post,
quantization
kind
of
era
of
ops
and
the
onyx
model,
zoo
example
for
end-to-end
training
that
is
identified
as
a
long-term
intercept
and
the
next
one.
With
respect
to
federated
learning
for
onyx,
there
is
no
new
onyx
operators
that
are
required
so
exploring
different
solutions
which
are
very
framework
agnostic
over
there
for
the
improvement
to
shape
inference.
C
C
The
next
fall
please
so,
as
we
saw
like
there
is
a
lot
of
a
couple
of
releases
that
have
happened.
1.11
1.12,
that
mayank
went
over
and
also
prashanth
highlighted
like
how
the
community
has
been
growing,
but
we
still
need
a
lot
of
volunteers.
It's
a
community
project,
so
there
is
a
lot
of
prs
that
need
to
be
managed
so
request
again
to
the
broader
community
to
have
more
people
from
your
company
join
this
effort,
and
we
can
take
this
further
and
much
faster
out
into
the
industry.
C
C
So
again,
please
continue
to
stay
engaged
and
your
contributions
to
onyx
and
its
related
projects.
So
there
are
different
resources.
The
website
github,
as
well
as
slack
channels,
slack
channels,
you'll,
see
like
so
many
of
them
related
to
different
topics
and
the
onyx
calendar
also
has
a
meetup
for
our
steering
committee
meetings,
as
well
as
the
other
sick
working
group,
so
that
can
be
picked
up
from
there
and
do
sign
up
for
the
mailing
list
as
well
as
follow
on
twitter.