►
Description
Event: LF AI & Data Day - ONNX Community Meetings, October 21, 2021
Talk Title: Model Zoo / Tutorials SIG Update
Co-Chairs: Wenbing Li & Mark Hamilton (Microsoft)
A
A
A
In
the
last
meeting
up
meetup
that
quantization,
the
mobile
model
is
our
fox
area
in
in
our
next
milestone.
So
and
now
we
have
a
thanks
for
the
some
inter
intel
guys
contribution.
Now
we
have
several
quantized
models,
upload
in
a
model
zoo
like
the
red
net
50
and
vg
and
chef
net,
and
this
is
a
spec
particularly
very
useful
for
the
mobile
apps.
As
you
can
see
the
phone
table,
we
take
the
resnet
50,
as
example,
you
can
see
with
without
too
much
loss
on
the
accuracy.
A
The
footprint
and
the
latency
are
both
greatly
improved
and
those
modes
are
generated
by
the
inter
neural
compressor
and
the
model
can
be
influenced
by
only
runtime
without
any
extra
configuration.
A
Also,
we
have
some
new
kind
of
the
model
which
we
called
out
of
box
model.
This
model
will
fills
all
the
pre
and
post
processing
into
the
model.
So
when
you
do
the
inference,
you
don't
need
extra
step
on
the
pre
and
post
processing.
A
This
will
greatly
simplify
the
model
inference
and
it
is
more
portable
because
you
don't
need
to
worry
about
the
those
pre-processing
libraries
in
any
platform
and
especially
useful
for
the
mobile
apps,
and
in
this
milestone
we
have
checking
the
gbe
tool
you
in
most
of
people.
If
you
are
familiar
with
gb2,
you
always
require
the
some
model
generation
algorithm.
A
typical
is
a
beam
search
to
get
the
final
android
result.
A
A
Then
we
talk
about
a
little
bit
about
the
root
map
in
the
next
half
year.
Firstly,
we
always
welcome
the
model
contributions,
especially
for
this
model
state
of
art
model
and
the
model
for
the
mobile
apps,
because
it
becomes
much
more
and
more
popular
now.
Another
way
that
we
have
a
whole
discussion
about.
How
do
we
deal
with
the
next
operation
in
the
models,
because,
when
on
module,
zoo
was
funded
away
several
years
ago?
At
that
time,
the
offset
water
is
still
very
low
compared
to
the
latest.
A
B
We
hope
that
this
will
enable
people
to
build
flexible
applications
on
top
of
collections
of
onyx
pre-trained
models
that
wouldn't
be
possible
without
a
unified
api
to
connect
them
all.
And
finally,
one
of
the
goals
of
this
work
is
to
maintain
parity
with
other
ecosystems
such
as
pytorch
and
tensorflow,
which
already
contain
model
hub
abstractions.
B
B
We
wanted
to
support
multilingual
clients,
because
onyx
is
not
just
an
ecosystem
and
python,
but
in
a
variety
of
different
languages.
This
drove
us
to
investigate
a
language
agnostic
protocol
layer
called
the
manifest
it
should
support
user
and
private
hosted
hubs
so
that
anyone
can
deploy
their
own
model
hub
and
use
it
within
their
closed
infrastructure,
but
still
be
able
to
leverage
the
common
apis.
B
It
should
be
secure
and
efficient.
You
know
it
should
stop
man
in
the
middle
of
attacks
with
checksums
and
it
should
support
local
caching
so
that
subsequent
calls
to
the
apis
don't
require
redownloading
the
model.
And
finally,
we
wanted
to
make
this
easy
for
the
onyx
team
to
maintain.
So
we
wanted
to
generate
this
manifest
from
the
existing
onyx
models
repository
so
that
this
could
be
done
automatically
and
that
this
collection
of
models
can
be
curated
without
any
human
input.
B
So,
to
quickly
sketch
out
the
main
steps
of
the
onyx
model
hub,
it
all
starts
with
the
onyx
hub
python,
client
that
is
built
and
at
the
onyx
onyx
repository
users
install
that
on
their
machines
and
then
can
type
onyx.hub.load
and
pass
in
the
name
of
a
model
along
with
additional
parameters
by
default.
This
will
point
to
the
main
onyx
models
repository
where
all
these
different
pre-trained
models
are
currently
hosted
and,
in
particular,
we'll
point
to
a
manifest
file.
That's
stored
in
that
onyx
model
repository.
B
B
If
you
wanted
a
particular
version,
you
can
pass
in
the
offset
keyword
and
if
you
wanted
to
go
even
further
and
pin
the
entire
thing
to
a
particular
fixed
hash
of
the
onyx
model
repositories
for,
say,
reproducibility,
that's
possible
as
well,
and
you
can
also
download
from
user
repositories
enabling
private
or
custom
style
deployments
in
addition
to
downloading
models.
We
also
contain
some
utility
functions
for
say,
setting
the
cache
directory
for
inspecting
the
information
of
these
models
prior
to
downloading
and
even
querying
the
models
by
semantic
tags
like
vision
or
detection.
B
A
lot
of
the
functionality
of
the
hub
relies
on
a
manifest
which
contains
metadata
and
locations
of
trained
models,
and
in
particular,
this
manifest
is
not
hard
to
create.
We
were
able
to
automatically
generate
it
using
the
existing
markdown
tables
on
onyx
models,
and
this
allows
us
to
suck
in
over
120
different
models
into
our
initial
version
of
the
hub,
with
metadata
for
check
summing
and
inspecting
the
input
output
structure
of
these
models
and
a
variety
of
other
features
in
this
work.
B
One
of
our
big
next
steps
is
to
create
a
jvm
based
client
for
the
onyx
model
hub.
We
currently
have
implemented
a
python
client,
but
the
jvm
based
client
will
allow
us
to
make
a
really
nice
interface
for
our
new
apache
spark
based
distributed
onyx
inference
code,
which
allows
you
to
evaluate
deep,
deep
learning
models
within
apache
spark
clusters
very
easily,
and
simply
and
second,
we
would
love
to
incorporate
some
responsible
ai
information
into
the
ontology
and
the
metrics.
B
If
you
have
any
feedback
about
the
project
or
want
to
kind
of
continue
moving
it
forward
and
thanks,
as
always
to
the
the
onyx
steering
committee
and
everyone
who
helped
along
the
way
to
provide
us
feedback,
help
us
navigate
the
build
system
and
help
us
get
this
contribution
into
the
main
onyx
repository.
So
thank
you
all.