►
From YouTube: wasmCloud Working Group - Machine Learning 04/14/22
Description
wasmCloud is a platform for writing portable business logic that can run anywhere from the edge to the cloud, that boasts a secure-by-default, boilerplate-free developer experience with rapid feedback loop.
https://wasmcloud.com
A
Welcome
to
the
wasm
club
machine
learning,
weekly
bi-weekly
call.
This
is
april
the
14th
2022.,
so
kristoff
andrew
steve.
I
mean,
I
guess
we
should
probably
start
with
the
state
of
the
demo.
Kristoff
you've
been
making
a
ton
of
progress,
and
we've
got
now
pre-processing
and
post-processing,
as
well
as
the
sort
of
core
work
you're
working
and
it's
working
with
a
bunch
of
different
models.
Steve
you
did
a
demo
on
the
wasm
cloud
call
on
wednesday.
A
Just
yesterday
is
there
anything
else
that
we
should
review
as
far
as
what
the
state
of
the
demo
is
now
and
where
things
are.
B
One
of
the
there's
some
minor
tweaks,
but
it,
but
I
think
those
are
probably
lower
priority
than
moving
ahead
with
the
next
features
for
the
demo.
I
think
so.
Tweaks
we
can
clean
up
the
output.
B
Another
tweak
is
doing
performance,
benchmarks
and
figuring
out
places
to
optimize,
but
I
don't,
I
think
it's
a
little
too
early
to
do.
Try
to
pay
a
lot
of
effort
into
optimization.
B
We
could
do
a
web-based
demo,
but
that
sort
of
goes
in
line
with
features
of
the
demo.
So
so
I
think
it's
pretty
good,
as
is,
and
until
we
decide
on
the
next
steps
forward,
we'll
have
to
hear
what
other
people
think.
A
Christoph
what
do
you?
What
do
you
think
yeah
you've
started
to
socialize
this
bit
with
some
broader
communities?
What
was
their
response
or
where
or
what
was
your
perspective
on
that.
C
A
That's
great
andrew,
I
think
you
kind
of
asked
an
interesting
question
around
like
what
does
this
demo
mean
for
us
in
its
current
state,
and
so
I
think
what
we
have
now
is.
A
The
second
thing
I
think
that
distinguishes
this
demo
andrew,
is
that
the
ability
to
mount
the
capability
providers,
whether
they're,
local
or
remote,
so
you
could
acquire
an
image
from
a
local
device
but
mount
a
capability
provider
that's
remotely
and
in
another
way
how
you
would
do
this
is
you
know
you
would
create
some
client
server
structure?
A
You
know
just
submit
it
via
a
rest
call
or
something
like
that,
but
I
think
this
way
is
incredibly
simple,
because
the
developers
don't
need
to
change
any
of
their
code
in
order
to
execute
you
know,
design
a
process
for
consuming
machine
learning
and
then
implement
it.
You
know
deployment,
topology
becomes
a
detail,
you
know
becomes
a
deployment
detail,
it
doesn't
become
an
architectural
criteria
that
I
have
to
bring
into
my
code.
A
So
that's
the
the
second
thing
that
I
think
that
it
means,
and
then
I
think,
there's
a
broader
third
thing
here
around
you
know.
The
connectivity
of
this
in
you
know
like
this
is
now
so
easy
to
take
a
set
of
code
and
because
it's
already
tied
into
all
of
the
other
ones.
I
think
that
this
really
opens
up.
You
know
an
architecture,
that's
very
popular,
you
think,
like
nest,
type
architecture
where
you've
got
a
lot
of
edges
and
end
points.
This
just
makes
building
those
things
very
easy.
A
So
I
think
the
combination
of
distributed
with
wasmcloud
plus
machine
learning
does
unlock
a
few
powerful
demos.
Are
there
other
things
that
this
demo
means
today
that
that
I
didn't
accurately
account
for
the
my
list?
Was
pre
post
processing
all
this
stuff
around
the
machine
learning
is
now
has
portability
the
ability
to
locally
or
remotely
mount
capability
providers
and
the
sort
of
architectural
collaboration?
That's
unlocks?
Is
there
anything
else
that
you
guys
would
highlight.
D
Sounds
about
right
to
me,
can
I
ask,
can
I
rewind
to
ask
a
couple
questions
about
the
original
not
to
derail
too
much
here,
but
with
the
pre-processing
that
you
guys
have
built
so
far?
D
D
B
Okay,
I've
tried,
with
I've,
tried
with
about
15
different,
varied
images
from
the
web,
jpeg
and
png,
and
it's
worked
what
looked
fine
great
on
all
of
them,
so
I
I
would,
I
would
add,
a
caveat
to
liam's
idea,
liam's
statement
about
where
we
are
in
terms
of
architecture.
B
I
think
this
architecture
is
still
in
the
prototype
stage
and
even
the
block
diagram
that
I
showed
in
the
community
call
yesterday.
I
think
that
could
change
once
we
get
further
along
and
into
a
demo,
that's
more
realistic
and
more
full
featured
or
plugged
into
a
real
world
application.
I
I
don't
think
we
would
want
to
declare
that
we've
decided
on
the
right
architecture
entirely
and,
and
people
should
just
go
and
run
with
this
now.
I
think
it
needs
a
little
more.
B
It
needs
some
more
proving
out
in
real
world
and
that's
partly
related
to
performance
issues.
I
know
there's
some
some
things
that
kevin
and
I
have
been
talking
about
about
response
time
that
will
probably
rear
their
head
when
the
models
get
a
little
more
complicated
and
their
response
time
increases.
B
So
there's
a
few
subtle
variations.
I
think
we're
going
in
the
right
direction.
I
don't
think
we've
made
any
I'm
not
aware
of
any
completely
wrong
turns,
but
we
might
make
a
30
course
correction
in
the
architecture
once
we
get
more
experience
with
it.
A
A
B
Yeah,
so
I
guess
the
term
architecture
is
a
little
ambiguous.
So
what
I
was
what
I
was
thinking
when
so
nothing
you
just
said
is
wrong.
What
I
was
thinking
about
was
the
topology
of
so
the
the
way.
The
way
the
message
passing
works.
Now
we
we
send
the
image
to
the
api
actor.
B
B
So
some
of
those
kind
of
things
might
change,
and
those
are
the
kind
of
things
that
if
we
were
to
tell
people
10
developers
to
go
off
and
build
your
application
with
this,
the
variations
in
those
would
cause
a
lot
of
duplicate
work
and,
and
maybe
some
frustration,
and
so
those
are
things
where
it's
not
just
that.
B
I
want
to
test
with
bigger
models
or
more
complex
models,
but
I
also
want
to
test
in
some
real
world
applications,
ideally
in
more
than
one
domain,
so
image
plus
something
else,
but
getting
that
real-world
application
in
there
and
all
the
other
things
that
that's
going
to
require
for
the
flow
to
work.
I
think
is
important
before
we
say
that
we
have.
C
You
open
for
a
feature
request.
Maybe
this
is
a
good
time
to
introduce
again
what
I
maybe
once
called
a
low
latency
mechanism.
So
I
dream
of
of
something
where
maybe
there's
an
intermediate
actors,
something
which
intermediates
between
our
today's
capability
provider,
which
does
the
inferencing
and
the
new
capability
provider,
which
represents
a
data
source
like
regardless
of
if
it
be
a
camera
or
something
else.
B
Yeah,
so
that
so
that
work
is
that
work
is
in
progress
so
for
for
the
other
folks
on
the
call.
B
We
still
have
a
default
two
second
response
time
anywhere
in
the
chain
for
for
any
rpc
message,
and
that
includes
from
the
http
server
to
the
first
api
actor,
because
then
all
of
the
other
things
that
happened,
the
pre-processing,
the
inference
the
post-processing,
those
all
together,
cumulatively,
have
to
complete,
within
two
seconds
before
the
the
api
actor
returns
to
the
http
server.
And
if
it
takes
longer
than
two
seconds
it'll
get
a
timeout.
B
You
can
increase
that
timeout
with
a
system-wide
configuration,
but
what
kevin
and
I
have
talked
about
a
way
to
send
fire
and
forget
messages
from
actors
and
that'll
that'll,
allow
different
kinds
of
technologies
and
it'll
allow
paths
that
take
longer
than
two
seconds.
So
that's
that's
in
progress.
We
should
see
that
that
change
soon,
probably
within
a
couple
of
weeks,
but
that's
separate,
and
so
that
might
actually
influence
the
topology
that
we
end
up
recommending
and
there's
also
some
other
things.
B
I
know
andrew
had
brought
up
the
idea
of
trying
to
get
zero
copy
in
the
tensors
that,
if,
as
we
dive
into
that,
so
I
put
that
under
the
category
of
performance,
optimization
as
we
get
into
that
the
api
will
probably
change
a
little
bit.
So
that's
another
thing
that
we'd
like
to
get
a
little
more
practice
with
before
we
before
we
tell
a
bunch
of
people
to
go
run
with
it.
A
No,
no
so
kristoff.
What
I
hear
from
steve
is
before
we
go
down
the
zero
copy
approach
that
we
should.
Maybe
try
to
you
know
steve.
I
don't
know
if
you
have
a
specific
proposal
here,
like
you
know,
do
you
think
we
should
spin
up
a
couple
models
ourselves
and
just
run
them?
You
know
like
point:
a
camera
out
my
window
run
the
dog
recognition
model
and
figure
out.
What's
the
most
popular
you
know,
dog
walked
in
the
suburb
of
washington
dc.
A
You
know
what
what
do
you?
What
do
you
suggest,
or
we
try
to
find
and
recruit
a
couple
folks
that
are
building
things
like
kristoff?
Maybe
you
have
some
concepts
that
we
could
lean
in
and
help
out
to
get
up
and
going
to
try
to
generate
some
some
data
and
some
experience
with
this.
C
C
There
there's
a
lot
of
potential
benchmarks
out
there.
I
would
just
grab
some
existing
stuff,
for
example,
the
other
week
I
switched
on
for
the
first
time
that
rather
new
coral
dev
board
and
run
the
demos
which
are
in
that
which
are
shipped
with
that
board,
and
there
are
some
nice
things
and
many
of
them
deploy
also
that
mobile
net
model.
So
one
of
these
we
also
have
here,
we
could
more
or
less
easily
compare
the
results.
Maybe
that's
a
good
start.
A
So
so
you
want
to
use
some
of
the
existing
benchmarks
that
are
already
shipped
with
the
coral
we
set
them
up
in
blossom
cloud
and
we
see
what
the
difference
is
running
them
through
there.
What's
the
test,
harness
that
the
coral
has
is
it
just
like
some
python
python
scaffolding
or
anything
like
that.
C
Yeah,
it's
c
plus
plus,
and
it
has
a
python
wrapper,
but
the
core
is
c
plus
plus
so
of
course,
tensorflow
lite,
that's
a
c
plus
library.
C
C
D
Yeah
that'd
be
pretty
interesting
to
see,
see
what
that
looks
like
in
terms
of
your
question
on
the
slack
channel.
Are
you
envisioning
that
at
the
wasm
cloud
level,
users
will
be
able
to
see
the
different
hardware?
That's
performing
the
inference
right
or
like
with
the
fact
that
the
inference
is
being
performed
on
a
tpu,
be
exposed
at
the
wasm
cloud
level
or
not.
B
It
should
be
transparent,
but
we
could
we
could
expose
it
in
terms
of
adding
information
to
the
response.
If
you
want
to
use
that
information
for
routing
the
request,
we
could
also
try
to
come
up
with
something.
That's
somewhat
abstract
like
like.
Maybe
a
choice
of
optimize
for
speed
or
cost,
or
something
like
that
and
then
have
some
router
decide.
B
We
we
try
to
keep
architecture
specific
stuff
out
of
the
messages,
but
if
there's
any
kind
of
some
goal
parameters
that
might
make
sense
to
add
to
the
request,
we
could
figure
those
out.
D
I
think
actually
I
was
leaning
the
other
way
towards
not
exposing
that
type
of
information
to
at
the
wasm
cloud
level
that
you
know
I
was
just
checking
to
see
and
it
sounds
like
what
the
proposed
path
would
be.
Well,
the
capability
provider
just
runs
either
on
the
on
the
coral
dev
board
or
on
some
other
machine.
D
Wasmcloud
really
doesn't
know
the
capabilities
of
that
machine,
but
the
capability
provider
itself
does
and
it
can
choose,
it
can
say.
Oh
I
see
I
have
a
tpu,
I'm
going
to
use
the
tpu.
Oh,
I
don't
have
a
tpu
over
here.
I'm
not
gonna,
I'm
gonna
use
the
cpu
right
is
that
sort
of
the
model
we're
looking
at
here,
yeah.
B
Yeah,
that's
correct,
so
the
provider
could
decide
which
which
of
the
available
resources
on
the
current
host
where
it's
running
it
wants
to
use.
In
addition,
there
is
some
there
is
some
network
smarts
built
into
built
into
wasm
cloud,
and
so,
if
your
edge
tpu
is
closer
like
connected
to
the
same
nas
leaf
node,
then
all
requests
will
go
to
that
by
default.
B
But
if
the
edge
tpu
is
not
on
the
network,
but
there
is
something
in
the
cloud
is:
if
the
same
capability
provider
is
running
in
a
cloud
server,
then
it
will
route
to
that.
So
we'll
get
the
kind
of
failover
scenario
that
you
and
minku
were
talking
about.
Yeah.
D
A
Okay,
so
I'm
kind
of
feeling,
like
we
have
a
consensus
here
on
a
path
forward
here
for
next
steps.
It
sounds
like
the
the
demo
kristoff
that
we're
going
to
do
is
we're
going
to
maybe
focus
on
the
coral
dev
board,
steve.
A
If
you
don't
have
one,
I
can
send
you
one
if
I
didn't
already
and
we're
going
to
pull
some
of
the
sample
examples
that
they
have
and
we're
going
to
try
to
configure
the
same
examples
in
wasmcloud
locally
and
we're
going
to
just
compare
them
side
by
side
and
use
that
as
a
baseline
for
how
we
think
we're
doing,
and
then
we'll
use
the
results
from
that.
In
order
to
understand
what
we
think
the
next
direction
is
for
the
demo
kristoff.
A
I
guess
you
didn't
hear
that
you
may
have
stepped
away
for
just
a
minute.
I
said
it
sounded
like
we
have
some
consensus
that
we're
going
to
try
to
recreate
our
clone.
Some
of
the
default
coral
dev
board
examples
using
the
tpu
we're
going
to
configure
them
to
wasm
cloud
and
just
run
them
side
by
side
and
see
just
what
the
initial
pass
at
the
differences
between
the
two
and
then
we'll
use
the
output
of
that
to
understand
where
we
need
to
pursue
and
where
we
invest.
C
That's
a
very
good
idea,
I'm
very
curious
not
only
to
run
models
on
the
edge
dpu,
but
also
on
that
these
arm
cores,
because
I
read
that
the
inference
engine
we
use
that
track.
That
shall
be
optimized
for
armed
course,
because
that
guys
who
wrote
it
they
had
arm
in
mind.
So
I'm
really
you
can
on
the
dev
pod,
also
run
the
inference
on
the
cpu
right.
So
I
I'm
really
interested
to
to
compare.
A
All
right,
steve
andrew,
are
we
all
in
agreement
here
on
what
our
next
steps
are.
D
A
Well,
andrew,
I
think
we're
heading
towards
this
is
you
know,
sort
of
gives
us
the
kind
of
like
test
for
your
use
case,
which
is.
C
C
A
Routing
as
needed,
do
you
do
you
feel
this
is
a
prereq
that
this
will
get
us
one
step
closer
to
the
demo
that
you
that
you
want
to
see
done
with
this.
D
Yeah,
I
think
so
because
that's
a
couple
weeks
ago
or
a
month
ago,
I
know
I
was
talking
about.
We
need
the
data
right,
we
need.
We
need
to
understand
what
is
the
performance
at
the
edge
and
what
is
the
performance
in
the
cloud
and,
as
you
guys
are
running
these
experiments,
if,
if
there's
a
way
to
capture
that
in
a
wasm
cloud
accessible
way,
then
that
can
inform
the
qos
attributes
that
we
were
talking
about
earlier.
A
I
like
what
we're
thinking
about
here
like
the
qos
of
ml,
like
yeah
yeah,
there's,
there's
something
there
like
it's.
D
D
A
C
A
A
Adaptive
ml
would
be
super
super
extra
credit
right
because
then
we
would
get
to
call
it.
Aml
and
we'd
have
a
conflict
with
anti-money
laundering
right
on
the
appreciation
just
always
attack
when
you
could,
like
you
know,
just
muddle
the
bottle,
the
communication
even
more
all
right.
I
love
this.
This
is
great.
Okay,
so
do
we
need
to
talk
about
who's,
gonna
be
steve
and
kristoff?
The
two
of
you
seem
to
be,
like
you
know,
doing
a
great
job
managing
the
work
now
between
the
two
of
you.
A
Do
we
need
to
talk
about
any
of
that,
or
do
you
guys
want
to
rely
on
anything
or
do
you
guys
just
gonna
hand
wave
and
just
continue
to
make
magic
happen?
Sort
of
like
when
the
free
godmother
comes
down
and
just
waves,
your
wand
and,
like
the
pumpkins,
get
up
and
dance?
That's
what
I
feel
like
happens
here.
You
guys
just
it's
amazing,
yeah.
B
Yeah
christoph
christoph,
and
I
can
follow
up
on
slack-
to
fine-tune.