►
Description
Join the .NET Machine Learning team as they kick off the Machine Learning Community Standup series!
Community links: https://www.theurlist.com/mldotnet-standup-2020-08-26
A
A
Hi
everyone
we
are
now
live
and
welcome
to
the
first
machine
learning
for
net
community
stand
up.
I
am
bree,
I'm
a
pm
on
the
dotnet
team
working
on
ml.net
and
we're
starting.
This
series
super
excited
to
talk
to
you
about
machine
learning
with
net
and
ml.net
and
all
the
community
contributions
and
all
these
things
that
are
coming
up
with
the
framework
and
tooling
so
I'll
hand
it
over
to
jake
and
louise
to
introduce
themselves
as
well.
A
B
C
A
Awesome
so
yeah.
So
since
this
is
our
first,
our
first
introduction
our
hello
world.
I
will
talk
to
you
all
a
little
bit
about
what
is
machine
learning
very
quick,
five-minute.
Intro
talk
to
you
a
little
bit
about
the
ml.net
framework,
then
we'll
start
digging
into
some
community
links,
and
then
we
actually
have
a
super
passionate
community
member
on
who's,
going
to
talk
about
his
project
mlops.net
so
bear
with
me.
This
is
my
first
time
working
all
these
tools
for
streaming
to
you
all.
So
let
me
get
started
here.
A
Awesome
so
just
a
few
slides,
I
promise
this
is
my
five
minute:
what
is
machine
learning,
and
hopefully
by
the
end
of
it,
you
are
all
experts.
No
I'm
just
kidding.
You
all
understand
that
the
basics,
if
you
don't
know
them
already
and
understand
all
the
different
things
that
you
can
do
with
machine
learning.
A
So
first
we'll
start
off
with
artificial
intelligence.
So
this
is
the
ability
of
a
computer
to
perform
tasks
normally
associated
with
humans
or
intelligent
beings.
So
this
can
be
things
like
reasoning,
generalizing
and
usually
this
starts
as
a
rule
or
a
logic
based
system.
So
you
think
a
lot
of
if
statements,
and
so
if
I
go
to
this
next
example-
a
problem
to
solve
if
you're
looking
at,
is
this
bread
or
not
bread?
A
If
you're
using
artificial
intelligence
to
solve
this,
you
might
have
a
bunch
of
different
if
statements,
so
you
know
if
it
has
ears,
then
it's
obviously
not
bread.
If
it
has
eyes
it's
not
red,
but
you
can
see
that
this
can
be
really
really
hard
to
scale
because
there's
a
lot
of
different
features
that
can
make
something
bright
or
not
bread.
A
It's
getting
computers
to
make
predictions
without
actually
explicitly
programming
them
so
without
having
all
these.
If
statements,
so
what
happens
here
is
computers
will
find
patterns
in
past
data?
Learn
from
this
experience
and
then
use
their
past
experience
to
act
on
new
data
and
make
predictions
so
again,
it's
used
to
solve
problems
that
might
be
difficult
to
solve,
with
traditional
rules
based
ai
and
as
an
example
with
the
bread
or
not
bread
instead
of
having
all
these.
A
If
statements,
what
you
would
actually
have
is
a
bunch
of
examples
of
bread
and
a
bunch
of
examples
of
not
bread
and
you'll
use
an
algorithm
to
train
a
model
and
then
use
these
this
model
to
use
its
past
experience
to
predict.
If
this
new
thing
is
bread
or
not
bread,
another
way
to
look
at
it.
If
you're
doing
looking
at
artificial
intelligence,
the
input
is
going
to
be
the
rules
like
those.
A
If
statements
and
the
data
and
what
you're
going
to
get
from
that
is
answers,
whereas
with
machine
learning,
you'll
actually
give
it
data
and
answers
and
then
the
output
would
be
the
rules
which
you
can
use
for
new
data
and
then
there's
deep
learning.
So
deep
learning
is
a
subset
of
machine
learning
and
imitates
the
way
the
human
brain
learns
and
thinks
processes
data.
This
is
used
for
more
complex
tasks
like
image,
classification
or
object,
detection
or
natural
language
processing,
and
how
this
all
fits
together.
You've
got
artificial
intelligence.
A
A
You
can
predict
the
price
of
a
house
based
on
house
features
that
would
be
a
regression
example
or
predicting
a
number
or
predicting
sales
based
on
past
sales
recommending
products,
users,
so
there's
a
whole
bunch
of
different
things.
You
can
do
with
machine
learning
so
from
here
I'll
start
talking
about
ml.net
and
I'm
actually
going
to
show.
A
This
is
the
main
landing
page
for
ml.net
and
we
will
hand
out
all
the
links
that
we
have
up
here.
All
of
our
community
links
as
well
at
the
end
or
maybe
after
we
show
all
the
links,
we
can
put
paste
it
into
the
chat,
but
what
we
have
here
ml.net.
So
what
this
is?
It's
an
open
source
and
cross-platform
machine
learning
framework
for
net
developers.
A
A
You
can
stay
in
the
net
ecosystem
and
use
the
tools
that
you're
familiar
with,
and
also
we
also
emphasize
it's
for
developers,
or
at
least
you
know,
we
target
developers
because
we
know
that
a
lot
of
developers
don't
have
data
science
experience
and
we
want
to
make
it
as
easy
as
possible
for
them
to
be
able
to
add
machine
learning
to
their
dot-net
applications,
so
the
tools
that
we
have
enable
enable
developers.net
developers
to
to
easily
do
this.
So
you
can
see
we
have.
A
You
know
a
few
of
our
value
props
here,
a
little
bit
of
the
history,
so
ml.net
actually
started
as
a
microsoft
research
project,
I
think
about
10
years
ago,
and
it
was
an
internal
framework
called
the
learning
code
that
was
used
by
a
ton
of
different
machine
learning
features
and
popular
microsoft
products.
So
that
would
be
power.
It's
using
power,
bi,
microsoft,
defender
being
ad
predictions
windows,
hello.
So
it's
been
used
for
a
really
really
long
time
and
then
in
2018
it
was
actually
released
at
build.
A
It
was
open
source,
may
crop
cross
platform
and
made
a
little
bit
friendlier
for
external
usage
and
it
was
renamed
to
ml.net.
So
the
first
preview
release
was
2018
of
build
and
then
last
year
2019
build.
It
was
actually
the
first
ga
release
and
also
our
tooling
releases
so
I'll
scroll
down
here,
and
you
can
see
here's
some
of
the
examples
of
things
that
you
can
do
with
ml.net
again
this
these
kind
of
map,
to
the
the
slide
that
I
showed
with.
A
What
can
you
do
with
machine
learning
and
there's
a
whole
bunch
of
different
things?
These
are
just
some
of
the
examples
and
actually,
if
you
follow
this
link
here,
it
takes
you
to
our
samples,
repo
and
actually
have
it
loaded
right
here
and
there's
a
bunch
of
different
samples
that
you
can
use
to
get
started.
Another
great
way
to
get
started
is
our
tooling,
which
louise
and
jake
will
actually
show
in
a
little
bit
so
yeah,
that's
the
very
basics.
A
All
right,
so
now
that
we
have
the
did
we
have
did
I
heard
something:
nope,
okay,
cool.
So
now
that
we
have
these-
and
maybe
I
forgot
to
zoom
in
so
you
all
can
see
but
we'll
again
be
handing
out
the
links
after
let's
jump
into
some
community
links
to
show
what
people
have
been
doing
with
ml.net,
so
actually
from
our
machine
learning
samples,
repo
we
actually
have.
A
If
you
scroll,
let
me
zoom
out
if
you
scroll
down
to
the
bottom,
we
actually
have
some
community
samples,
so
these
are
ones
that
are
maintained
by
our
team
and,
of
course,
we
all
always
love
having
community
contributions
to
these,
but
we
have
actually
community
samples
which
are
created
and
maintained
by
community
members.
So
I
will
show
this
so
here
are
just
some
of
them
that
we
have.
One
of
my
favorite
ones
is.
Let
me
see
if
I
can
find
it
here.
Is
this
ml.net
and
visual
desktop
uwp
app?
A
It's
because
it's
really
easy
to
see
all
the
different
applications,
and
this
is
by
github
aliases
sample
brewer,
and
I
believe
it's,
oh
god,
does
anyone
know
how
to
pronounce
dietrich?
Maybe
I
feel
like
that's
right
I'll
I'll
confirm
that
later,
but
I
will
show
so
you
can
see
there's
a
bunch
of
different
examples
here,
where
they
implement
ml.net
into
a
uwp
app
and
make
it
really
really
visual.
A
So
you
can
see
what's
happening
and
they
have
a
bunch
of
different
examples:
clustering
multi-class
classification,
binary
classification,
so
you
can
see
that
they
go
on
and
even
some
of
the
features
that
ml.net
has
they.
They
showcase
in
this
very
visual
way.
So
that's
one
of
my
favorite
community
samples,
another
one
which
is
up
at
the
top
here.
A
Is
this
coven
19
exploratory
data
analysis,
so
this
is
actually
praveen
who
presented
at
the
first
virtual
ml.net
community
conference,
which
we
had
earlier
this
year
and
he
added
it
as
a
sample.
So
he
does
some
data
analysis
with
jupiter,
notebooks
and
net
and
then
actually
uses
ml.net
to
make
some
some
predictions
so
yeah.
If
you
go
to
this
page,
you
can
see
some
more
of
the
community
contributions
on
the
samples
repo
and
so
then
I
move
on
to
youtube.
My
links
are
a
little
slow,
so
this
is
actually.
C
Oh
sorry,
I
was
just
saying
that
one
of
the
takeaways
right
is
that
there's
multiple
ways
to
contribute
to
these
samples.
We
certainly
will
welcome
them.
If
you
already
have
an
existing
sample,
you
can
either
submit
and
add
it
to
the
list
of
existing
net.
You
know
machine
learning,
samples
or
you
can,
of
course,
just
you
know,
provide
the
link
and
we'd
be
happy
to
add
it
to
this
community
samples
sort
of
collection
here,
yeah.
A
A
That
ranges
from
how
do
I
start
with
ml.net
to
more
end-to-end
scenarios
or
more
advanced
scenarios
and
he's
always
adding
new
ones,
so
he
I'm
sure
we'll
have
him
on
the
of
the
show.
At
some
point
he
was
one
of
he
was
one
of
the
organizers
of
the
community
conference
that
we
had
earlier
this
year.
So
definitely
encourage
you
to
check
this
out
if
you're
just
learning
or
have
a
specific
thing
that
you
want
to
learn
with
ml.net
and
he's
on
twitter
as
well.
So
I'm
sure
if
you
had
a
specific.
A
A
So
we're
starting
to
actually
see
a
lot
more
youtube,
videos
and
and
tutorials,
and
lessons
online
about
ml.net,
which
is
really
really
awesome
to
see.
So
you
can
see
here,
we've
got
kaushik
roy
chowdhury,
who
created
this
introduction
to
ml.net
and
again
has
a
ton
of
different
ml.net
videos
here
and
it
looks
like
he
creates
some
udemy
courses,
so
you
may
see
an
ml.net
once
soon
in
the
future.
A
If
there's
not
one
already
and
then
another
example
which
louise
actually
found
was
they
used
ml.net
to
predict
how
nice
are
your
discord
users,
specifically
it's
a
bot
that
looks
at
messages
on
discord
and
it
sends
those
messages
to
the
machine
learning
model
and
will
predict
if
it's
nice
or
rude,
which
I
think
is
really
cool,
sentiment,
analysis
or
predicting
the
sentiment
of
the
comment.
If
it's
positive
or
negative,
it's
actually
really
popular,
especially
when
showcasing
machine
learning.
C
A
A
So
then,
then,
we've
got
alexander,
who
is
actually
on
the
show
he'll
be
on
later
today,
he's
made
a
ton
of
different
contributions
to
ml.net,
including
his
mlops.net
project,
which
he'll
talk
about
later,
but
he's
also
just
started
on
this
ml.net
templates
project,
which
creates
new
ml.net
templates
using
the.net
new
command.
So
this
is
really
really
cool.
I
believe
he
just
started
this
yeah
last
week,
so
super
excited
to
see
how
that
develops
and
see,
if
you
know
eventually
integrating
it
into
vs
as
well,
would
be
really
really
awesome.
A
This
is
actually
a
showcase
of
our
customers
and
how
they're
using
ml.net.
So
I've
actually
written
these
stories
so
that
they're,
you
know
pretty
short
and
you
can
kind
of
get
the
business
problem,
what
they're
trying
to
solve
why
they
chose
ml.net
and
if
they
were
wanted
to
share
what
their
architecture
solution
architecture
looked
like.
So,
let's
go
into
scan
cam.
A
For
instance,
they
are
an
australian
company
who
wanted
to
solve
the
problem
of
fuel
theft
in
australia
and
they
are
actually
using
ml.net
with
so
for
object,
detection
to
detect
cars
and
images
and
then
the
license
plate
and
then
search
as
a
database
of
previous
offenders.
And
you
can
see
here
they
they
talk
about,
why
they
chose
ml.net
and
of
course,
the
main
reason
we
usually
get
is
well
or
netshop
or
we're
all.net.
A
A
So
if
you
want
to
check
out
some
real
life
examples-
and
I'm
always
adding
or
trying
to
always
add
new
stories
to
here
as
well
or
if
you
even
have
a
story
yourself
and
want
me
to
add
it
to
the
showcase
reach
out,
and
we
can
definitely
get
it
added
on
there
and
from
there,
I'm
actually
going
to
turn
it
over
to
jake.
To
talk
about
some
exciting
features
that
we
just
released.
So
let
me
switch
over
all
right,
jake!
Take
it
away.
Can.
B
B
One
thing
is
that
when
it's
due
when
it's
with
ml,
when
it's
trying
to
you
know
figure
out
and
learn
the
algorithms,
it's
computationally
intensive
and
so
gpus
can
actually
do
that
much
better
than
cpu
and
so
right
now
we
only
have
gpu
training
lit
up
for
image
classification,
that
image
classification,
because
it
uses
the
deep
learning
the
neural
networks.
It
is
even
more
computationally
intensive,
so
you
kind
of
get
the
most
gain
by
adding
gpu
training
to
that.
B
B
Sure
so
yeah,
I
don't
know
if
we've
really
stepped
through
this
much.
There
are
definitely
videos
that
brie
has
done
for
kind
of
walking
through
this,
but
I
can
navigate
a
little
bit
so
on
this.
On
this
first
page,
you're
just
going
to
kind
of
choose
your
scenario
so
text
classification,
we
kind
of
touched
on
one
of
the
examples
earlier
with
sentiment,
analysis
it'll,
take
in
like
free
form,
text
and
categorize
it
so
with
sentiment.
B
One
of
the
classic
examples
is
like
having
all
the
features
of
a
house
and
trying
to
predict
its
value
image
classification,
that's
kind
of
one
that
I
can
click
through
later
and
show
you
exactly
how
it
works
through
the
product,
but
it
just
yeah.
Categorizing,
it's
kind
of
like
text
classification
in
this
in
the
sense
that
it's
classification,
so
you're
still
just
kind
of
labeling
it,
but
with
image
classification,
it's
based
on
just
image
data
and
not
not
text
and
then
yeah
for
recommendation.
B
That's
you
know
trying
to
trying
to
give
you
recommendations
just
like
netflix
would
give
you
recommendations
and
object.
Detection
is
a
new
feature
that
luis
is
actually
going
to
talk
about
in
a
little
bit,
but
it's
somewhat
related
to
image
classification,
but
I'll.
Let
I'll
let
luis
kind
of
talk
about
it
more
in
a
bit.
B
So
for
a
lot
of
those
yeah,
there's
only
going
to
be
the
cpu
training,
but
you
would
come
in
here
choose
your
different
options.
For
for
this
specific
example,
we
have
we
have
local,
local,
gpu
and
azure,
so
in
azure
you're
going
to
get
even
more
performance
improvement,
it's
going
to
search
over
more
models,
and
it's
going
to
you
know
free
up
your
resources
on
your
machine.
So
you
can
keep
you
know
developing.
So
if
you
have
something
you
know,
I
I
gave
you
numbers
earlier
about
the
80
000.
B
You
can
imagine
if
you
don't
have
a
compatible
gpu
it'd
be
nice
to
not
tie
up
your
machine
for
four
hours
and
instead
you
know
push
it
out
to
azure
and
get
a
result
back
pretty
close
to
the
same
as
the
gpu
training
here.
But
over
time
the
azure
automotive
team
is
doing
a
great
job
of
sort
of
ramping
up
how
well
it
scales
and
and
the
speed
of
of
training.
B
So
then
you
train
in
here,
so
I
already
have
this
train,
but
normally
you
just
click
start
training
and
it
would
take
so
this
one
was
302
seconds.
So
I
can't
do
that.
It's
like
five
minutes
and
then
once
you
once
it's
done
training
you
come
to
this
evaluate
screen,
and
here
you
can
pick
a
different
image
in
the
case
of
image
classification
for
the
text
ones.
There
would
be
text
fields
here
and
you
would
choose
between
them
and
then
it'll
it'll
go
ahead
and
do
a
prediction.
B
I
think
it's
running
a
little
slower
since
I'm
sharing
my
screen
and
tell
you
that
it's
copy,
so
this
is
just
meant
to
give
you
some
insight
into
the
the
model
that
you've
trained
start
to
give.
You
build
some
confidence
that
it
actually
trains
a
model
that
can
do
what
you're
you
were
intending
and
then
from
there.
B
A
A
There
we
go
so
before
we
model
builder
only
had
cpu
training
for
image
classification
and
it
was
really
really
slow
and
especially
as
you
started,
to
add
more
images
to
your
data
set
and,
as
you
start
to
add
more
classes,
then
we
added
support
for
azure
training,
which
did
help
that
a
ton,
but
we
still
wanted
an
option:
a
faster
option
for
local
training
for
those
who
didn't
want
to
use
at
the
cloud
or
wanted
to
test
it
out
beforehand.
So
now
we
have
gpu
training,
which
is
really
awesome.
A
C
While
I
share
all
right,
I
think
it's
sharing
the
right
screen
there.
We
go
all
right,
so
this
screen
should
look
familiar
to
you
and
basically
the
other
scenario
that
jay
kind
of
alluded
to
earlier
was
object.
Detection
right,
so
object
detection
for
folks
who
may
not
be
familiar.
It's
also
a
computer
vision
problem,
except
that,
as
opposed
to,
for
example,
image
classification,
we're
just
you're
just
trying
to
determine.
C
What's
you
know,
basically
classifying
the
images
with
object,
detection
you're
actually
able
to
pinpoint
where,
in
the
image
certain
objects
are
right
and
you're
also
able
to
classify
the
different
types
of
objects.
So
you
can
imagine
this
being
very
useful
in
scenarios
such
as
self-driving
cars
right.
You
need
to
be
able
to
detect
pedestrians
street
signs
and
things
like
that.
Another
scenario
would
be,
for
example,
if
you're
in
manufacturing.
C
Right,
perhaps
you
you
know
you
have
an
assembly
line
and
you
need
to
find
if
whatever
it
is,
that
you're
making
your
assembly
line,
perhaps
if
so
there
may
be
cracks
or
there
may
be
defects
in
the
product
that
you're
making
object.
Detection
is
actually
a
really
good,
a
really
good,
a
way
to
sort
of
use
machine
learning
to
solve
that
type
of
problem.
And
that
way,
you
don't
necessarily
have
you
know,
sort
of
human
errors,
or
you
know
you
can
minimize
the
the
human
error
involved
in
the
process.
C
So
again,
this
is
something
that's
available
again,
it's
very
familiar
or
will
be
available
in
model
builder,
and
this
looks
very
familiar
right.
You
start
off
where
you
normally
would
with
going
to
your
scenario
in
this
case,
I
went
ahead
and
clicked
object,
detection
here,
and
you
can
see
that
you
have
the
option
to
go
ahead
and
train
out
train
on
azure
all
right
and
it's
very
similar
to
the
image
classification
scenario.
C
You
also
have
that
options
trend
on
azure
and
the
way
that
this
works
right
is
you
have
this
machine
learning
workspace?
You
can
think
of
a
workspace
as
this
place
that
essentially
collects
all
of
your
resources
that
you
would
need
in
natural
machine
learning.
So
that
includes
your
computes
right,
so
so
the
actual
vms
that
are
going
to
be
used
to
train
your
models
and
once
you
go
ahead
and
provide
these
things
right
sort
of
go
here,
workspace
name
right.
C
You
can
create
one
if
you
don't
have
them
already
and
once
you
have
gone
ahead
and
created
an
experiment,
you
go
ahead
and
select
your
data,
so
currently
the
way
that
the
data
format
is
in
is
in
this
sort
of
json
format.
That
being
said,
though,
this
is
again
it's
still
a
work
in
progress,
so
we
would
greatly
encourage
you
to
provide
your
feedback
if
you're
working
with
object,
detection
scenarios,
you
know
what
ways
do
you
prefer
to
format
your
data?
You
know,
how
are
you
storing
it?
C
B
That's
actually
what
we
said
about
the
data
is
true
for
all
of
our
scenarios
like
if
you
can
reach
out
to
us.
Let
us
know
like
where
your
data
is
like
how
your
it's
currently
stored,
so
we
can
help
make
that
process
easier
for
you
to
kind
of
integrate
it
into
model
builder
right
now.
We
we
pretty
much
just
support.
You
know
text
formats
and
some
a
little
bit
of
sql
connections,
but
if
we
can
make
that
process
that
early
process
of
getting
started
in
multiple,
easier,
we'd
love
to
we'd
love
to
hear
how.
C
Yeah,
certainly,
and
and
something
that
I
kind
of
skipped
over
right,
was
that
the
tool
that
it
might
make
it
easier
at
the
moment
to
do
this
formatting
right
is
this
tool
called
bot
right
and
and
by
you,
provide
images,
and
you
can
tag
your
images
and
then
eventually,
you
can
export
your
tags
of
those
images
into
the
json
file.
As
I
mentioned
earlier,
and
that's
gonna,
you
know
give
you
a
nice
format
for
that
model
builder
will
be
able
to
handle
now
here.
C
C
Okay
around
this
cat
right,
so
you
should
be
able
to
see
you
know
a
bomb
box
around
here
and
that's
kind
of
what
you're
seeing
here.
Whoops
there
all
right,
so
that's
kind
of
what
you're
seeing
there.
So
once
you
have
everything
set
up,
you
have
your
data,
you
have
your
environment,
you
go
to
the
train
and
then,
after
this,
it's
literally
the
exact
same
steps
that
you
would
otherwise
do
with
model
builder
that
you're
probably
used
to
which
is
you
train.
C
You
evaluate
kind
of
like
what
jig
was
demoing,
where
you're
able
to
make
sure
that
your
model
is
performing
like
like
it
should,
and
then
it
generates
a
code
for
you
so
that
you
can
then
consume
this
out
of
your
you
know,
asp.net
or
whatever
end
user
application.
So
definitely
stay
tuned
for
that,
and
we
know
that
it's
something
that
you
know
the
community
really
is
passionate
about
and
they
really
care
about
the
scenario.
A
All
right
awesome,
please!
Now
I
will
attempt
to
oh
wait.
It
put
it
back
for
me,
awesome
so
yeah.
Those
are
just
a
few
mini
demos
to
show
off
a
new
feature.
We
just
had
a
new
feature.
That's
coming
up
some
other
things
to
look
forward
to
besides
object
detection,
where
our
other
tool
that
we
have
is
the
ml.net
cli.
So
that's
our
cross-platform
tooling,
offering
we're
looking
we're
about
to
add
image
classification.
A
D
D
D
I
think
it's
a
great
library
and
I'm
trying
to
extend
some
of
the
capabilities
the
library
currently
has
and
one
of
the
things
that
I
was
working
on
when
I
was
looking
at
building
machine
learning
models
I
was
earlier
this
year.
I
think
you
know
I
think
liz
can
attest
to
to
my
questions
was
that
once
I
had
a
model
and
once
I
got
this
process
up
and
running,
I
had
no
real
good
way
of
putting
into
production
and
managing
it,
and
you
know.
Where
would
I
store
these
things?
D
D
That
can
all
help
with
all
these
things
and
the
general
you
know,
question
I
had
was:
how
can
I
automate
my
machine
learning,
training
and
deployment
just
like
we
do
with
devops
as
a
software
in
air
rights
and
kind
of
just
integrate
everything
very
nicely.
The
issue
was
struggling
with.
Is
that
neither
of
these
tools,
supported.net,
so
ammo
flow,
even
though
it
tested
supports
every
machine
learning
library
does
not
support
ammo.net,
which
is
unfortunate.
D
This
is
a
great
tool
otherwise,
and
I
also
want
other
capabilities
than
just
being
able
to
be
in
the
cloud.
So,
although
aws
and
azure
are
great
ways
of
training
your
models
on
on
clusters
and
managing
those
in
azure
machine
learning,
I
also
wanted
the
ability
to
do
it
locally
and
as
well
as
in
the
cloud
so
I
set
up
in
like
may,
I
think,
to
start
building
this
a
tool
called
amalops.net
luis.
D
D
D
So
it's
a
pretty
daunting
task
when
you
start
looking
at
this
on
like
how
how
to
banish
envelopes
and
how
to
make
sure
you
get
the
best
out
of
your
model
and
successfully
deploy
to
production
and
continue
adding
value
for
your
company.
So
what
is
tool
is
really
supposed
to
do
right
now
and
you
know
we're
building
on
it
and
we'd
love.
D
Your
support
is
specifically
experiment
tracking,
so
tracking,
you
know
an
experiment,
meaning,
for
example,
as
release
pointed
out
here,
detecting
cats
faces
on
pictures
and
then
for
each
time
we
try
to
train
a
model.
We
have
a
run
which
have
things
associated
with,
like
you
know,
evolution
metrics
how
good
it
was
hyper
parameters
so
how
we
fine-tune
that
model,
but
also
things
like
tracking.
You
know
where's
the
data.
Where
do
you
get
it
from?
Has
it
changed
since
last
time?
If
we're
using
classification?
D
For
you
know,
fraud,
things
like
data
distribution
is
very
important
to
ensure
we
have.
You
know
balanced
data,
so
tracking
those
things
are
good
and
when
we
have
a
model,
that's
fully
trained.
Where
do
you
put
it?
So
we
need
some
kind
of
model
repository
and
currently
we
store.
We
support,
azure,
blob
storage
and
aws
s3
buckets
as
well
as
local
file
shares
if
you
wanted
to.
But
this
model
repository
is
useful
because
you
can
get
these
versioned
models.
So
if
you
have
a
run
that
is
better
than
the
previous
one.
D
We
can
then
add
versioning
to
this,
and
also
the
full
audit
trail
of
where
this
model
come
from,
who
trained
it.
What
data
was
used
for
it
and
so
forth?
So
you
get
this
audibility,
so
you
can
be
accountable
for
your
clients
and
whoever
consumes
your
model.
And
finally,
what
we're
currently
working
right
now
is
the
deployment
phase.
So
how
do
you
deploy
and
consume
your
model?
D
And
I
mean
that's
a
really
cool
thing
with
m
networks,
you
can
do
it
in
so
many
ways,
because
it's
the
net,
so
you
can
do
it
as
like
a
uri
based
deployment
right.
You
can
read
the
model
from
a
location
wherever
you
want
to.
You.
Can
embed
it,
you
can
put
it
in
a
docker
container,
there's
many
ways
of
doing
it.
D
Currently,
we
support
uri
based
deployment,
so
we'll
track
that
for
you
and
manage
all
those
things
for
you,
we're
currently
working
on
the
docker
deployment
to
automatically
generate
these
images
and
push
them
to
your
clusters
and
track
policies
for
you-
and
one
thing
I
think,
is
really
cool
with
this.
As
we
work
on
this
is
that
you
don't
need
to
supply
dependencies,
you
won't
need
to
supply
like
piping
stuff,
that
you
need
for
psychic
learner
python,
for
example
like
okay,
here's,
my
entry
point:
this
is
how
I
call
the
model.
D
This
is
how
I
transform
the
data
before
calling
the
model.
These
are
my
packages
you
need
to.
You
know,
do
like
I'm
hoping
we
can
skip
all
that
stuff,
because
we
can
use
reflection
to
figure
out
which
package
dependencies
we
have.
We
can
use
code
generation
in
c,
sharp
and
dot
new
templates
to
do
what
it
seems
for
us,
so
it
should
be
as
easy
as
saying.
Okay,
I
have
this
model.
Please
create
an
image
of
it
and
push
it
to
this
cluster,
and
I
don't
want
to
worry
about
all
the
the
complexity.
D
That's
around.
You
know
deploying
a
machine
learning
model,
so
there's
a
lot
of
work
being
done
here
and
hopefully,
at
the
end,
we'll
have
a
client
as
well
that
you
can
visualize.
You
know,
compare,
runs
and
and
all
these
things
managing
through
a
ui,
but
so
far
we
have
a
new
package.
That's
deployed
here.
You
can
use
that
lots
of
people
are
contributing
here.
I
would
definitely
want
to
highlight
brett
parker
would
highlight
aj
from
new
zealand
and
and
tons
of
others
as
well.
D
A
D
So
the
idea
behind
the
containerized
deployment
right
is
to
more
or
less
deploy
a
registered
model,
but
without
having
to
supply
exactly
what
the
schema
made
for
input
and
output
and
and
the
asp.net
core
api
that
the
model
should
be
wrapped
around.
So
we
needed
some
good
way
of
you
know
automatically
generating
an
asp.net
core
api
for
amazon
net
that
exposes
the
model
through
an
endpoint
and
all
those
things.
D
D
So
we
have
a
package
as
well
after
that,
and
we
only
have
three
templates
so
far,
but
they're
pretty
quick
to
create,
and
if
you
have
an
experience
with
dotnet2
templates,
you
can
definitely
chime
in
and
chime
in
anyway.
For
that
matter,
it's
it's
pretty
straightforward
and
we
would
love
more
examples.
D
So
if
there's
something
you
want
to
create
open
an
issue
and
and
start
collaborating
on
it,
that'd
be
awesome
or
just
open
an
issue
and
we'll
we'll
create
it
for
you,
but
the
way
you
can
use
it
just
to
kind
of
give
an
idea
here
we
can
use.
It
is
by
installing
the
templates
from
nuget,
for
example,
we
currently
have
pre-release,
so
you
can
hit
that.
U
install
and
the
template
version
here
and
that
will
install
them
and
just
like
any
other
template
here,
they
will
now
show
up
in
your
little
new
templates
here.
D
D
We
can
then
look
at
how
that
template
looks
like
and
what
comes
with
it.
So
as
as
normal,
it
comes
with
a
schema.
So
we
have
a
model
input
here
and
we
can
fill
that
out
ourselves
depending
on
our
schema,
looks
like,
and
it
comes
with
this
pre-generated
file
here,
which
includes
the
general
flow
for
your
training
model.
So
if
you've
done
this
a
bit
before,
you
probably
recognize
this
a
lot.
D
But
you
see
how
you
know
we
load
data
from
a
file
here
and
we
select
our
trainers
and
so
forth,
but
in
addition
to
that,
it
also
comes
with
mlabso.net.
So
we
can
create
a
run
here
down
here,
which
is
just
defining
you
know.
Do
we
want
to
use
sqlite,
sql
server
or
cosmos
for
backing
storage
for
data,
and
where
do
we
want
to
store
a
model?
Is
it
locally
or
an
s3,
bucket
or
somewhere
else,
and
based
on
that,
you
know,
we
can
then
use
that
ml.
D
D
How
does
it
look
like
in
general,
and
we
can
also
do
multiple
more
things
here
to
see
like
hyper
parameters
and
training
times
and
uploading
models
to
the
cloud
or
wherever
you
want
to
and
there's
more
ways
of
doing
this
as
well,
but
in
general,
the
library
has
been
designed,
so
it
feels
like
ambled
net
and
it
respects
ml.net
by
it's
doing
it.
Natively
right
so
passing
in
data
views
everything
and
we're
passing
in
just
a
trainer
in
this
case
here
and
automatically
figure
out
the
hype,
parameters
and
log
them
for
you.
D
A
A
A
D
A
D
Yes,
absolutely
so
we
had
a
great
reception
of
the
first
conference,
which
is
awesome.
We
had,
I
think,
700
sign
up
across
the
world
and
you
know
when
we
started
doing
this.
It
was
john
wood
put
on
twitter
one
day
before
I
even
talked
to
him
that
that
wouldn't
be
fun
with
the
conference
on
emerald
net.
D
D
Anyone
who
wants
support
with
solving
a
specific
machine
learning
problem
with
emble.net
of
their
choosing,
I
think
and
then
we'll
kind
of
score,
the
creativity
and
the
accuracy
of
those
models,
and
I
think
we'll
have
you
know
different
kind
of
scoring
cards.
I
believe,
depending
on
how
creative
you
are,
because
you
know
in
many
cases
you
may
have
an
awesome
idea,
but
to
solve
something
in
three
days
without
proper
data.
D
A
A
D
What
would
it
be?
Oh,
I
have
so
many
things,
and
you
know
I.
I
speak
a
lot
about
this
and
I
meet
a
lot
of
people
from
the
community
and
I
hear
about
what
they
want
as
well.
There's
a
couple
things
that
I
think
are
cool
factors
which
I
don't
necessarily
always
think
they
will
add
much
like
firm
value
for,
and
there
are
things
that
will
add
some
actual
value.
D
So
some
of
the
cool
factors
right
would
be
a
keras
like
api
for
deep
learning
that
you
can
kind
of
create
your
own
neural
network.
You
can
decide
how
many
layers
you
have
your
optimization
functions
and
dimensions
so
forth,
so
you
can
do
that
yourself
a
bit.
So
we
have
full,
like
you,
know,
keras,
like
api
for
nets
and
I
think
that'll
be
really
cool
and
definitely
use
cases
for
that
as
well
things.
That
would
be
useful.
D
A
That's
awesome
thanks,
alexander
I'll,
actually
keep
you
on
we're.
Gonna
move
to
q,
a
and
alexander
feel
free
to
answer
any
questions
as
well.
So
I
think
I
will
s
go
ahead
and
if
anyone
has
questions
feel
free
to
post
them
in
the
comments,
I'll
try
and
go
back
louise,
if
you
see
any
that
you
haven't
been
able
to
answer
in
the
chat,
feel
free
to
read
it
out
and
answer.
B
So
I
saw
I
saw
one
question
on
there
that
was
about
a
code
first
approach
and
having
a
little
bit
more
control,
I
have
a
recommendation
which
might
not
be
exactly
what
you
were
thinking,
whoever
whoever
was
that
asked.
I
think
they
lost
it
in
the
chat
here,
but
I
would
actually
still
recommend
you
kind
of
start
with
model
builder,
the
cli
inside
so
I'll
share
my
screen
again.
E
B
B
A
Yeah
also
add
in
our
samples,
repo
there's
a
ton
of
different
examples
that
if
you
clone
that
project,
you
can
also
modify
it
based
on
your
own
data
or
your
own
scenario
or
swapping
out
algorithms.
So
that's
another
good
way
to
start
our
docs
has
a
ton
of
tutorials
starting
from
scratch
as
well,
but
if
you
all
want
next
week
or
in
two
weeks
when
we
have
our
next
stand
up,
we
can
totally
do
a
deep
dive
into
the
code.
If
you
want
to
explain
the
different
parts
of
the
code.
D
I
think
just
add
as
well.
I
think
the
model
build
is
fantastic
start
in
many
cases
like
especially
for
new
ml
tasks
like
I'm
not
so
experienced
with
ranking.
So
if
I
looked
at
that
model
builder-
or
something
would
be
great
just
to
to
you
know,
get
familiar
with
what's
expected,
but
then
you
know
when
you
want
to
take
it
to
production.
D
You
definitely
want
to
fine-tune
the
code
for
your
business
needs
because
you
will
know
more
about
the
data
than
model
builder
will
ever
know,
because
that's
just
the
nature
of
things
and
you
can
add
more
context
to
that.
But
it's
a
great
start.
I
think,
and
and
also
gives
you
a
good
structure
of
the
code.
A
B
I
can
kind
of
touch
on
it
and
then
maybe
people
can
add
stuff
to
it.
I
don't
think
we
really
have
support
for
that
at
the
moment
like
like
nothing
native,
we
haven't
really
thought
about
it
specifically,
but
what
you
described
would
work
so
like,
as
you
start
getting
more
data,
you
can
go
back
iterate
on
the
models
and
perform
fit.
I
think
I
I
read.
B
Maybe
I
read
another
question
by
the
same
same
person
who
who
asked
like,
if
you
were,
while
you
were
typing
what
you
could
do
so
like
if
you
would
just
want
to
like
predict
the
next
word
or
something
like
that
that
wouldn't
actually
require
retraining.
There
might
be
something
you
can
do.
It
would
be
a
pretty
big
model.
I
don't
know
if
we
have
something
great
for
that
right
now,
but
I
have.
A
B
B
D
D
C
D
C
D
No
exactly
right,
that's
kind
of
one
of
my
wishlists
as
well.
It's
the
cares
like
api,
and
I
I
can
definitely
feel
that.
Maybe
it's
not
always
much
for
business
value,
but
it's
there's
definitely
a
possibility
for
it
and
it
definitely
brings
the
cool
factor
to
the
library
as
well,
especially
considering
you
know
that
the
the
net
bindings
on
top
of
something
like
tensorflow
is
going
to
be
a
lot
faster
than
those
on
top
of
python.
So
that's
really
exciting.
I
think.
B
And
we
also
have
sorry
go
ahead,
I'm
just
gonna
say
I
mean
it's
on
it's
on
our
wish
list
as
well
internally.
I
think
that
you
guys
can
all
help
us
kind
of
push
it
to
the
top
of
our
priorities.
If
you,
you
know,
come
to
our
our
github,
perhaps
we
can
create
some
sort
of
survey
that
we
can.
We
can
add
to
the
community
links
to
this,
to.
B
Let
us
know
how
you
want
to
use
it
like,
especially
if
you
can
give
us
like
a
a
business
use
case
and
a
a
story
like
and
or
just
you
know,
if
everyone's
saying
they
they
really
want
it,
then
we'll
really
try
to
try
to
prioritize
it,
but
yeah
it's
it's.
It's
super
cool,
we'd
love
to
have
it
in
in.net,
and
I
think
it's
it's
inevitable.
Sometime
we're
just
trying
to
trying
to
try
to
figure
out
when
and
how.
B
Going
on
within
within
microsoft,
around
neural
networks
and
such
that
we're
hoping
to
leverage
and
build
on
top
of
so
some
of
this
is
like
how
much
do
we
do
now
in
a
world
in
a
technology?
That's
like
changing
rapidly
versus
waiting
for
some
of
the
other
cool
projects
that
are
happening
to
stabilize
a
little
bit,
and
then
we
jump
on
those.
But
yes,
it's
definitely
not
a
problem.
We're
like
avoiding
we're
investing
in
it.
It's
just.
C
Yeah,
and
with
this
one,
what
you
could
do
is
for
feature
extractions,
you
can
use,
and
I
added
it
to
the
chat
as
well,
but
it's
the
dnn
featurized
image,
transform
right
and
that's
gonna,
do
kind
of
what
I
what
I
believe
that
you're
alluding
to
your
dean,
which
is
you
want
to
extract,
features
from
your
images
and
then
there's.
You
can
use
the
image
analytics
package
nougat
package
to
perform
a
lot
of
transformations
and
and
any
operations
that
you
may
need
to
perform
on
images.
B
Example
for
like
taking
the
features
from
from
an
image
and
then
tossing
it
into
a
classification
model
and
that
actually
performs
fairly
well
to
like
do
image
classification,
but
you
can
imagine
that
you
could
take
that
data
as
one
set
of
features
but
then
bring
in
other
features
as
well.
So
the
only
example
I
can
come
up
with
my
brain
right
now
is
like.
B
If
you
were
going
back
to
that
like
housing
example,
you
could
imagine
you
could
take
a
picture
of
it
and
pass
that
into
like
classification
or
something
and
then
take
the
output
of
that
and
pass
it
into
a
regression
model
to
kind
of
help.
You
know
try
to
incorporate
the
quality
of
appearance
of
a
home
into
the
price
or
something
like
that,
but
that
sort
of
model
composition.
You
could
do.
A
This
is
the
example
I
was
talking
about
before.
Let's
see
my
computer's
running
very
slow,
though
so,
oh
no
come
on,
you
could
do
it
so
normally,
so
normally
this
little
ticker
would
go
left
and
right,
depending
as
you're
typing
in
real
time.
This
is
actually
in
our
samples
repo.
So
you
can
try
it
out,
I
think,
just
with
oh
there
we
go
we're
trying
to
share
my
screen
and
everything.
A
A
C
D
Exactly
right
and
join
in
to
all
the
fun
I
mean,
if
you
see,
I
also
encourage
your
community
here.
If
you
see
something
that
is
not
built
by
them
of
that
team,
there's
no
reason
not
to
build
yourself,
and
you
know,
there's
tons
of
things.
We
need
to
build
an
ecosystem
and
not
everything
going
to
be
built
on
amazon.
D
A
E
B
There
was
one
more
question
I
saw.
I
don't
know
if
it
got
answered
on
data
processing
before
tossing
into
the
model.
I
think
that
that's
it's
kind
of
outside
of
the
scope
of
our
tooling
as
far
as
like
what
what
model
boulder
does,
but,
yes,
generally,
it
does
like
you
need
to,
at
the
very
least
get
it
into
like
a
csv
tsv
format,
and
it
generally,
you
will
improve
your
your
model
performance
with
some
manipulation
of
the
data
to
try
and
you
know
make
it
easier
to
to
train
on.