►
From YouTube: Weekly Sync 2021-03-30
Description
Meeting Minutes: https://docs.google.com/document/d/16u9Tev3O0CcUDe2nfikHmrO3Xnd4ASJ45myFgQLpvzM/edit#heading=h.mkpbq33u707t
A
A
A
That's
good
great,
let's
see
natesh
anything
from
you
today.
D
Okay,
all
right
so
questions
so
oops.
Okay,
we
are
waiting
until
accuracy.
Changes
are
complete
because
that.
A
A
A
A
All
right,
so,
let's
talk
about
commit
message,
formatting
first,
because
that
is
okay
and
this.
This
issue
work
issues
yeah
great,
and
this
is
okay,
this
guy
all
right.
Okay,
actually,
let's
talk
about
this
first,
so
basically,
the
thing
what's
going
on
here
is
that
we
don't
have
when
you
add
something.
So
when
you
add
a
new
model,
you
usually
will
need
to
add
it
within
a
new
plug-in
right.
A
So
if
you
are
using
this
cap
boost,
that's
from
pi
pi
pie
right
so
and
I'm
assuming
this
is
something
you
install
from
pipe
pipe.
I
write
yes,
pip
install
great,
so
this
is
something
that
you
would
need.
You
would
need
to
create
a
separate
plug-in
for
now
we're
we're
sort
of
in
the
middle
of
how
plug-ins
work
is
changing.
So
this
is
a
great
you
know.
A
This
is
a
great
great
great
implementation
within
a
file
right,
which
is
you
know
what
that
first
tutorial
sort
of
takes
you
through
and
then
you
can.
You
know,
use
this
from
the
command
line
and
stuff.
If
it's
just
in
this
single
file,
you
can
reference
it
using
its
python.
The
entry
point
path
to
the
model
right.
A
The
thing
is,
if
you
want
to
reference
it
from
this
entry
point,
then
you
have
to
make
it
a
package
and
to
run
it
in
the
ci
and
to
add
it.
You
know,
and
all
basically,
if
you
have
for
every
distinct
dependency
that
we
have
so,
for
example,
cap
booths
is
a
new
distinct
dependency.
So
it's
not.
You
know
it's
not,
for
example,
the
source
mysql
stuff,
it's
dependent
on
the
mysql
library.
A
You
know
the
tensorflow
stuff,
it's
dependent
on
tensorflow,
and
so
we
try
to
create
a
new
plugin
or
a
new
package
for
every
for
every
every
p,
every
set
of
code
that
has
a
distinct
set
of
dependencies
right,
and
this
enables
us
to
you
know
not
install
because
the
thing
is,
if
you
install
all
these
machine
learning
libraries,
you
end
up
with
like
gigabytes
and
gigabytes
and
gigabytes
of
stuff,
and
so
not
everybody
wants
to
install
everything.
A
So
this
is
why
we
have
this
plugin
system,
and
so
you
you,
we
we
create
a
plug-in
anytime.
We
have,
you
know
a
new
distinct
dependency,
and
this
would
be
a
case
where
we
would
like
you've
done,
create
dfmo
model
capabilities
and
model
capus.
That's
exactly
correct.
There
would
just
be
a
lot
more
files
under
there.
You
know
things
like
the
setup
up,
config
and
stuff
like
that
right.
The
setup
config
is
how
we're
defining
our
packages
so
right
now
we're
actually
in
the
process
of
changing
things.
A
So,
and
I
mean
I
wish
it
would
give
me
the
titles.
Sometimes
it
wants
to
convert
them.
Sometimes
it
doesn't
cowboys
classify.
This
is.
A
A
A
Until
yash
figures
out
how
to
do
this,
second
party
plug-in
thing
that
we've
been
talking
about
figures
out
how
second
party
plugins
are
going
to
work,
and
essentially
what
that's
going
to
be
is:
is
we're
going
to
try
to
host
a
bunch
of
plugins
github.com.
A
We're
gonna
try
to
host
a
bunch
of
plugins
under
this
org
and
so,
for
example.
This
is
the
model
transformers,
one
that
we
split
out
and
this
was
contributed
to
us
by
himachu
and
and
then
himachu.
Unfortunately,
none.
A
To
update
the
apis
and
and
humachu
is
busy
with
a
full-time
job
now,
so
it
got
moved
into
here
and
and
sort
of
the
idea
is
we'll.
Have
this
support
matrix
of
understanding?
You
know
what
plug-ins
work
with
each
other,
because
sometimes
you
get
these
version
conflicts
and
I
think
whoever
did
the
orion
model
recently
ran
into
this,
but
you'll
run
into
these.
A
These
contextual
version
conflict
errors
where
package
resources
will
will
throw
up
exceptions
when
you
load
code,
because
certain
thing
each
certain
packages
depend
on
versions
of
things
that
are
not
what
is
installed
so,
for
example,
if
if
you
have
a
version
range
restrictor
on
numpy-
and
you
said
well,
let
me
just
pull
up
this
as
an
example,
so
tensorflow
greater
than
2.0
but
transformers
less
than
3.1.
A
So
when
you
require
transformers,
it
means
that
you
can't
use
tensorflow,
2.4
and
it'll,
throw
up
an
exception
and
there's
this
is
you
guys
will
run
into
this
all
the
time,
because
all
these
libraries
have
a
bunch
of
conflicting
stuff,
so
we're
trying
to
figure
out
this
second
party
plug-in
situation
so
hold
off
on
on
doing
more
work
here,
unless
you
have,
unless
you
wanted
some
specific
feedback,
I
can
give
you
some
specific
feedback
if
you're
trying
to
use
this
for
something
right
now,
if
something's
not
working,
but
otherwise
we're
going
to
hold
off
on
this,
and
I
will
mark
it
as
second
party
plug-in
so
that
we
know
that
we
can
get
back
to
it.
A
B
Actually,
I
my
target
was
to
merge
this
pr
so
that,
along
with
this
pr
I
can
I
am
I
am.
I
will
see
that
how
the
things
are
working
on
this
plug-in.
A
A
B
I
have
one
query
in
the
just
for
like
how
the
things
are
working
over
here.
So
can
you
move
to
the
pier.
B
In
the
class,
the
cadbury
cad
boost
classifier
class.
We
are
taking
lots
of
pairing
parameters.
I
am
taking
the
references
from
the
exhibits
like
this.
They
both
are
almost
similar.
A
B
So,
even
when,
even
in
xj
posting,
we
are
not
using
all
the
parameters
like
we
are
using
l
learning
rate
and
underscore
estimators,
but
along
with
this,
there
are
various
other
parameters.
So
we
are
not
including
that
in
this
class,
in
the
exhibits
thing
so
like
that.
B
A
Just
that
no
one
got
around
to
doing
it.
Let's
see,
I
think
this
was
just
meant
as
more
of
a
quick
example
at
the
time
to
show
how
to
how
to
get
it.
Working,
let's
see
so
yeah
all
of
these
you're
saying
we're
not
using
some
of
them.
B
Yeah,
there
are
many
more
parameters
which
we
have
not
included
here.
A
Yeah,
you
know
so
so
this
is
a
good.
This
is
a
good
catch.
So
where
is
let's
see
if
we
can
find
the
docs
for
xgb
regressor.
A
So
I
think
the
thing
is
that
that
we
have
hard-coded
those
config
structures
for
this
guy
and
there
are
several
functions
that
allow
us
to
so
we
we
hard
coded
the
the
config
values
for
those
guys.
Let's
see,
where
is
the
api
reference
and
util
config.
A
A
Okay,
let's
just
go
to
xg
boost,
so
we
have
these
two
functions
and
and
and
the
make
config
numpy
will
parse
the
numpy
style,
doc
string
to
build
the
appropriate
config
parameter
dictionary
or
the
profit
config
structure,
and
it
will
make
sure
that
everything
that's
referenced
like
everything
here,
anything
that's
listed
for
in
the
documentation
string
becomes
part
of
the
you
know
it.
It
automatically
creates
properties
for
those
so
and
does
does
anybody
remember
where
we?
Where
are
we
doing
that?
A
A
F
D
F
A
C
E
C
E
D
A
Yeah
yeah
perfect.
Thank
you
good
job,
natasha,
all
right!
So
in
this
we
have
so
this
config
so
make
number.
No.
Yes,
we
call
make
numpy
config
and
we
pass
it
lgbm
model
and
where
is
okay?
Hopefully
this
links
me
to
the
code
source
great
so
and
you'll
notice
so
make
numpy.
Config
is
gonna.
Look
at
this.
It's
gonna
parse
out
this
stock
string
and
it's
gonna
create
appropriate.
You
know
it's
going
to
create
appropriate
values
of
the
types
listed
with
the
default
values
listed.
A
A
Where's
make
notepad
config.
So
we
call
this
make
config
numpy
and
that
parses
this
docs
string
and
it
will
build
the
config
structure,
that's
appropriate
and
then
we
can
add
any
other.
You
know
properties
that
we
may
need
to
add
there,
and
so,
for
example,
you
know
we
want
everything,
that's
that
would
come
from
parsing
this
and
we
want
features,
predict
and
directory
then
so,
then,
what
we
can
do
is
we
can
basically
convert
it
to
a
dictionary.
A
We
can
delete
the
things
that
we've
added,
and
so
now
we
only
have
things
that
came
from
the
from
the
you
know
from
the
possible
parameters
here,
and
we
just
do
variable
expansion
to
pass
them
all.
So
this
is
the
ideal
way
to
do.
Do
things
like
this,
like
what
you're
talking
about
where
you're
saying
you
know
why
haven't
we
used
all
of
them?
It's
because
we
didn't.
A
I
don't
think
we
had
that
function
at
the
time
this
code
was
written
so
or
either
that
or
either
xbg
regressor
did
not
have
a
numpy
style
docstring
or
it
did
not
have.
A
A
type
hints
on
it
because
you
also
you
can
use,
make
config
inspect
if
there's
type
hints,
and
we
need
to
to
to
document
all
this
more
so,
and
let's
make
a
note
of
that.
A
So
we
need
to
document,
make
config
inspect,
slash,
numpy
and
when
to
use
in
the
new
model.
Tutorials.
A
A
Okay,
we'll
make
sure
we
document
that,
because
that's
that's
important,
because
as
you're
going
through
and
creating
these
models,
you
know
this
is
the
easiest
way
to
do
it.
So,
let's,
let's
make
sure
we
do
that
and
let's
say
light
gbm
is
a
good
example
all
right.
So
then
what
what
else
were
you
saying
on
this?
So
you
wanted
to
figure
out
the
plug-in
system
and
you
wanted
to
you
wanted
to
you
know
you
were
just
playing
around
with
writing
a
model.
A
I
would
suggest
going
through
the
rest
of
that
packaging
tutorial
we're
going
through
the
packaging
tutorial
because
you'll
end
up
with
you
know
you,
so
you
did
the
writing
a
model.
I
assume,
and
now
now
you
know
going
now.
You
want
to
go
through
this
packaging
tutorial,
because
this
will
show
you.
This
command
will
give
you
that
whole
directory
structure-
and
you
basically
take
this
and
you
move
it
under
model
slash
whatever
now.
A
The
thing
is
this
is
this
is
going
to
change
when
we
have
the
second
party
plugins,
but
right
now,
that's
sort
of
the
way
things
have
worked
is.
Is
you
would
move
it
under?
Is
that
what
it
says
down
here?
Yeah,
it's
it's
mentioned,
so
you
move
it
under.
You
know
that
path,
and
then
you
submit
a
pr
so
so
that
would
be
sort
of
the
extension
of
what
what
you've
done
there,
and
this
will
show
you
how
to
write
tests
and
things
which
is
going
to
be.
Let's
see.
A
A
The
packaging
tutorial
but,
like
I
said
we're
going
to
wait
on
hold
merging
those
until
we
get
to
the
third,
the
second
party
plugins,
because
that
will
that
will
sort
of
make
everything
make
more
sense
all
right.
I
have
already
completed
that
packaging
tutorial.
You
have.
Okay,
do
you
have
any
questions,
then?
If
on
the
plug-in
system
or
figuring
out,
you
know,
you
said
you're
trying
to
figure
out
the
plug-in
system.
A
Okay,
just
let
us
know
just
let
us
know
and
that's
that
if
you
come
up
with
any
questions,
that's
a
good
thing
to
cover
in
the
meeting,
because
it's
I
think,
we've
covered
it
in
a
few
meetings
actually
and-
and
I'm
not
sure,
I
think
it
might
be
in
the
meeting
minutes
where
we've
talked
about
entry
points
and
stuff
and
the
recordings
there
can
give
you
more
insight
into
how
the
whole
thing
works.
A
B
Actually,
according
to
the
idea,
we
we
have
to
first
hyper
tune,
the
hyper
parameter
and
we
have
to
choose
a
model.
That's
our
target,
but
my
question
was:
do
we
need
to
implement
a
plugin
type
thing,
as
there
are
some
automatic
frameworks
available,
such
as
auto
scale
and
auto
pytorch,
or
we
need
to
create
a
new
automobile
framework?
That
was
my
question.
We're.
A
Creating
a
new,
auto
ml
framework
because
we
already
have
like
auto
sklearn,
and
I
think
that
h2o
model
does
some
auto
ml2,
so
we're
basically
trying
to
you
know
so
we
have
this
this
ecos.
We
have
this
like
ecosystem
of
plug-ins
right
and
so
now.
The
goal
here
is
to
be
able
to
use
models,
to
optimize
other
models
right
and
and
and
and
figure
out.
You
know
what
what
how
do,
how
do
I
tune
type
of
printer
parameters
of
the
models
that
I
we
have
in
our
ecosystem?
A
A
How
do
we
integrate
the
data
flow
part
and
data
flow
cleanup
part
into
this?
So
but
you
know
that's
sort
of
an
extension
of
this
whole
thing.
Any
other
questions
on
that.
B
No,
that's
it.
Okay,
great.
A
Yeah
we
are
implementing
from
ourselves
not
wrapping
another.
A
H
Today,
yeah,
I
wanted
to
ask
you
if
we
are
already
doing
multi-output
models.
A
I
don't
believe
we
have
any
models
that
that
predict
multiple
outputs.
That's
that's
something
that
that
would
be
would
be
good
to
look
into,
though
I
don't
think
there's
anything
that
restricts
us
from
doing
that.
At
the
moment
we
just
don't
have
any
implemented
that
do
that
and
that
that
actually
yeah,
if
you're
thinking
about
use
cases,
that's
a
good
one.
I
had
run
into
that
a
while
ago
wishing
we
had
that.
I
can't
remember.
H
All
right
cool
yeah,
I
have
curated
some
other
use
cases
as
well
I'll,
send
you
those
offline
other
than
that.
I
don't
have
anything
to
discuss
at
the
moment.
A
G
A
Okay,
let's
see
all
right,
which
prs
are
those.
G
A
Okay
and-
and
I
think
that's
basically
on
hold,
though
too,
with
the
second
party
situation,
so.
A
Okay,
sudanshi.
B
Yeah,
so
I
actually
wanted
to
show
you
like
what
I
have
been
up
to
with
the
ice
cream
demo.
Oh.
A
A
Okay,
so
let's
see
okay
so.
A
A
A
All
right
so
and
everybody
I
so
there
was
a
recent
refactor
of
all
of
the
cash
download
stuff,
and
so,
if
you
need
to
do
internet
stuff,
sync,
url,
open
and
sync
url
retrieve
do
the
do
the
protocol,
validation
and
then
sync
url,
retrieve
and
validate,
does
download,
plus
hash
validation.
So
this
is
very
interesting.
So
if
you
don't
call,
maybe
this
is
not
returning
what
is
call
doing
return
wrapped,
okay,
so
wrapped.
A
I
A
A
This
is
one
where
we
probably
need
a
better
way
to
do
this,
so
should
I-
which
one
was
it
it
was
there's
a
few
of
these
and
basically
it
goes
into
text
vulnerabilities
and
there's
was
a
few
of
them
where
it
was
detecting.
It
was
doing
a
comparison
on
the
exact
number
and
we
should
just
be
doing.
Compare
greater
yeah
assert
equal.
We
should
just
be
doing
a
certain
greater
because
the
vulnerability.
I
A
I
And
one
more
thing
about
this
issue
was
that
the
code
coverage
stopped
by
a
very
small
amount.
So
should
we
write
a
test
for
this.
A
Yes,
we
should
add
a
test
for
this
and
you
can.
A
So
use
you
look
at
look
at
the
way
that
the
test
for
the
release,
stuff
works
and
you
and
and
the
way
that
that
unit
test
mock
works
there.
You
know,
I
think
you
can
do.
Let's.
A
D
A
Yeah,
I
think
you
can
you
can
you
can
look
at
the
way
that
the
release
stuff
for
unit
test
mock
works
and
and
I'll
just
pull
that
up
right
now,
so
you
can
see
it,
but
your
mouse
urban
test
service
test,
dev.
A
So
this
has
a
few
things
where
we're
mocking
sub
process
stuff,
and
this
is
a
good,
a
good,
pretty
good
example
of
let's
see
a
fake
process
and
fake
process
sort
of
does
different
things
based
on
you
know
what
what
the
command
that
was
run
was
supposed
to
be
so
there's
like
there's
there's
a
few
commands
that
get
run
in
the
release
like
by
subprocess
exec,
and
so
we
fake
that
exact
creation
call
and
use
this
make
exec,
and
then
we
pass
different
kinds
of
processes
and
these
fake
processes.
A
So
we
create
this
fake
process
class
and
when
we
wait
for
it
to
finish
executing,
we
say
you
know
we
can
switch
based
on
what
the
command
was.
That
was
run
and-
and
we
can
say,
okay
well
if
it
was
not.
If
if,
if
it
was
the
archive
command,
create
this
archive,
because
that's
one
of
the
steps
that
happens,
and
so
you
can
do
similar
things
right.
So
if
you
need
to
run
something,
then
you
can.
A
You
can
fake
running
that
or
you
know,
for
example,
if
you
need
to
check
the
sequence,
this
is
probably
you
probably
need
to
do.
A
similar
thing
where
you
need
to
check
the
check
the
sequence
of
things
that
ran
right
and
that
would
be
you
know
something
like
this,
so
so
this
is
going
to
be
a
good
reference
for
you,
probably
and
and
yeah
unit
test
mock.
Is
your
friend
here
to
avoid
network
interaction
and
interaction
with
you
know,
other
other
things.
A
Also,
you
probably
want
to
make
sure
that
you
do
this
in
some
kind
of.
Let's
see
you
do
you
don't
want
to
run
this
test.
So
when
you
do
those
link,
creations
and
stuff,
you
want
to
check
that
you
created
all
the
right
links,
so
I
would
add
tests
for
that
and
I
would
add,
support
for
doing
this
to
a
target
directory.
A
So
that
way
you
can
create
a
temporary
directory
so,
for
example,
all
of
these
docs
path
right,
so
I
would
say
that
this
root
so
take
take
pages
path
as
an
argument
kind
of
like
because
you
have
yeah
so
take
pages
path.
As
an
argument
take
this
as
an
argument
via
the
config
structure.
A
I
Okay,
so
can
we
test
it
like
against
the
old
dog
starters.
A
Test
it
against
the
old
docs.sh.
No,
because
that's
going
to
go
away,
but
what
you
sh
so
so
take
take
the
page's
path
as
an
argument
and
then
make
sure
that
all
of
the
you
know
the
right
hand.
Side
of
these
is
exists
right,
so
have
a
separate
list
that
you
maintain
to
make
sure
that
these
these
exist
and
if
you
look
at
the
link,
the
link
command,
it
does
similar
things
where
it
like
tests.
It
does,
it
runs
the
create
or
this
is
the
create
command.
A
It
runs
the
create
command
and
it
tests
that
the
correct
files
are
present,
and
so
you'll
you'll
want
to
check
that
those
links
are
present
in
that
in
that
directory
that
you
pointed
at
and
let's
see
what
else
do
we
have
yeah
that
that
way,
we
we
know
that
we're
creating
all
the
correct
links
and
then
you're
also
going
to
want
to
let's
say
yeah
you're
going
to
compare
these
output
calls
to
make
sure
that
the
outputs
was
correct
right
and
see
because
yeah
you're
doing
the
you're
doing
the
print
right
so
you're
going
to
grab
it
just
this.
A
This
you're
gonna
have
a
very
almost
exactly
similar
comparison
situation
going
on
there
as
as
to
the
release
tests,
and
I
think
that
should
be
sufficient
here,
because
yeah
you'll
want
to
said
images
path.
Hater
editor,
does
it
recurse.
A
A
Yeah,
it
does
not
recurse,
I
don't
think
we
have
any
recursive
stuff
going
on
there,
but
you
know
we
may
eventually
okay.
This
is
the
kind
c
I
found
oh,
and
this
is
what
I
might
say
we
add.
We
want
to
add
this
as
a
failing,
so
so
need
to
update
fall
comparison
to
greater
than.
A
A
My
mouse
is
slowly
drifting
away.
Okay,
if
you
submit,
if
you,
if
you
find
an
issue,
that's
with
the
ci.
That's
not
yours!
The
way
that
we
can
communicate
that
to
everyone
else
that
we
know
that
something's
failing,
that's,
not
you
know,
caused
by
rpr.
It's
just
random.
A
Like
you
know,
some
kind
of
weird
network
issue
add
a
bug
that
says
kind,
ci
failing
and
then
that
way
when
other
people
are
saying,
I
wonder
if
this
is
my
fault
or
if
this
is
something
else
you
know,
then
they
can
go
and
look
at
this
tag
and
we
know
which
ones
might
be
might
be
might
be,
and
this
is
not
a
very
descriptive
issue
chase.
So
this
is
sometimes
the
npm
audit
tests,
for
example.
A
This
is
another
example
so
fail,
because
the
endpoint
for
the
npm
audit
api
is
down
and
that's
what
this
issue
is
about,
and
so,
if
you
see
that
in
the
tests,
you'll.
A
A
You'll
know
you
know
that
that
it's
not
your
fault
and
then
you
know
that
your
pr
will
be
merged
anyways.
If
you
don't
worry
about
fixing
that
okay
and
then,
let's
see
yeah,
so
the
images
to
copy
that,
I
think,
might
need
to
be
add,
target
to
args
and
validate.
A
A
Yeah
yeah
yeah.
The
thing
is,
this
is
a
context
manager.
So
when
you
enter
the
context-
and
this
is
the
tricky
part
about
it
being
both
a
you
know-
maybe
we
just
need
to
decorate
this
function
with
it,
because
if
we
decorate
the
function
then
at
all.
A
A
This
must
much
that's
the
function.
That's
called
on
entry,
yeah,
that's
sort
of
very
awkward
looking
you
know
I'm
and
I'm
I'm
thinking
is
there
a
way
that
we
can
just
this
is
their
cash
down
like
this
cash
download
decorator?
Is
there
a
way
that
we
can
also
make
it
used?
A
Yeah,
that's
annoying
that
they
can't
do
it
all
in
different
ways.
Maybe
it
doesn't
even.
I
Retrieve
but
it
would
not
check
for
the
file
yes
already.
A
Yeah,
you
know,
maybe
this
doesn't
even
need
to
be
a
decorator.
It's
like.
Why
does
it
need
to
be
a
decorator
now,
I'm
realizing,
because
we
could
just
you
know
you
could
one
could
just
call
it
at
the
beginning
of
the
of
the
function?
No,
it
needs
to
be
decorated
because
of
all
this.
Should
I
test
that's?
Why?
A
Because
we
have
this,
should
I
test
in
the
cache?
Basically
there's
this
is
what's
going
on,
so
yes,
yes,
varshney
a
while
ago
added
all
this
stuff
to
should
I
to
make
it
so
that
we
can
run.
Where
is
it
examples
to
make
it
so
that
we
can
run.
A
A
We
might
want
to
change
these
to
not
be
decorators,
because
it's
really
just
a
synchronous
operation.
There's
really
no
reason.
This
thing
needs
to
decorate
anything.
Was
there
a
reason?
Does
it
it
passes
it?
I
don't
underst
yeah,
you
know,
I
don't
know
why
we
made
this
decorator.
It
doesn't
need
to
be
a
decorator
that
needs
to
change
at
some
point.
So,
let's
just
change
that,
let's
make
a
note,
can
you
open
an
issue,
an
enhancement
issue
that
says
to
change
that
cache
download?
A
A
Okay
and
the
rest
of
this,
you
know
we'll
all
all
sort
of
cover
more
in
depth
offline.
So
I
think,
there's
a
there's
a
couple
more
things
here,
but
yeah
sweet
great
and
then
let's
see
10
40,
and
this
was
ci
job
to
validate
comment
message:
format.
Okay!
So
this
what?
What?
What
was
the
question
here.
E
I
A
Yep,
I
think
that's
a
great
idea,
and
I
think
that
we
should
you
know,
support.
So
basically,
it
has
to
match
if
you've
changed,
if
you've
changed
so
sub
module
file
must
match.
A
A
Blah
blah
blah
if
any
paths
under
sub
module
are
changed,
so
you
should
you
can
go
higher
right.
So
if
you
wanted
to
be
more
generic
right,
you
could
just
say
sub
module
right.
If
you
change
multiple
things
under
there,
if
you
only
change
one
thing,
then
you
should
be.
You
know
specific
here.
So
if
you
only
change
sub
module,
slash
file,
then
we,
you
know
we
should
see
sub
module
colon
file.
If
use
change
said
module
file
and
sub
module
util,
then
we
could
see.
A
A
Yes
well,
and
so,
if
you're
changing
change
log,
then
you
know
change.
Log
is
your
path
right,
so
if
if,
if
you're
changing
something
at
the
root,
then
that
is
your
identifier
right,
then
there's
also
special
cases
like
if
we
change
things
for
all
of
the
plugins,
so
special
cases
to
support,
and
let's
just
look
at
so
git
log,
one
line.
A
How
do
I
get
just
messages?
Okay,.
E
A
I
you
know,
look
just
look
at
this.
You
know
look
at
the
log
for
sort
of
examples
and
and
we'll
go
from
there,
but
yeah
here
sort
unique
okay.
So
what
do
we
have?
Yeah
util
entry
point,
I
think.
Sometimes
we
reference
by
class
name,
yeah
change,
log
ci
base
yeah.
So
if
it's
in
so
yeah
okay,
so
if
and
if
it's
in
the
root
directory,
then
we
support
it
as
if,
if
it's
in
so
if
it's
in
dff
from
also
the
way,
I
usually
hear
people
think
referencing
things
to
me.
A
C
A
This
might
be
dfml,
and
then
we
have
dfml
slash
dfml
right,
which
is
actually
the
main
plugin.
So
if
it's
in
the
main
plugin
directory
right,
that's
how
we
usually
reference
it
is.
Is
you
know,
as
if
you
know
we
don't
say,
dfml
colon
based
right?
We
just
say
base,
because
this
is
the
main
plugin
right
and
when
we
start
to
split
out
all
the
other
plugins
that
will
make.
You
know
a
lot
more
sense.
So
and
then,
let's
see
so
yeah
ci
docs.
C
A
Yep
yep,
but
generally
you
know
it
follows
that
that
file
structure
and
then
for
plugins.
We,
you
know
we
prefix
with
the
plugin
directory
so
and
then
oh
yeah.
If
we
do,
if
we
do
modifications
to
all
the
plugins,
we
we
just
put
plug-in
yeah.
So
you
this
is
probably
a
good
command
to
to
get
a
good
idea
of
what
what
kind
of
stuff
we've
had
and
we
want
to
so
change.
Log
plugins,
oh
clean
up
change,
log.
A
Okay
and
then
this
will
help
you
get
an
idea
all
right,
all
right
great.
So,
let's
I
want
I
want
to.
I
want
to
get
the
I
want
to
talk
about.
I
want
it.
I
want
sutancia
to
show
everyone
the
ice
cream
demo
before
we
get
to
to
because
we
could
who
knows
what
will
happen
with
the
data
flow,
since
that's
some
debugging
and
this
is
it
show
and
tell
so,
let's,
let's
see
what
we've
got.
B
So
as
we
like
discussed
previously,
so
what
I
have
done
right
now
is
so
I
have
so.
I
have
actually
got
the
name
of
the
cities
and
their
respective
urls
and
their
respective
hash
values
and
similarly
for
all
the
other
cities
and
for
the
population
data
set.
So
I
actually
went
through
that
data
set
and.
A
B
So
I
found
that
those
data
sets
and
their
hash
value
so,
like
I
have
this
lookup
temperature
method,
so
it
will
take
the
city
month.
It
will
just
download
that
particular
city,
it
will
download
it
to
the
cache
directory
and
then
it
will
return
the
value
which
we
want
to.
D
B
Application
we
are
also
having
another
cache
directory,
and
so
here
what
I
was
thinking
like
we
can
use
pandas
for
it
to
return
the
values
for
the
city,
because
in
the
csv
what
is
happening
is
we
have
the
city
name
and
the
state
is
there?
A
I
mean,
I
think,
if
you
just
pass
this
the
file
name
to
load
the
load
function,
it
will
instantiate
the
csv
source
and
load
all
the
records
for
you,
because
using
pandas
to
read
a
csv
file
seems
like
a
little
bit
overkill
to
include
a
third-party
module
for
that,
since
we
already
have
support
for
that,
because
you
know
like
what
what
if
you,
if
you
were
to,
I
think,
if
you
just
said,
load
file
path.
It
should
load
the
csv
file
with
the
csv
source
and
return
all
the
records
right.
B
So
actually,
the
problem
here
is
this
part.
B
E
B
But
the
thing
which
I
want
to
like
like
say,
is
like
for
this
value.
We
are
actually
taking
the
city
itself,
but
given
the
city,
we
will
not
be
able
to
find
what
is
the
state
value.
A
B
A
No,
I
think
that
was
in
reference
to
you
know.
We
had
we'd
gone
down
that
path
with
with
the
csv
sources
and
then
the
json
source,
and
we
realized
okay.
If
you're
going
to
use
json
source,
you
don't
need
to
use
the
json,
there's
no
point
in
using
the
json
source.
If
it's
in
a
different
format
right,
it
makes
sense
to
leverage
the
csv
source,
because
it's
already
in
csv
format,
right
so
so
and
and
yeah.
A
So
you
you
know
you
could
use
the
pandas
read
data
frame
stuff,
but
then
you
have
to
import
pandas.
I
don't
think
there's
any
need
to
to
to
to
start
including
third
party
modules
in
this
demo.
I
don't
at
least
see
that
at
this
point,
since
we
already
have
the
ability
to
read
the
csv
files,
you
know.
B
B
B
A
A
A
D
A
B
So
actually,
just
look
up.
Temperature
thing
will
actually
work.
B
So
like
in
the
inputs,
what
I
have
given
is
the
city
and.
J
B
A
B
So
this
actually
value
is
directly
this
value.
A
All
right,
fantastic,
that's
great,
that's
great,
very
good,
very
good!
Let's
see
it
seems
it
seems
a
little
bit
slow.
Is
it
it
running?
It's
running
the
cache
downloads
right
I
mean
it.
Let's
see.
A
B
A
B
Like
one
more
thing
to
ask,
so
in
the
data
flow,
what
happens
when
we
raise
an
exception.
A
A
All
right
city,
not
okay,
yeah,
so
and
and
yeah,
so
try
actually
try.
That
say
just
say
if
you
know
just
just
raise
the
exception,
remove
the
if
and
raise
the
exception.
So
we
can
see,
see
it
happen
or
you
could
pass
a
different
city
name.
You
know.
A
Yeah
great
there
we
go
and
then
it'll
dump
the
I'd
say:
operation
exception
yeah,
let's
tell
you
which,
which
what
the,
what
the
variables
that
were
it
was
called.
What
so?
Yes,
it
will
raise
the
exception
good.
If
you
pass
strict
equals
false,
it
will
not
raise
the
exceptions.
B
B
So
using
these
two
data
flows,
we
are
actually
going
to
generate
a
data
set
which
will
be
training,
so
I
have
to
ask
like
how
how
do
we
go
ahead
with
generating
the
data
like.
A
Yeah,
okay,
so
the
thing
was
we
were
going
to
what
were
we
going
to
do?
I
think
you
just
need
to
basically
pick
just
pick
the
cities
and
the
just
just
you
just
pick
the
cities
in
the
months
right
so
for
just
do
several
cities
and
several
months
and
each
month
for
that
city
right
and
then
all
of
a
sudden
you'll
have
your
data
set.
Because
what
is
what
is?
What
was
that?
Let's
see
what
was
the
problem
description
again.
A
A
Oh
yeah,
so
we
need
to
generate
the
sales.
Okay,
yeah,
that's
right,
yeah
yeah!
We
didn't
need
the
sales
yeah,
because
this
is
like
you
know.
This
is
what
we
know
is
that
ice
cream
sales
go
up
in
the
summer
right
when
it's
hotter
out
people
buy
more
ice
cream.
We
may
not
have
data
for
that,
but
but
I
think
that's
a
a
known
fact.
So
so
so
we're
doing
this
demo
under
that
assumption
right.
A
So
we're
just
going
to
add
some
data
right
that
that
roughly
corresponds
to
that
and
and
so
what
you
can
do,
is
you
can
right?
So
the
thing
is
the
machine
learning
will
predict
the
the
value
of
the
sales
right.
You
can
just
throw
in
numbers,
or
you
can
write
some
kind
of
function
right.
You
can
run
the
data
flow
and
then
write
your
ice
cream
generator
function
right,
which
is
basically
like
you
know
you
could
do
some
kind
of
linear
function
based
on
temperature
right
and,
let's
see
yeah
so
yeah.
A
You
could
just
do
some
kind
of
linear
function
based
on
temperature
or
you
could
just
go
and
go
through
your
csv
file
list.
Every
city,
with
every
month
that
you
have
in
there
you
know
and
and
just
write
some
values
right,
so
it's
sort
of
whatever
you
wanted
to
do
right.
If
you
write,
you
can
write
a
linear
function
that
generates
it
and
then
basically,
what
you
would
be
doing
is
you'd
be
training
a
model
to
predict.
You
know
to
you'd,
be
training
a
model
to
to
understand
what
function.
That
was
right.
A
That's
essentially
what's
going
on
here,
because
we're
just
making
up
the
data,
but
you
know
the
the
principle
is:
is
to
show
how
to
do.
You
know
additions
of
data
right
to
an
existing
data
set
so
so
yeah
you
could
write
an
operation
to
you
could
write
an
operation
that
that
is
just
a
linear
function.
You
know
like
blank
time
y
y,
mx,
plus
b
right
and
then
you
can
use
the
simple
linear
regression
model.
A
That's
actually
perfect,
because
because
that's
one's
in
the
standard
library,
so
the
whole
thing
doesn't
need
any
plugins
at
this
point,
which
is
great,
which
means
it
can
fall
under
the
fast
test,
fast
stock
test.
Let's
see
what
was
he
going
to
say
on
that?
Oh
yeah,
if
you
do
that,
what
you
can
use
is
the
merge
command,
so
you
can
use
the
merge
command
and
okay,
it's
the
merge
command.
It's
not
well
documented.
Is
it
we
changed
it,
and
so
we
removed
the
documentation
and
we
forgot
to
add
it
back.
A
But
if
you
do
dfml
merge,
oh
god,
okay,
I
have
a
giant
strike
trace
trace
back
from
my
contextual
version
conflict.
So
if
you
use
the
dfml
merge
command
on
the
command
line,
you
can
basically
take
a
source
data
set
and
destination
data
set
and
you
can
use
the
dataflow
preprocessing
source
and
you
can
run
this
your
lookup
temperature
and
lookup
population.
A
And
then
you
add
in
your
your
you
know
your
essentially
your
function
generator
and
then
you
make
the
output
source,
another
csv
source,
and
so
you
basically
you're
taking
the
data
flow
so
as
you're
turning
it
into
you
know,
use
the
data
flow
source
to
run
the
data
flow
that
you
have
right
and
then
add
in
one
more
operation.
A
Then
that
operation
is
your
little
your
linear
function
and
then
output.
You
know
grab
the
output
of
those
those
records
that
you're
doing
through
this
modification
and
dump
them
to
a
different
csv
source.
You
know
so
that's
the
destination
in
the
merge
command,
and
now
you
have
your
example.
Data
set
right,
yes,
cool.
This
is
gonna,
be
sweet.
This
is
yeah,
that's
gonna,
be
sweet
and-
and
let's
see
yeah,
if
you
wouldn't
mind,
if
you
want
my
document
that
command
and
then
we
can
have,
we
can
add
that
command.
A
Actually,
you
know
we'll
we'll
we'll
add
a
variation
of
that
command
to
the
to
the
cli
reference,
the
quick
reference
of
the
cli
page
for
for
how
to
do
the
merge
command
since,
like
I
said
that
the
reason
why
the
docs
aren't
there
you
you
should,
I
think
I
think
you
won't
have
too
much
trouble
figuring
out
what
what
you
know
figuring
out,
what
to
do
there,
because
it's
it's
basically
just
you
know
it's
the
same
syntax
we
always
use.
You
know,
define
your.
A
You
can
pull
from
the
example
that
does
the
data
flow
source,
but
yeah
it
should
be
good,
great,
great
work.
I'm
excited,
I'm
excited!
This
is
great.
This
is,
I
think,
yeah
one
of
the
first
ones.
This
is
the
first
thing
we're
doing
that's
actually
adding
do
showing
the
modification
to
an
existing
data
set.
You
know,
rather
than
entirely
creating
a
new
one,
so
this
would
be.
This
is
a
great
example,
good
job,
good
job,
okay,
so.
A
So
I'm
just
taking
notes
here,
so
you
can
use
the
merge
command
with
data
flow
progressively
source
to
take
the
lookup
operations
and
add
in
an
operation
with
man.
You
know
this.
This,
the
command
line
argument
says:
config
files
is
going
to
be
really
great.
We
need
to.
We
need
to
go
get
that
done,
because
that
saves
so
much
copy
pasting
of
giant
command
line
things
if
you're
in
the
terminal-
and
you
use
a
terminal
editor
if
you
do
control
x
or
what
is
it
okay,
crap
crap
crap.
Now
I
forgot.
A
Oh
no
ctrl
e
ctrl
e
will
pop
open
your
editor
and
you
can
write
your
command
line.
You
know
whatever
command
line,
you're,
trying
to
type
that
might
belong
into
your
text,
editor
and
then
it'll
just
run
it.
So
if
you
have
an
environment
variable
set,
if
you
have
the
all
caps,
editor
environment
variable
set
and
you
type
at
a
blank
prompt
or
actually
at
any
prompt,
even
if
you
have
something
existing
or
no
wait.
Okay!
No,
not
if
you
have
something
existing.
If
you
type
control
control,
are
you.
A
Serious,
I
think
I,
of
course,
as
soon
as
I
go,
to
communicate
what
the
key
combination
is.
I
forget
it.
Well,
I
swear
it
was
control,
you
there's
a
way
to
do
it,
and
then
you
can
write
long
command
line
things
so
use
merge
coming
out,
you
df
to
take
a
look
up
operations
and
in
an
operation
with
a
linear
function
that
will
generate
the
sales
data.
D
Function
essentially,
but
we
don't
need
to
talk
about
this.
A
The
example
is
about
so
yeah
we
don't.
We
don't
need
to
talk
about
this
in
the
demo.
We
can
mention
how
we
did
it
in
like
a
you
know,
in
an
aside
after
we
can
mention
how
we
did
it
in
another
document,
but
not,
but
that
will
distract
from
the
demo
if
we
include
it.
A
The
main
example
document
all
right,
great
anything
else
on
that
suit.
Aren't
you.
B
A
And
I'm
gonna
try
to
rebase
the
accuracy
stuff
with
master
to
see
what
happens
there
and
then
I
think
we
will
be
on
to
phase
eight.
So
I
just
wanted
to
give
that
a
shot.
A
Yeah
yeah
and
I
believe,
that's
sort
of
just
updating
a
lot
of
our
examples
which
luckily
we
all
we
have
most
of
them
tested
and
I
think,
there's
a
they
think,
there's
an
issue
that
documented,
which
ones
are
auto,
have
automated
tests
and
which
ones
don't
so.
Hopefully
we
will
catch
it
all.
We
will
see,
though,
all
right
yeah,
that's
a
big
one,
all
right
so
shaw
data
frame,
pr.
J
Yeah
I'll
just
present
my
screen.
If
we
could
just
go
through
a
test
once
it
would
be
great.
A
Yeah,
okay,
sounds
good
and
and
who
just
just
joined.
I
believe
this
is
who
was
on
the
mailing
list.
A
Yeah,
you
use
user
user.
Whatever,
who
are
you.
A
Well,
you
can
speak
up
when
when,
when
shaw
is
down
here-
and
we
will
get
to
your
your
question
on
that-
you-
you
posted
on
the
mailing
list,
all
right.
E
E
J
This
is
a
test
file
and
I've
written
the
tests.
There's
only
a
couple
of
tests
and
they're
both
running
well
right
now,
great
I'll
just
go
through.
J
A
Yeah
all
right,
and
what
can
you
turn
logging
on
and
let's
see
what
happens
with
logging,
so
just
prefix
the
whole
command
so
hit
the
up
arrow
and
go
to
the
very
beginning
and
then
just
type
all
caps
logging
equals
lowercase,
debug
so
very
beginning,
so
we're
going
to
set
it.
This
is
actually
this
is.
This
is.
C
A
Set
a
temporary
environment
variable
for
the
lifetime
of
the
command
that
you're
about
to
run
as
you
prefix
it
with
whatever
environment
variables.
You
want
so,
okay
great.
So
now
we
get
to
see
the
logging
we
instantiate,
the
data
frame
source,
feature
columns.
You
know
we
enter
the
context.
This
is
the
double
context,
entry
thing
and
then
we
enter
the
so
into
the
main
object
and
we
enter
the
context.
A
J
Okay,
so
for
the
first
test,
what
we're
doing
is
we're
trying
and
saving
some
records
in
the
data
frame
source
and
checking
that
works
well
for
the
second
one,
we're
just
checking.
If
that
particular
feature
is
there
in
the
record
or
not?
Is
there
in
the
rate
of
game
or
not
for
a
particular
record.
A
Okay,
so
zero,
okay,
so
record
zero
record.
One
and
record:
let's
see
my
addict
okay
record,
one,
okay,
so
the
data
frame?
Okay.
So
let's
not
leave
this
data
frame
over
the
in
the
global
namespace
because
I
think
that
yeah,
that
makes
it
confusing
which
test
does
what
to
it
right.
So,
let's
instantiate
it
per
test
case
and
then
it
becomes
really
clear.
What's
going
on
and
let's
also
make
feature
calls
just
be
features:
lowercase
features
yeah
and
then,
let's
you
know,
I
think
you
know.
A
There's
there's
there's
one
for
predictions
too,
and
let's,
let's
so
yeah
just
predictions
and
features
instead
of
feature
calls
and
and
and.
J
Actually,
I
actually
thought
we
could
do
away
with
the
predictions
argument,
because.
J
Yeah
because
it
was,
I
could
do
the
same
thing
with
the
features
argument
only
so
I
thought
we
could
do
it
with
the
prediction
argument.
A
A
Okay,
so
what's
what
is
easier?
I
guess.
Okay,
we
also
thought
about
remember
how
we're
going
to
implement
that.
Maybe
that
read
html
and,
like
you
know
you
can,
you
can
say,
read
html
from
from
a
pandas
data
frame
source
and
it
will
go
and
grab
an
html
table
and
turn
it
into
a
source,
and
so
I'm
thinking
about
that,
as
maybe
a
common
use
case.
That
will
be
an
extension
of
this,
and
if
we
are
thinking
about
that,
then,
if
we're
thinking
about
that,
then
would
it
be?
A
Would
people
be
specifying
in
that?
In
that
case
people
probably
say
you
know
they
want
every
column
right
and
they
don't
want
to
go
and
write
every
column
in
the
features
right.
They're
like
this
is
my
features.
Right
is
this
table
and
they
maybe
have
one
prediction
column
in
the
table
right.
So
from
that
standpoint
it
makes
sense
to
define
the
prediction
columns,
because
you
have
to.
G
A
J
A
No,
I
don't
have
any
inputs,
just
let's,
let's,
let's
make
sure
this
test,
it
was
hard.
I
can't
follow
the
test
with
the
fact
that
the
global
namespace
is
active
or
the
the
data
flow
is
in
the
global
namespace.
So
please
run
the
the
style
stuff
and
please
please
move
you
know,
make
sure
the
data
frames
get
instantiated
on
a
yeah
and
then,
unless
you're,
calling
that
setup
data
frame
or
setup
source
function
somewhere
else.
A
Let's
just
get
rid
of
that,
because
if
we're
not
calling
it
multiple
times,
you
know,
let's
see
either
we
should
be
calling
it
calling
it
in
both
test
cases
or
we
should
not
have
it
at
all.
I
don't
know
because
that
also
you
know
in
test
cases
it's
it's
yeah,
okay,
so
yeah.
So
let's
do
the
predictions
thing
and
let's
run
the
style,
and
that
sounds
good.
A
A
Okay
cool.
Can
you
spell
your
name
for
me
for
the
meeting
minutes
real,
quick.
K
A
A
Yeah
yeah
great
so,
let's
see
support,
support,
archive
storage
and-
and
I
think
you
know
I've
had-
we've
had
trouble
trouble.
I've
had
trouble
getting
back
to
you.
Sorry
if
you
chat
on
gitter.
That
is
usually
the
best
way
to
reach
everyone,
that's
sort
of
where
we
all
talk
the
most
just
just
as
a
future
future
way
to
get
less
less
less
long
response
time.
So
sorry
about
that,
it's
just
hard
to
track
things
that
that
ends
up
being
the
best
place
to
track
conversation.
D
A
We
want
to
talk
about
the
support.
Let
me
write
down
the
stuff
for
the
data
frame
test,
so
we're
going
to
accept
prediction,
columns
that
store
predictions
in
config
instead
of
features
and
then
we're
going
to
move
the
data
frame
outside
of
the
oh,
and
I
just
realized
we're
going
to
get
into
a
tricky
situation
with
this
move.
The
data
frame
outside
of
the
tests
error
of
the
move
to
test
scope
rather
than
global
scope.
A
Okay,
okay
support,
so
support
archive
storage.
So
at
a
high
level
the
what's
going
on
here
is
we
want
to
so
right
now.
A
Yes,
okay,
great,
can
you
guys
see
all
right
so
so
at
a
high
level?
What's
going
on
here?
Is
all
the
all
the
models
take,
a
all
the
models
take
a
where's,
the
plugins
page.
A
So,
for
example,
so
the
models
all
take
a
directory
which
they
store
their
con,
that
that's
where
they
store
the
saved
model
to
that's
pretty
much
the
way
all
of
them
work.
So
the
goal
of
this
project
is
to
make
that
into
and
take
to
basically
do
a
massive,
find,
replace
and
say
directory
is
now
location.
So
now,
instead
of
all
the
models
have
a
property
named
direction.
Now
they
all
have
a
config
property
named
location
and
that
location.
A
Now,
since
it's
no,
you
know
since
we're
so
first
first,
first
step
change
the
name
next
step
change,
the
functionality
right,
and
so
now
the
the
the
goal
is
to
make.
You
know,
do
the
find
replace
make
sure
everything
works
right,
run.
All
the
tests
make
sure
everything
still
works
then
go
through
and
say
all
right.
A
How
do
we
start
making
it
so
that,
if
a
location,
that's
given
is
like
a
zip
file
or
something
like
that
that
the
zip
file
is
opened
and
extracted
to
a
temporary
directory,
and
then
the
model
uses
that
temporary
directory
and
then
when
the
model
is
done
with
whatever
it's
doing,
then
the
contents
of
that
temporary
directory
gets
packed
back
into
the
zip
file
that
was
referenced
by
location
and
that's
essentially
the
entire
project,
and
you
know
so
there's
some
some
stretch
stuff,
that's
defined.
A
On
top
of
this,
you
know
where
you
know,
if
you
get
done
with
all
of
that,
then
then
you
want
to
start
thinking
about
okay.
Well,
how
would
we
make
it
so
that
this
might
interact
with
you
know
something
that's
stored
at
like
an
http
address
right?
Would
we
go
and
download?
You
know
that
model
etc,
right
and
and
think
about
okay?
Well,
now?
What?
If?
What?
If
you
know,
how
could
we
even
use
it?
This
is
where
I'm
going
with
this
is:
could
we
use
a
data
flow
to?
A
Could
we
use
a
data
flow
to
to
specify
the
could
the
location
be
a
data
flow
and
if
it
is
a
data
flow?
How
would
that
work
if
we're
packing
it
back
up
too?
So,
because
if
you
specify
a
data
flow,
then
you
can
do
much
more
complex
things.
A
For
example,
you
know
you
could
do
you
could
integrate
authentication
to
some
kind
of
arbitrary
third-party
resource.
It
doesn't
constrict
the
code
to
being
you
know,
it's
it's
it's
much
more
dynamic.
So
did
you
have
any
questions
on
any
of
that.
K
A
K
A
A
One
second:
I'm
writing
that
in
the
notes
directory.
A
Okay
further
reading
for
this,
so
I
think
that
that
there
was
some
stuff
in
the
this
is
the
auto
ml
project
ideas,
archive
storage
for
models?
Okay,
so
I
think
there's
some
yeah
there's
some
links
here.
So
basically
I
would
look
at
the
model.
Plugins
and
just
you
know
play
around
so
play
around
with
implementing
so-
and
this
is
yeah
so
play
around
with
the
model
tutorials
go
through
the
model
tutorial,
because
the
model
tutorial
shows
you
well,
you
know
it
doesn't
really.
A
Does
it
really
yeah,
it
doesn't
really,
unfortunately,
show
the
a
enter
a
exit
method.
What
would
be
good
for
that?
So
this
simple
source
tutorial
or
the
complex
source
tutorial
might
be
actually
a
good
thing
to
look
at
too,
because
it
has
these
a
inter
and
a
exit
methods,
which
is
where
you
know
this
type
of
thing
might
happen
right,
because
so
all
of
the
all
of
this
all
everything
in
dffml
has
this
double
context:
entry
pattern,
and
so
this
is
a
good
document
to
go
over,
because
this
sort
of
shows
you.
A
What
might
be.
You
know
what's
happening
here,
and
so
this
is
with
an
example
with
a
with
a
simple
source
and
a
simple
model,
and
now,
when
we
implement,
you
know
the
way
that
these
are
implemented
and-
and
you
may
want
to
just
dive
into
the
dfml-
slash
model,
slash
model
code,
and
that
would
be
a
good
thing
to
look
at.
Maybe
you
know
look
at
the
look
at
code
for
simple
model
in
dfm,
slash
model,
slash
model
you're,
primarily
interested
in
the
a
inter.
A
A
A
Etc,
so
because
these
get
called
every
so
this
double
context,
entry
pattern
that
happens.
C
A
A
with
block
in
python,
it
triggers
the
inter
method,
the
underscore
underscore
inter
method.
If
you
do
an
async
with
block
it
triggers
the
a
enter
method,
and
so
that
is
basically
that
allows
us
to
do
setup
and
tear
down
around.
You
know
like
a
con
yeah,
so
you
can
do
setup
and
tear
down,
and
I
would
also
look
at
the
context
manager
page
in
the
documentation
or
context
lib,
so
context.
A
And
it
lets
you
do
all
this
context
manager
stuff.
So
I
would
check
this
out.
This
is
fun
stuff,
and
this
will
tell
you
more
about
you
know,
sort
of
yeah.
You
know
context
managers
and
what
what
you
can
do
with
them
so
and
that
will
tell
you
also
you
know
what
you
know:
how
does
this
stuff
work,
because
this
is
you
know
this?
This
is
how
python,
how
you
do,
how
you
do
setup
and
teardown
in
python,
essentially
and
so
play
with
the
context.
A
Managers
stuff
from
the
from
the
main
python
documentation,
understand
how
this
stuff
works
and
then
and
then
look
at
the
a
n
or
a
exit
methods
and
think
about
okay.
How
do
I,
if
I
had
a
zip
file
because
you'll
see
this
self-config
directory
in
the
simple
model
and
you
could
think
about?
Okay,
if
I
had
a
zip
file
there?
How
would
I
turn
that?
A
If
say,
how
would
I
turn
if
someone
you
can
look
at
the
way
it
is
now
at
self.config.directory
in
a
inner
method
of
simplemodel,
and
you
can
say:
okay,
I
think,
there's
something
that
creates
the
directory.
If
it
doesn't
exist,
you
could
say
hey
what
happens
if
I,
if
I
provide
a
zip
file
path
to
an
existing
model
and
you
can
start
changing
the
code
right
now,
for
example,
like
the
simple
model
code,
you
could
start
changing
the
test
case
for
the
slr
model
and
you
could
mod.
A
You
could
write
some
test
cases
which
provide
a
directory
which
is
a
zip
file
or
a
tar
file,
and
then
you
can
start
working
with
that.
A
intro
method
in
model
py
of
simple
model
and
modifying
the
code
there.
That
was
creating
the
directory.
If
it
not,
it
didn't
exist
to
instead
now
trying
to
use
the
zip
file
module
or
the
tar
file
module
to
extract
that
that
that
directory.
A
That
archive
that
you
gave-
and
you
can
you
know-
create
the
archive
by
running
the
model
and
zipping
up
the
directory
or
tarring
up
the
directory,
and
that
that's
how
you
can
get
your
essential.
You
know
your
your
thing
to
work
with
there
to
start
with.
Does
that
make
sense
sound
like
a
good
plan?
Yes,
it's
sweet,
sweet,
great
and
then
yeah
like
getter
gear
is
probably-
and
let
me
just
say
this
again
for
everyone,
because
I've
been
realizing.
This
getter
is
probably
the
best
way
to
reach.
A
You
know
me
and
default
your
your
fellow
students
and
and
other
mentors,
because
everybody
is
on
there
and
if
one
of
us
doesn't
see
it,
someone
else
can
see
it
and
there's
been
a
lot
of
great
discussion
happening
on
there
logarithm
across
people.
So
so
that's
really
great
thanks.
Everyone,
for
you
know
it's
always
good
to
help
each
other.
This
helps
things
move
forward
faster,
well,
cool
is,
does
anybody
have
any
other
opens
today?
Did
we
get
everybody?
Does
anybody
have
anything
else.
A
So
that's
on
the
milestones
page,
so
the
the
milestones
page
says
that-
and
I
believe
it's
this
weekend.
Where
was
that
milestones
so
yeah
drafts
do
so?
Basically,
if
you
get
the
drafts
to
the
to
to
the
gator
channel
by
by
april
3rd,
then
we
can
guarantee
that
we
can
give
you
some
sort
of
feedback
on
your
draft
before
the
proposals
are
due.
But
if
you
don't
get
it
to
us
by,
then
you
know
we.
We
can't
can't
guarantee
that
we'll
have
time
to
look
at
it
before
before
drafts
are
due.
A
So
if
you
can
get
to
us
at
any
time,
you
know
before
then,
but
that's
just
the
last
date
and
and
if
you
do
get
at
us
after
that
date,
then
you
know
we'll
we'll
we'll
we'll
try
to
get
to
it,
of
course,
but
no
guarantees
right,
we'll
guarantee
that
we
get
to
it.
If
we
get
it
to
us
on
that
date,
great
any
other
questions
from
anyone.
B
One
last
question
from
my
side
like:
I
am
targeting
the
idea
for
auto
ml
thing
and
like
how
shall
I
start
with
it
like?
Shall
I
create
some
issues
or
what
kind
of
issues
can
I
create
so
that
I
could
start
with
this
automatic
project?
I
have
gone
through
the
tutorials,
which
are
mentioned
in
the
ideas
portion.
A
Well,
I
think,
sort
of
starting
the
pro
you
know
starting
to
to
to
prototype
what
it
looks
like
right,
because
the
thing
is
you're
going
to
be
implementing.
You
know
how
how
to
start
so
you're
going
to
be
implementing
a
model
right
and
that
model
is
essentially
going
to
sit
on
top
of
other
models,
and
if
you
looked
so,
we've
got
some
stuff.
If
you
look
at
model
and
then
load
models
dynamically,
so
this
is
kind
of
you
know
how
and
actually
I
think
they
realize
this
needed
to
change
the
other
day.
A
So
looking
at
at
this
so
start
writing
a
model
load.
Other
models
model
dynamically
based.
Actually
you
know
yeah
load,
other
knowledge
to
interface.
A
All
right,
you
can
load
all
bypassing
nothing
to
do
load
or
a
specific
list
by
passing
or
a
specific
model.
By
passing
its
entry
point
string.
A
So
if
you
see
any
of
these
models,
if
you
see
vw
model,
you
can
pass
it
to
load
here.
Well,
I
can't
paste
it
on
that.
You
can
pass
it
to
load
and
then
it
will
load
that
vw
model,
if
that
package
is
installed,
so
you
can
play
it
play
around
with
loading
a
few
models
and
if
you,
if
you
don't
pass
anything,
this
is
why
this
needs
to
be
changed.
A
If
you
don't
pass
anything
it'll
load,
every
model
that
is
installed
and
so
and
this
is
that
should
be
a
reference,
but
it's
not
yeah
it'll,
and
so
you
play
around
with
that
and
and
figure
out,
okay.
Well,
how
would
I
dynamically
instantiate
models
and
then
what
you
know?
What
are
the
steps
that
that
I
need
to
go
through
to
create
automl
right
like
how
do
I
do
parameter
tuning
right?
A
How
would
how
would
I
know
which
parameters
that
I
can
tune
and
which
ones
that
I
can't,
because
you
gotta
you
have
to
think
about
for
all
of
these
projects
right,
the
entire
code
base
is
is,
is:
is
you're
able
to
change
it
right
so
for
anybody
who's?
Thinking
about
doing
a
project
right,
you're,
you
you
have
complete
license
to
change
anything.
Should
it
need
to
be
changed
right.
So
if
you
think
that,
for
some
reason
you
know,
if
you're
like
okay
wait,
we
need
more
information
on
which
parameters
are
tunable.
A
We
need
to
go
through
and
add
that
type
of
thing
to
each
each
each
field.
Then
you
that's
that's
you
know,
that's
that's
up
for
grabs
right,
but
just
if
you're
gonna
right,
if
you,
if
you're
going
through
and
you're,
designing
your
project
and
you're
like
oh,
we
need
more
information
from
the
models
right
like
I
need
to
know
which
models
are
regression
or
richer
classification
right
to
be
able
to
choose
what
I
want
for
this
problem,
then
you're
going
to
find
out
that
okay,
we
need
to
go.
Have
some
self-identification
in
the
models.
A
For
example,
himachal
found
that
he
needed
to
have
the
length
of
a
source
when
he
was
doing
his
project,
and
so
he
implemented
the
you
know.
A
He
said
that
we
need
to
go,
implement
a
length
method
on
all
the
sources,
and
so
so
he
did
that,
and-
and-
and
you
know
now
when
now
we
need
to
go-
implement
the
length
method
on
all
the
sources
right
because
things
like
you
know,
like
you
know,
if
you,
if
you
need
to
find
all
possible
values
of
a
feature
for
example,
then
you
have
to
go
through
all
the
all
the
different
records,
so
we
found
that
there
are
definitely
operations
that
require
that
we
go
through
the
entire
data
set
to
to
enumerate
possible
values
and
things,
and
so
we've
discovered.
A
Okay,
we
need
to
change
the
apis
of
everything
in
here
right
all
the
sources
right
you
may
need
to
change,
find
that
you
need
to
change
something
about
all
the
models
right,
just
figure
out
what
those
things
that
you
need.
A
You
will
try
to
come
up
with
as
clear
as
a
path
if
possible,
right
because
I've
said
this
to
a
few
people,
but
if
you're
designing
your
schedule
right,
this
is
the
schedule
that
you're
you're
being
graded
on
right,
so
design
a
schedule
fit
you
want
to
do
you
want
to
do
a
good
amount
of
work
to
figure
out
what
is
a
schedule
that
you
would
want
to
be
held
accountable
to
right
because
you're
you're
you're
designing
your
own
accountability
schedule
here,
so
so
make
sure
that
it's
realistic
and
make
sure
you
know
what
kind
of
work
you're
cutting
out
for
yourself
so
yeah
any
any
other
questions
on
anything
like
that
or
anything
at
all.
B
H
Yeah,
I
have
a
question
like
about
the
load
function.
H
I
think
it
only
allows
you
to
either
send
a
single
either
load
a
single
model
or
all
models.
A
A
I
think
yeah,
that
might
it
might
be
good
if
we
started
to
allow
multiple
models.
That
would
probably
be
a
good
idea.
I
I
don't,
I
don't
think
we
have
that
implemented.
So
you
know
that
might
be
one
of
the
changes
one
wants
to
make.
H
Yeah
one
more
thing:
I
have
a
pending
pr,
if
you
could,
you
know,
oh.
A
Yes,
the
class
yeah
that
guy
yeah
I
haven't-
I
haven't
quite
gotten
to
that
yet
and
so
that
and
the
light
gpm
are
both
still
on
my
docket.
So,
and
there
was
one
more
thing
I
was
going
to
do
after
this
meeting.
Does
anybody
remember
what
was
that?
Okay.
A
Okay,
all
right!
Well,
I
think
I
dropped
something
else,
unfortunately,
but
so
light
gbm
and
we'll
see
h2o
model
and.
A
Confidence,
jesus
all
right.
Okay,
let
me
write
those
down.
I
cleaned
up
my
desk,
so
my
sticky
notes
are
gone.
I
got
to
put
them
all
back
great
well!
Thank
you
thanks.
Everyone
have
a
good
one.