►
From YouTube: Weekly Sync 2020-07-28
Description
Meeting Minutes: https://docs.google.com/document/d/16u9Tev3O0CcUDe2nfikHmrO3Xnd4ASJ45myFgQLpvzM/edit#heading=h.rd7xvcwokf5w
A
So
I
need
a
review
on
the
data
floating
pr.
It's
done
and
I
was
going
through
the
chatbot
example
in
a
new
which
to
land
and
it
doesn't
detect
the
initials.
So
it
needs
that
to
load
the
secrets.
So
how
do
you
install
that.
C
A
E
Yeah,
I'm
still
waiting
on
your
reviews
for
my
full
requests
same
as
last
the
last
meeting,
and
I'm
also
catching.
B
B
B
E
B
Yeah
I
mean,
I
don't
think
that's
I
mean
I
think
I
think,
being
consistent
with
what
you
know.
People
think
the
name
is
is
good
to
do
right,
so
they've
named
it
with
underscores,
then
we
might
as
well
be
consistent
right.
E
Also
like
have
you
had
the
chance
to
like
work
on
the
layer
support
tutorial.
You
were
talking
about
last
meeting.
B
All
right,
let's
see,
okay,
so
we've
got.
B
E
B
G
Tensorflow
2.3
is
installed,
but
it's
not
supporting
numpy
1.191.
B
C
B
Okay,
so
you
will
update
the
file
classification
example
after
that
and
then,
let's
see.
E
Not
no
just
I'm
just
waiting
on
that
layer,
support
example.
Then
I'll
move
forward
with
the
all.
E
B
H
E
Yeah,
I'm
really
excited
to
add
that
object,
detection
with
bounding
box
examples.
E
Funny,
I'm
still
learning
this
stuff
right
now.
G
Yeah
so
I
made
the
changes
and
that
you
mentioned
in
the
space
model:
okay,
but
then
yeah
so
numpy
version.
That
is
the
concrete
tenth
and
that's
where
ci
is
filling
okay
and
then
yeah.
Then
I'm
working
on
example
uses
and
another
specie
model.
So
these
are
the
things.
B
B
Okay,
sweet
yeah;
okay.
So
after
we
change.
B
B
B
Cool
cool-
let's
see
great
all
right,
let's
see,
let's
start
running
through
this,
so.
B
A
D
B
A
Oh,
so
do
we
need
to
mention
that
in
the
tutorial,
so
currently
I
just.
B
B
Well,
so
that's
right,
so
that
will
work
when
it's
the
released
version
right
so
yeah.
So
that's
sort
of
the
thing
you
gotta
you
gotta
be
thinking
about
all
right,
so
we've
got
the
current
development
version
and
then
until
we
release
the
next
version,
then
all
this
stuff.
You
know
when
you
do
install
the
fml
you're,
getting
the
last
version
right
and
so
we're
gonna
make
so
yeah
so
give
it
the
path
to
well
you're
gonna
actually
need
to
uninstall
first,
because
just
that's
that's.
B
I
usually
want
like
dash
y,
oh
okay,
yeah
dash
y
will
make
it
not
ask
you
for
that
and
then
do
dffml
anything
you
might
have
installed.
I'm
not
sure
if
you
can
do
star
like
diff
or
pip
uninstalled,
if
well,
because
any
source
is
within
dffml,
so
yeah
try
that.
B
H
B
A
Yeah
there's
the
nas
files
are,
I
can,
if
I
does
it
work.
If
I
delete
it.
A
B
B
Yeah
yeah,
of
course,
that's
like
one.
That's
one
of
the
things
that
I've
noticed.
This
is
the
harder
part
about
writing
to
the
tutorials.
Is
that
like
you're,
you've
got
one
copy,
you
want
to
display
another
copy,
and
then
you
don't
you
you
know,
especially
if
you
use
the
lines
thing,
then
you
it's
you.
The
chances
are
you're
going
to
end
up
forgetting
some
of
the
lines
and
then
they
copy
paste
it
and
it
doesn't
work.
So
it's
a
typical,
difficult
thing
to
run
into
so,
let's
see
so.
B
Yeah,
okay,
so
let's
see.
B
C
B
B
B
B
B
B
G
Yeah
so
tensorflow
has
2.3
and
they
don't
support
numpy
1.19.1
very.
B
B
B
B
All
right,
okay,
let's
see
so,
we
need
to
have
a
new
version.
Okay,.
B
Yeah
yeah,
you
remember
when
we
had
to
pin
it
yeah.
This
is,
let's
see
it's
very
similar
when
we've
been
we've
been
running
into
these
like
a
few
times
now,
and
this
is
like
yeah.
This
is
very
annoying.
Let's
see
if
we
can
find
a
stack
trace,
because
it
would
be
good
to
record
what
the
hell
is
going
on
here
because
yeah
okay.
So
how
did
you?
B
B
F
B
Okay,
well
that
that's
yeah,
that's
much
clearer,
clearer
what
the
hell
is
going
on
see.
I
think
I
was
looking
at
like
what
was
I
looking
at,
who
I
was
even
looking
at
this
or
something
it
was
like.
It
was
running
an
example:
oh
yeah.
This
is
what
I
was
looking
at
and
I
was
just
like
how
the
hell
are
we
going
to
figure
out
what
is
incompatible
here
right,
because
this
is
just
for
some
reason.
It
didn't
get
it
on
install!
Oh,
I
guess
it
says
tensorflow,
but
I
didn't.
B
Yeah,
okay,
I
couldn't
figure
out
what
was
the
other
package
that
it
was
incompatible
with
right.
I
guess
numpy,
but
I
was
wondering
you
know
where
is
numpy
1.19.1
coming
from
was
my
main
thing,
because
I'm
assuming
somebody
is
pinning
numpy
19.1
right.
It
would
be
really
nice
to
understand.
I
mean
sudarsana.
You
found
that
tool
right,
that
that
gives
us
the
dependency
tree,
pip
depth
tree
right
or
something.
F
B
F
B
Yeah
that
issue
yeah.
That
issues
probably
needs
to
go
away.
I
mean
the
idea
is
basic.
Well,
the
idea
is
essentially
the
same
thing.
Well,
it's
okay!
So
the
idea
is
not
just
the
dependency
tree
like
that.
It's
it's
to
resolve
all
these
version
numbers.
That
was
the
idea
behind
this
one,
and
so
that
that
that
is.
That's
that's
why
it's
sort
of
still
still
desired
is
something
we
might
do,
but
I
haven't
talked
to
rahul
in
a
while.
B
He
wanted
to
do
that,
but
I
haven't,
I
don't
think
he's
had
any
bandwidth,
for
he
was
going
to
do
some
kind
of
solver
on
it
to
figure
it
out,
but
yeah.
That's
I
don't.
I
haven't
heard
from
him
on
that
in
a
while,
so
anyways,
but
I'm
going
to
copy
paste
this
giant
stack
trace
in
here
or
just
this
line
so
that
we
have
an
issue
that
tracks
what
happened
here
and
then
we'll
close
it
because
we'll
just
pin
it.
B
B
B
C
H
B
Brian
all
right,
I
put
this
guy
in.
B
Anymore,
okay,
let's
see
so
still
need
revenge.
Okay,
let's
check
out
the
dataflow
one
run
command.
C
B
B
B
D
H
B
What
is
it
with
the
other
one?
It's
input
def
or
something
or
no.
I
guess
this
does
get
converted
to
have
a
hyphen,
let's
waste,
something
to
fix
that,
never
mind.
B
B
B
All
right
great,
so
this
looks
good
just
like
a
couple
things.
Oh
and
then
there
was
yes
gotta
watch
out
for
spaces
here
at
the
end
of
this
line.
We
got
one
here,
but
it
looks
like
we
missed
one
here.
B
B
B
Okay,
let's
see.
B
Okay,
I
figured
that
was
what
happened.
So
I
think,
let's
see,
let's
see
well,
it's
a
bummer
about
your
mic,
but
that's
okay.
I
think
I
mean
you
could
try.
You
could
try
rejoining,
but
let's
see
or
we
can
just
we'll
watch
the
chat
so.
A
B
B
B
I
was
playing
around
with
heroku
recently
and
just
like
how
how
nice
that
is-
and
I
was
just
I
was
thinking
about
this
and
how
nice
this
is
and
how
we
need
to
expand,
expand
upon
this,
make
it
make
it
even
more,
like
you
know,
figure
out
how
to
do
I've.
Have
you
guys
seen
that
then?
The
next
thing
I
want
to
do
is
is
use
caddy
server.
If
you
guys
think
caddy
server.
B
Okay,
yeah:
we
should
check
this
out,
so
it's
it's
like
it's
great.
Basically,
they
they
have,
let's
see
where
is
they
have
auto
tls?
This
is
why
I
really
like
it,
because
they
do
automatic
tls
on
everything
and
it
functions
as
a
great.
It's
a
great
reverse
proxy,
easy
reverse
proxy.
B
That
will
do
tls
on
all
your
sub
domains
and
things
and
set
up
all
the
certificates
and
automatically
renew
them
and
everything
for
you,
and
it
can
reverse
proxy,
like
all
of
the
things
like
http
2,
websockets,
whatever,
which
is
the
main
reason
why
I
like
it
so
much
along
with
the
tls
thing,
and
then
it
can
also
serve
static
files.
It
can
do
it
can
do
lots
and
lots
and
lots
of
things.
B
It's
very
fully
featured
let's
see-
and
it
has
these
little
config
look
little
config
files
and
the
latest
thing
is
so
they
have
these
little
config
files,
but
it
also
is
configurable
via
an
http
port
like
another
port,
so
one
port
serves
and
then
the
other
port
configures.
B
So
you
can
sort
of
add
new
hosts
and
stuff
on
the
fly
and
yeah,
it's
very
it's
sweet.
It's
very
sweet,
so
basically
the
idea
would
be
we'd
take
the
stuff
that
we
have.
You
know
that
that
existing
tutorial
that
where
it
starts
the
docker
container
but
essentially
add
a
few
more
operations
and
these
operations
know
how
to
interact
with
caddy
so
throw
these
operations
in
the
data
flow
and
basically
say:
okay,
start
that
docker
container
but
start
it.
B
You
know
with
some
kind
of
like:
let's
see,
how
would
we
do
this
we'd
want
to
start
all
we'd
want
to
start
it
we'd
create
like
a
network
and
then
we'd
start
the
new
containers
in
the
network
and
the
main
container
would
be
this
caddy
container
running
and
that
one
is
the
one
with
the
port
exposed,
and
then
we
have
a
reverse
proxy
to
the
other
containers.
Within
this.
B
B
Exactly
exactly
every
yeah,
so
you
would
just
go
add
new
web
hooks
to
different
repos
and
it
would
say:
okay,
this
is
a
new
repo.
You
know,
let
me
go,
deploy
all
those
data
flows
and
have
caddy
reverse
proxy.
All
the
different
data
flows.
You
know
and
it'll
appear
as
one
server
you
know,
maybe
different
sub
domains
or
different
urls
or
getting
reverse
proxied
to
different
actual
containers
running
with
whatever
the
service
is.
That
is
that
data
flow
right.
B
So
that's
sort
of
like
the
next
step
in
building
out
that
that
that
tutorial
or
sequence
of
tutorials
that
you
have
there
right
because
right
now,
it's
it's
one.
It's
it's
the
one
container
yeah.
Can
you
make
an
issue
to.
B
B
B
B
So
the
idea
here
is
to
add
more
operations
to
deploy
operations
so
that
one
could
have
a
server
where
the.
B
Running
one
container
per
repo
web
hooked
repo
with
only
container.
H
B
B
Okay,
let's
see
these
aren't
really
check
boxes.
So,
let's
see
actually
waiting
on
layer
support
is
a
checkbox.
B
Yeah
thanks,
let's
see
yeah,
let's
let
me
let
me
finish
out,
because
I
had
some
comments
and
then
I
think
I
had
more
comments
and
I
never
got
around
to
getting
back
and
adding
more
comments.
So
sorry
about
that,
let's
see
yeah.
B
H
Yes,
so
actually
I
wanted
to
discuss
about
how
to
like
move
forward.
Okay,.
B
Okay,
let's
see
single
p
play.
Torch
needs
review.
Okay,
so
let's
just
finish
out
the
pie
torch
review
here,
real
quick.
B
E
You
think
that
I
I
don't
know
if
they
are
important,
but
I
guess
some
very
this
stuff
was
not
working
without.
B
A
B
It
it
has
to
do
with
in
python
2.5
when
we
let's
see
well,
I
guess,
let's
see,
I
know
what
it
used
to
do
in
python
2,
but
let's
see,
let's
just
find
out,
because
I'm
not
sure
it's
real
module
serves
three
purposes:
avoid
confusing
existing
tools
that
analyze
import
statements
and
specified
modules.
They're
importing
ensure
that
future
statements
run
under
releases,
private
run
one
runtime
exceptions:
okay!
Well,
that's
obviously
not
here
document
when
incompatible
changes
were
introduced
and
when
they
will
or
were
made
mandatory
executable
documentation
yeah.
B
My
understanding
of
this
was
this
was
let's
see
this
was
something
that
yeah
we
did
in
python
2
when
we
needed
the
print
statement
and
as
a
function,
and
then
I
don't
know
what
the
hell
division
is.
But
this
looks
like
it's
very
old,
so
I
don't
think
we
need
that
one
and
I
don't
think
we
need
this
one
either.
So
that's
curious
to
me
why
it
would
be
breaking
because
yeah
these
are.
B
These
are
things
that
that,
like
my
understanding
is
that
they
would
okay,
so
the
way
that
the
python
python
development
cycle
works
is
that
they
right
now.
You
know
there's
like
there's
a
bunch
of
different
versions
of
python,
3
right
and
then
the
whole
big
deal
was
that
we
stopped
no
more
work
on
python
2.7
right
and
what
that
means
is
basically
all
these.
B
So
when
they
come
out
within
the
first
version
of
something
it's
you
know
it's,
you
know
2.7.0
or
3.7.0
right,
it's
whatever
point:
zero
is
the
last
one,
the
dot
release,
and
so
each
everybody
has
their
own.
Let's,
let's
talk
about
this
because
this
might
be
good
to
have
in
the
notes.
Let's
see.
E
B
Here,
yeah,
that
makes
sense
okay.
So
so,
let's
just
go:
let's
we'll
finish
going
over
this
so
essentially.
B
B
Releases
when
we,
when
they
introduce
a
new
feature
in
so
when
they
introduce
a
new
feature
in
one
of
the
future
releases,
like
say
so
say
they
had
3.5
right
and
we're
there
we're
working
on
3.5
that
I
believe
they
work,
two
releases
behind
so
they're
working
on
3.5,
they're,
working
on
3.6
and
they're,
working
on
3.7
right
and
as
soon
as
they
finish
at
some
point
in
time.
B
I
think
they
have
a
defined
cycle
like
every
three
months
or
something
at
some
point
in
time
like
they
get
to
the
end
of
this
defined
cycle
and
now
they're,
only
working
on
3.6,
3.7
and
3.8,
and
so
now
3.5
is
just
whatever
it
was
right.
So
they're
they're,
not
gonna,
do
any
more
bug
fixes
on
that,
and
you
can
see
that
if
you
go,
if
you
go
to
the
python
bug
tracker,
for
example,
where's
a
good
one,
oh
yeah,
what
is
that
one?
B
I
just
looked
at
one
the
other
day.
That's
like
a
good
example
of
this.
Let's
see
windows,
it's
like
a
windows.
B
B
Let's
see
where
did
it
go?
I
am
sure
I
should
get
this
one
more,
but
it's
a
good
example,
because
I
know
what
it
is
at
least,
let's
see
where
it
is.
B
B
Okay,
so
basically
they'll
when
they
have
a
bug,
they'll
decide,
for
example
async.
I
o
like
in
the
standard
library
there's
no
way
to
do
that's
why
we
have
to
have
that
async
test
case
class
right
because
there's
no
async
test
case
in
the
standard
library.
Well,
so
basically,
I
submitted
this
bug
right
and
then
they
come
through
and
they
say
well
what
did
they
say
somewhere
in
here?
They
talk
about.
B
Basically,
they
decide
that
they
decide
which
versions
they're
actually
going
to
fix
this
in
right
and
it
depends
on
whether
this
is
a
new,
a
new
feature
right.
So
in
this
one
technically
it's
a
new
feature,
so
they
were
only
adding
it
yeah,
so
they
wanted.
They
wanted
to
add
it
in
3.8,
and
so
then
it
only
get
added
in
3.8,
because
I
think
they
were
still
working
on
3.8
at
the
time
and
now
they're
working
on
3.9.
B
So
basically,
since,
if
it's
a
new
thing,
it
only
gets
added
in,
like
whatever
the
current
one
they're
working
on
is-
and
I
believe,
they're
currently
working
on
3.9,
I
think
yeah,
and
so
basically
when
when
they
introduce
one
of
those
new
things
right
and
it's
one
of
these,
I
think
that's
what
they're
saying
when
they
say
no,
I
really
lost
that
tab,
but
when
they
said
a
like
it's
a
language,
breaking
thing
right.
So
that's
why
it's.
B
The
very
first
statement
is
because
the
first
thing
the
interpreter
is
going
to
see
is
okay.
Well,
this
is
going
to
be
like
the
print
statement
right.
So
the
the
print
statement
is
going
to
cause
a
lot
of
problems.
You
know
you
can't
import
that
later
on,
it
needs
to
be.
It
happens,
needs
to
happen
before
any
other
imports
right,
because
as
soon
as
you
you
make
that
change,
then
right.
B
If
you
weren't
to
do
that,
then
it's
going
to
treat
all
the
print
statements
as
if
as
if
they
shouldn't
be
a
function,
call
right
and
and
when
you
say
from
future
import
print,
then
it
was
making
it
as
a
function
call
and
so
when
they
added
it
in
a
new
version,
then
when
they
have
that
they
have
that
little
dot.
Release
version
right
this
little
though
what
the
z
right
and
so
that
that's
like
a
bug,
fix
thing.
So
they're
still
working
on
that.
B
If
you
look
at
like
the
downloads
yeah,
so
bug
fix
security,
security,
end
of
life
right,
so
they're
no
longer
working
on
the
the
2.7
series
right,
but
they
are
working
on
3.8,
3.7,
3.6,
3.5
and
then
3.9
is
the
active
development
version
right.
B
So
they,
let's
see,
let's
see
july
yeah,
so
they've
been
working
on
okay
and
3.8.
So
then,
you
can
see
they
come
through
here
and
on
about
a
four
month
cadence
they
release
a
new
dot
release
this
last
number
right
and
when
they
do
that,
if
they
added
a
language
breaking
statement
in
like
3.9,
for
example,
the
one
you
don't
see
here,
because
that's
the
active
get
branch
then
they
might
in
the
dot
release.
Add
some
like
that
thing.
The
ability
to
import
that
from
future
does
that
make
sense.
B
Yeah
so
yeah
then,
and
that's
also
that's
like
yeah,
that's
yeah,
that's
the
chest
of
that.
B
When
they
release
new
new
versions,
that.
B
Ul
end
of
life
they'll
and
it's
a
language
and
it's
something
that
would
break
all
the
code
unless
the
interpreter
knows
about.
B
Right
off
the
bat
it's
put,
then
they
add
it
to
future
module
of
the
oops.
I
deleted
to.
B
B
B
Okay,
so
yeah:
let's
remove
that
a
long
bit
of
discussion
for
let's
delete
that
line.
B
B
B
F
B
So,
okay,
and
so
the
idea
is
basically
just
that
you
end
up
with
the
best
model,
because,
whatever
the
okay
cool,
let's
see,
I'm
just
trying
to
think
because,
let's
see
what
is
the
way
that
they
data
loader
data
set
okay
and
once
again,
it's
like
god.
I
wish
there
was
a
better
way
that
we
could
take
the
yeah
effectively.
G
Yeah,
actually,
actually
we
can
put
this
under
the
same
issue
that
we
have
already
opened
regarding
tensorflow
tf.data.
So
we
can
use
that
for
this
input
and
validations
things
so
the
tensorflow
thing
yeah.
So
we
have
one
issue
right:
tf.data.
We
need
to
migrate,
tensorflow.
B
G
Yeah
so
so
the
tensorflow
is
all
started
subordinating
by
torch,
also
so
using.
F
B
G
That
may
be
of
that,
maybe
your
health-
I
mean
saksham,
you
have
to
check,
but
I
think
that
will
help.
B
Yeah
cool
because
yeah
that's
what
I'm
wondering
here
is
is
what
it
would
have
yeah.
Can
we
that
yeah
my
main
thing
that
I'm
thinking
about
so
so
we
can,
we
can
do
tensorflow
first
and
then
we
can.
We
can
do
this.
One
right.
F
B
The
main
thing
I'm
thinking
about
here
is
that
is
that
you
know
we've
got
I
mean,
and
this
is
just
the
thing
that
we
have
with
a
lot
of
models
right
now
is
that
we're
loading
everything
in
memory
right
and,
ideally,
we
we
wouldn't
load
everything
into
memory,
right,
yeah,
because
that's
just
some
machines,
don't
have
that
much
memory
right
and
then
a
large
point
of
this
that
with
the
sources
and
the
async
and
everything
is
the
fact
that
we
could
stream
everything
in
and
out
right.
B
I
think
right
now,
the
only
thing
that
we
actually
have
the
capability
to
do
that
with
is
the
dell
for
pi
thing
so,
but
that
happens
to
be
just
because
they
have
this
streaming
api
that
works
well
with
the
way
they
just.
That
was
a
happy
accident.
Most
other
libraries
are
not
quite
as
as
it's
nice
in
that
respect.
So
let's
see
yeah
okay.
Sadly,
that
has
no
neural
networks.
B
So,
okay,
so
yeah,
we'll
just
add
it
to
here
and
we'll
say,
notes,
or
let
me
just
put
it
in
the
description
so.
B
B
Features
applicable
features,
did
I
put
this:
where
did
I
okay
yeah?
So
let's
nix
I
mean
this
is
this?
Is
this
this?
I
mean
this
amounts
to
this
right.
So,
although
yeah,
let's
see
classification.
B
Let's
see
yeah
oh
yeah,
and
I
said
this
okay,
I
mean
it
works.
I
think
there's
a
bit
of
there's
a
few
things
that
could
sort
of
you
could
do
some
cleanup
on.
You
know
like,
for
example,
you
know:
do
we
really
need
a
set
or
indicator
for
this?
If
we're
not
gonna,
you
know
if
there's
no
functionality,
if
there's
nothing,
to
wrap
around
that
right.
If
there's
no,
if
you're
not
going
to
do
saving
or
loading
there
or
something
like,
I
don't
think,
that's
really.
B
E
So
I
just
use
some
of
the
code
from
from.
B
Transfer
for
transformers
yeah,
okay
yeah,
because
that
was
I
mean
that's
a
product
of
of
that
code's
been
around
for
so
long
now
that
that
it
could
also
be
cleaned
up
right,
though,
didn't
we
have
an
open
issue
for
tensorflow?
No,
you
did
that
tensorflow
refactor,
but
they're
still
there's
still
probably
stuff
in
there,
because
we
then
decided
we
need
to
get
rid
of
applicable
features
and
yeah.
There's
just
been
a
few
things.
B
So,
let's
just
sort
of
let's
say,
let's
just
put
a
comment
for
general:
let's,
let's
do
some
general
cleanup
and
just
because
I
mean
I
think,
there's
a
few
things
in
here
that
we
could
just
sort
of
tidy
up
and
then
do
you
have
the
example.
I
don't
believe
we
have
example
usage
on
these
guys.
E
Yeah,
I
realized
that
I
haven't
added
the
dog
stream.
Okay,
all.
B
Right
so,
let's
remember
and
then,
since
this
is
sort
of
like
a
model
scikit
thing,
your
doc
string
is
going
to
go
in
init.py,
so
check
out
model
scikit
for
that
so
similar
to,
since
this
is
similar
to
model
psychic.
B
B
B
Okay,
sorry
just
finishing
that:
okay,
so
confusion
on.
Let's
say.
B
Okay
and
then,
let's
see
oh
yeah,
okay,
so
now
the
question
is
integration,
cli
test
and
basically
so
what
is
the
difference?.
B
You
think
test
case
and
dealie.
Okay,
no,
that's
a
dumb
answer!
Okay,
my
my
dumb
answer
is
ideally
nothing
because
I
remember
a
few
weeks
ago.
I
think
I
threw
up
a
branch
and
let's
see
where
did
that
branch
go.
B
Yeah,
okay,
so
a
few
weeks
ago,
wow,
it's
probably
gonna,
be
a
few
months
ago
now,
but
we
threw
a
branch
up
here
and
yeah
here
during
a
meeting
and.
B
Ch2
tempter,
okay,
now
oh,
this
was
okay.
I
guess
we
merged
this,
so
yeah,
you
guys
remember.
So,
basically
we
were
like
all
right.
Okay,
let's
have
it.
We
know
we're
always
making
temporary
directories
for
things.
So
let's
just
have
async
test
case,
be
the
first
thing
that
it
does
it
you
know,
makes
a
temporary
directory
for
that
test
case
right.
B
So
then
anything
you
do
is
in
this
new
temporary
directory,
and
so
then
we
were
talking
about
how
we
had
to
reference
files
by
their
relative
paths,
to
the
file
that
you're
you're
running
with
underscore
underscore
file
underscore
underscore
and
getting
the
relative
path
from
there.
B
And
then
we
talked
about
how
this
is
sort
of.
Like
you
know,
our
goal
here
is
that
we
want
to
take.
We
want
to
take.
B
Basically,
we
want
async
test
case
to
be
the
same
thing
as
integration
cli
test
case,
although
the
problem
right
now,
I
think,
is
because
and
why
we
want
this
is
because
it
would
have
these
stacks,
which
is
sort
of
something
that,
and
this
is
part
of
why
we
need
to
go
through
and
make
sure
that
this
is
all
this
is
something
that
has
to
be.
It
has
to
be
like
consciously
done,
because
there's
some
test
cases
that
have
stack
in
a
stack
and
so.
E
Yeah,
what
are
these
stack
tests
case.
B
Basically
a
lot
of
the
cli
tests
had
a
bunch
of
with
statement
usage
right
and
you
can
condense
a
with
statement
down
to
a
call
to
stack.enter
context
and
and
that,
and
that
way
you
don't
have
this
indented
with
block
right.
And
therefore
it
was
the
main
reason
why
this
was
helpful
was
for
these
mk
temp
file
and
mk
temp
dirt
right,
because
we're
always
creating
these
stupid,
temporary
files
right
and
temporary
directories.
B
And
in
this
case
I
believe
the
main
thing
was
yeah,
that
we
wanted
non-existent
temporary
files
and
to
do
that,
we
created
a
temporary
directory,
and
maybe
it's
not
in
here
yeah.
We
created
a
temporary
directory
and
then
created
a
file
name
within
that
temporary
directory,
because
if
you'd
use-
and
I
believe
this
was
because
of
the
sources
and
how
with
you
know,
we
have
the
allow
empty
flag
and
stuff
right,
and
so
we
were
having
to
create
sources
that
had
file
names
but
were
empty
files.
B
And
so
there
was
there's
a
bunch
of
tests
that
do
this
kind
of
stuff,
and
basically
I
got
it
all
condensed
down
here,
and
I
mean
the
main
thing
to
answer.
Your
question
is
that
there,
the
only
difference
is
that
async
test
case
or
integration
cli
test
case
has
it'll
it'll
check
for
the
required
plugins,
which
is
another
thing
right.
So
if
you
have
other,
you
know,
if
you,
if
you're
not
in
the
standard,
it'll
skip
the
test
case
unless
those
other
plugins
are
installed,
and
then
it
derives
from
this
async
exit
stack
test
case.
B
So
you
have
all
the
you
know
you
can.
You
can
create
temporary
files
and
temporary
directories
without
using
the
with
statement
so
that
you
don't
end
up
with
you
know,
giant
indented
blocks
in
your
test
cases,
and
then
it's
going
to
throw
you
into
a
new
temporary
directory
like
by
default
you're,
going
to
be
operating
in
a
temporary
directory
rather
than
you
know.
Whatever
the
you
know,
the
top
level
of
dffml
would
usually
be
your
working
directory.
B
So
but
the
goal
here
is
eventually
to
go
through
and
audit
the
existing
test
cases
and
figure
out
which
ones
have
usage
of
stack
or
a
stack
and
then
or
you
know,
don't
call
super
dot,
setup
or
super
dot,
tear
down
and
have
a
setup
and
tear
down
and
migrate
them
over,
so
that
you
know
all
this
functionality
is
essentially
in
async
test
case,
because
if
we
we
use
that
temp,
I
mean
there's
just
so
many
places
where
we
create
temporary
files
that
this
it
should
all
be
in
one
class.
B
But
so
there's
really
no
difference
other
than
the
fact
that
you
get.
You
have
access
to.
You
know
to
whatever
methods
are
in
here
right
but
like
fundamentally
right
but
but
and
conceptually
we're
hope.
B
Hopefully
they
become
the
same
thing
so
don't
like
if
you're
using
one
over
the
other
you're
using
you
should
be
using
it
because
you
know
you
want
the
mk
temperature
function
or
something
or
you
want
to
have
this
exit
stack
created
for
you
or
to
you
know
both
of
the
kinds
of
exit
stacks
or
something
you
know,
and
that's
something
that
that's
something
that
we
need
to
I
mean
we
might
need
to
document
that,
but
more
likely
we
just
need
to
go
and
make
the
change
right
and
then
then,
there's
not
even
confusion
of
you
know
missing
that
in
the
documentation
right,
because
the
documentation
is
huge
now,
so
it's
not
always
like.
B
B
B
So
conceptually
not
much,
we
need
to
merge
them
into
one
class
eventually
check
the
implementation
for
details
on
for
extra
methods
and
our
okay
temperature
and
okay
temp
file
methods
and
exit
stacks.
B
So,
let's
see
all
right,
okay
and
then
I
think
I
submitted
this
review.
Okay,.
E
B
B
Yay
trailing
white
space
yeah
we
need.
We
have
a
oh
jesus-
oh
god,
oh
god,
okay!
Let's
look
at
this.
What
happened.
B
Well,
I
guess
I'll
just
click
anywhere.
Is
this
going
to
show
up,
or
is
this
what's
going
on
here
so.
H
B
It
just
like
this
okay,
it
must
be
a
problem
with
the
docs.py
script,
because
it's
generating
headers
that
are
not
the
right
or
no.
Oh,
it's
because,
oh
it's
because
of
the
okay
yeah
yeah.
This
is
the
problem.
The
problem
is
that
we
need
to
switch
the
headers
around
in
that
template
and
we
need
to
go
through.
When
did
we
last
deal
with
this?
We
dealt
with
this
before.
B
B
B
Operation
output
right
so
take,
for
example,
yeah,
so
this
guy
right,
these
are
clearly
like
not
the
same
style
of
headers
that
we're
using
other
places.
You
know
the
plus,
plus
plus
plus
plus.
If
we
look
at
the
model
or
let's
see,
is
this-
I
think
we
have
the
same
well.
We
don't
really
do
that
with
the
model.
Stewie.
B
Yeah,
because
that
would
be
the
same
thing
to
do.
The
problem
is
that
the
template-
god,
damn
it
really?
What
is
this,
which
one
is
this
well
it'll,
tell
us
later.
I
guess
the
template
docs.
B
Use
the
or
no
oh,
it's
the
docs,
it's
either
docs.py
and
template.
I
believe,
because
you
used
operations,
nlp
operations.
G
Yes,
numpy
dog
strings
use
this,
so
I
forgot
that
we
don't.
B
Use
these
well,
that's
that's
I
mean
we
should
change
it
because,
let's
see
what
are
we
doing,
I
mean
we
have.
B
I
mean
we
need
to
change
it
because
this
is
annoying
as
hell,
but
we're
not
going
to
be
able
to
fix
it
very
we're
going
to
have
to
go
through
and
change
them
all
right,
so
yeah,
but
this
is
this
is
the
main
problem.
Let's
make
some
notes,
so
the
main
main
issue
here
is
that
script,
slash,
docs,
dot
py
is
creating,
is
using
the
underline
somewhere
and
therefore
rst
sphinx
rst
person
is
deciding.
B
That
is
a
a
bunch
of
hyphens
is
a
larger
like
a
second,
it's
like
a
second
level
header
or
something
second
level
header.
Let's
see.
B
B
See:
okay,
it's
a
lot
of
pluses.
It's
just
in
our!
Let's
see
it's
just
in
the
main
modules
and
then,
where
else
do
we
have
it?
Okay,
just
the
main
module
all
right.
Well,
that
means
that
we
haven't
don't
have
documentation
for
the
other
operations,
which
is
sad,
but
it
does
mean
we
have
less
less
things
to
worry
about.
So,
let's
see.
B
So
is
there
like
a
fast
way
to
change
that?
Let's
see,
let's
see,
we
need
to
change
all
those
pluses
in
there.
B
B
B
B
You
know
just
three
pluses,
so
we
want
to
go
through
and
we
want
to
we're
writing
this
for
loop,
that's
going
to
print
out
the
long
bit
and
then
you
know
down
the
way,
and
so
so
we're
going
to
go
through
and
we're
going
to
grab
every
python
file
in
dffml
operations
and
we're
going
to
replace,
let's
see
and
hopefully
plus
it's
not
some
kind
of
weird
regex
thing
we're
going
to
replace
that
with
yeah.
We
need
some
way
to
multiply
that.
I
guess
we
could
just
use
python.
B
B
All
right
so
we're
going
to
look
for
all
those
pluses,
reverse
sort
them
so
that
we
print
one
and
like
print
print
the
longest
one
first,
because
we
got
to
replace
the
longest
one
first
and
then
I
don't
know
what
we're
gonna
do
all
right,
import
os
and
then
this
is
like
always
what
I
do
when
I
just
can't
figure
out
how
to
do
something
that
bashes.
I
just
do
it
in
python
right,
because
then
we
know
how
to
do
that
right
and
then
we
can
print.
B
And
then
you
got
a
double
escaper:
let's
see
we
can
do
single
quotes
dash
times.
Okay,
so
we're
getting
our
our
bunch
of
pluses.
We
export.
I
because
you
have
to
export
the
variable,
otherwise
it's
just
a
variable
into
bash
to
get
it
into
the
environment
of
subprocesses
and
then
we're
going
to
create
this
python
process,
which
is
going
to
say,
okay,
grab,
that
environment
variable
and
make
a
bunch
of
dashes.
B
B
B
Okay.
So
now,
if
we
do
get
diff
okay,
so
we
should
have
changed
them
all
right.
So
if
you
have
to
do
massive,
find
or
replace
things,
that's
this
like,
for
example,
well
suit
honshu.
I
guess
I
don't
know
you're
doing
live
yours.
Your
stuff
is
going
to
be
mostly
like
pretty
code
heavy,
but
sometimes
sometimes
you
have
something
right
where
you're
working
on
the
whole
code
base-
and
you
need
to
do
stuff
like
this
and
so
find
an
sed.
B
B
Okay,
the
python
workaround,
so
basically
the
point
of
using
python
here
was
just
because,
if
you
don't
the
point
of
this
is,
if
you
don't
know
how
to
do
something
in
bash,
do
python
dash
c
and
just
write
it
in
python.
That's
that's
the
point.
B
If
you're
writing
some
shell
stuff-
and
you
don't
know
if
you
and
you
can't
figure
out
how
to
do
it
in
bash,
you
can
always
just
write
it
in
python
and
you
can
use
python
dash
c
and
you
can
use
semicolons
to
separate
lines,
and
the
point
of
this
was
really
just
to
say.
Okay,
so
I
our
variable
I
here
is
all
of
these.
B
Our
variable
I
here
was
oops.
Well
now,
they're
gone.
B
It
was
all
of
these,
it
was
just
every
single
it
formed
a
you
know.
It
was
a
bunch
of
where
did
it
go
yeah?
It
was
all
of
these
right
and
then
we
cut
it
down
and
then
what
we
wanted
was
we
wanted
to
take
these
and
for
each
one
of
these
strings.
We
wanted
to
replace
it
with
the
same
number
of
minuses
right.
B
We
wanted
to
take
the
number
of
pluses
and
replace
it
with
the
same
number
of
minuses,
and
so
we
had
to
reverse
it,
because
or
else
if
we
did
this
for
one
first,
then
we'd
end
up
with
minus
minus
minus
and
then
a
few
pluses
right,
and
so,
but
I
don't
know
how
to
do.
I
don't
know
how
to
convert
that
number
of
things
in
bash
like
I
just
don't
know
how
to
do
that.
So
my
my
go-to
is,
I
take
it
and
I
write
it
in
python
right
and
so
to
get
it
into
python.
B
Yeah
yeah
right,
yeah,
okay,
so
I'm
glad
I'm
glad
there
was
something
there
was
something
useful
here.
There's
also
this
thing
dash
x.
B
B
Yeah
there's
some
way,
there's
some
way
to
define
yeah,
oh
dash,
yeah,
maybe
not
I
don't
know
I
swear
there
use
there
was
something
set,
implementation,
specific
options.
I
think,
if
you
just
set
something
it
just
shows
up,
I
can't
yeah
you
can
they
they
end
up
somewhere.
B
They
end
up
in
like
some
cis
somewhere,
maybe
that's
not
actually
as
useful
but
yeah.
I
guess
you
know
the
environment
variables.
You
can
always
use
and
that's
just
an
export
with
bash
but
yeah.
Okay.
So,
let's
see
what
happened
here.
B
Okay,
hopefully
it
did
the
right
thing
right,
let's
see,
okay,
yes,
I
think
we
did
the
right
thing
so
now,
yes,
and
now
we're
all
using
the
standard,
numpy
doc
strings.
Now
our
we
need
to
go
to
make
sure
that
everything
yeah
okay,
so
we
must
not
be
using
orgs
there.
Okay,
this
is
all
good.
This
is
all
good
services.
This
is
just
like
a
listing.
Okay,
so
we're
good
great.
H
B
B
B
Fix,
okay,
so
by
dark
strings.
B
C
Okay
and
then.
E
B
B
All
right,
okay,
so
sorry,
I
know
that
took
a
while.
So
let's
see
and
we
wanted
to
talk
about
accuracy,
plugin,
okay
and.
H
Yes,
yes,
so
right
now,
what
I
have
done
is:
I
have
basically
removed
the
accuracy
method
in
all
of
the
models
so
yeah,
so
I
was
actually
like,
except
for
the
tensorflow
hub
and
the
trans
transformer
models
or
the.
I
have
not
removed
the
accuracy
method
and
yeah.
Okay,.
E
H
Also,
like
I
was
thinking
of
wrapping
scikit-learn
model
evaluation
like
metrics
plug-in.
I.
B
I
think
yeah,
but
I
think
we
need
to
mainly
get
you
know,
get
the
get
the
whole
the
big,
the
big
project,
part
of
this
one
done,
because
because,
or
else
it's
gonna
I
mean
you
could
go,
do
that
right,
but
and
and
then
have
them
in
there.
But
you
know
you
won't.
B
H
B
Okay,
so
you've
been
looking
at
wrapping
the
scikit
metrics.
H
Yes,
so
I've
actually
shared
a
link
on
getter,
so
you
can
look
at
it.
So
that's
what
I
was
like
planning
to
wrap
in
the
accuracy
plugin,
okay,.
D
B
Sweet
sweet,
so
are
you
gonna
do
this
kind
of
like
we
did?
Let's
see,
are
you
going
to
do
this
kind
of
like
we
did
the
scikit
models.
B
H
So
it's
like,
should
I
work
on
the
psychic
part
or
like,
should
I
get.
B
I
mean
you
could
do
either
or
right.
I
think
I
think
that
I
think
that
you
know
for
this
to
at
the
end
of
the
day.
You
know
you
could
do
you
could
end
up
doing
a
bunch
of
work
on
that
and
but,
like
you,
need
to
to
have
that
work.
Get
used.
You're
gonna
have
to
finish
the
other
one
right.
B
So
you
know
it's
sort
of
it's
sort
of
whatever,
whatever
you're
feeling
on
now
right.
Yes,
since
since
since
you're
yeah
you're,
it's
it's
it's
all.
It's
all
up
to
you
on
that
one
right.
So.
H
B
One
thing:
are
you
also
looking
for
so
you
said
you
removed
some
of
the
models
right
so.
H
B
Okay
and
the
main
reason
why
I
was
suggesting
that
we
get
guidance
from
himachu
on
this
is
because,
because
I'm
not,
I
can't
well,
why
was
it
oh
because
of
the
the
main
when
you're
gonna
you're
gonna?
Well,
I
guess
the
next
step
is
modify
the
test
cases,
but
mainly
yeah.
Is
there
just
just
and
now
is
a
good
time
to
do
it,
but
so
let
me
go
to
the
main
issue
here.
B
B
B
Right,
like
I'm
like
do
do
these
do
these
psychic
metrics
measures
of
accuracy
also
make
sense
for
nlp
models
and
because
I'm
not
I'm
not
familiar,
let's
see,
I
guess
you're
yeah
do
do
they
make
sense,
or
does
your
validation
data
look
a
lot
different,
because
I
can't
remember
where,
let's
see.
B
B
B
Yeah,
okay,
that's
kind
of
this
is
kind
of
what
I
was
thinking
here.
So
that's.
This
is
mainly
the
point
of
this
is
to
have
a
check-in
right
and
say:
what
are
we
going
to
need
to
do
here
because
we
may
need
to
we?
You
know
we
should.
We
should
update
this
issue
and
say
that
we
may
be
taking
out
these
right
because
we're
about
we're
removing
this
this
code
right.
B
B
B
A
new
class-
and
we
could
even
do
it
in
the
same
file
for
now,
but
basically
just
take
the
code
out
and
and
put
it
you
know,
don't
don't
remove
the
code
completely
right
because
now
now
we
know
that
you
know
for
then
let's
enumerate
that
list.
So
we
had,
you
said
well
pi
torch
we
haven't
merged
yet,
but
we'll
have
to
do
that
so
because
that
will
get
merged
soon
here.
So.
B
Pie,
torch,
okay,
so
for
vision
and
nlp
models
we
need
to.
We
need
to
take
or
move
the
existing
code
into
an
accuracy
scoring.
B
Score
class
and
for
use
in.
B
B
Okay,
yeah
no
worries.
Thank
you
for
joining
today.
Bye
have
a
good
one,
okay,
so
yeah
so
for
nlp
and
vision,
yeah,
so
take
those
put
them
in
their
own
class
right.
So
you
still
have
the
code
around
right
and
then
you'll.
When
you
go
to
everything
except
ones
where
they
we
took
out
the
accuracy
method
in
phase
three.
So
does
this
make
sense.
H
B
H
B
Sweet
okay,
yeah
and
then
basically
yeah
just
for
just
for
now.
That's
all.
We
really
need
to
do
right
and
then,
within
the
test
cases,
we're
just
going
to
make
sure
that
we're
using
that
class
as
well
and
we're
probably
going
to
need
to
you
know,
register
those
classes
as
as
entry
points.
So,
let's
see.
B
Let's
see
yeah
okay,
so
then
the
next,
because
the
next
big
thing
here
is
that
yeah,
okay,
so
we've
updated
yeah
we
updated
all
of
this.
We
didn't
okay,
we
may
need
to
go
through
an
update.
We
probably
should
have
done
that
a
little
bit
earlier,
but
we
need
to
go
through
an
update,
the
command
line,
the
command
line
stuff
in
dfml
cli
ml,
because
you're
gonna
need
to
update.
B
Well,
let's
see,
I
guess
that
will
be
in
phase
four,
because
that's
to
be
when
you
start
updating
the
tests.
So
let's
see
yeah.
If
you
want
when,
when
you
think
I
need
a
break
from
this
massive
of
accuracy
in
models,
then
you
go.
You
go
and
implement
some
psychic
accuracies
right.
So,
let's
see
okay,
so
okay
so
need
update.
B
Diff,
slash,
cli,
slash,
email,
the
py
first,
so
that
cli
tests
and
examples.
I
will
be
able
to
specify.
H
No
thanks.
That's
it.
B
Okay,
great
cool!
Thank
you
all
right!
Himachu!
Do
we
still
have
you?
Are
you
oh
yeah,
okay,
great,
so,
okay,
we
just
needed
to
upload
the
changelog.
Did
we
do
that
already?
I
feel
like
I
remember
you
saying.
G
Really,
no,
no,
no
I'm
working
so
there's
nothing.
B
Okay,
cool
and
then
again
I
will.
I
will
review
let's
see,
so
let
me
just
make
sure
because
I
I'm
still
I'm
I'm
in
need
of-
I
mean
I'm,
I'm
I'm
behind
and
behind
so
let's
see
so
these
are
okay.
B
Good
single,
let
me
just
make
sure
so
let
me
everybody.
Please
give
me
a
recap
of
of
what
I
owe
everyone.
So
I
will
look
over
all
of
these
pull
requests,
but
is
there
anything
else
that
I'm
not
thinking
of.
E
Okay
yeah
this
in
last
week,
you
were
talking
about
creating
the
tutorial
for
points.
B
Yeah,
that's
a
big
thing!
Okay
that,
because
I
know
that's
gonna
give
so
so
pr
review
and
entry
point.
Okay,
sweet
all
right!
Well,
thanks
everyone
and
then
so
I'll
get
on
those
and
let's
see
yeah,
okay.
Okay,
that's
the
only
actual
item:
okay,
yeah!
That's
still
still
waiting,
all
right
sweet!
Thank
you!
Everyone
and
I'll
talk
to
you
on
friday
or
on
twitter
have
a
good
one.
Thanks,
yeah
bye.