►
From YouTube: Weekly Sync 2021-07-06
Description
Meeting Minutes: https://docs.google.com/document/d/16u9Tev3O0CcUDe2nfikHmrO3Xnd4ASJ45myFgQLpvzM/edit#heading=h.js6785aou89
A
A
A
B
Yeah,
I
posted
a
fix
on
the
other
pr
and
it
should
be
fixed
now.
If
you
really
run
the
test.
There
was
some
issue
with
pillow
and
it
was
failing
due
to
that.
They
have
merged
a
fix
now,
so
it
should
be
fixed.
A
Okay,
great
all
right:
oh
okay,
and
we
didn't
get
a
review.
B
B
But
I
think
the
issue
itself
should
be
fixed
now,
because
the
one
the
pr
that
was
posted
for
this
issue
has
been
merged.
Now.
B
Oh
yeah,
but
it
won't
be
released.
I
don't
know.
A
Okay,
so
so
we
have
a
so
you're
telling
me
that
there
is
a
fix
in
pie
torch
for
this,
but
it
hasn't
been
released.
C
B
A
Let's
see,
release
pillow
july,
okay,
I'm
yeah,
I'm
wondering
okay
has
it
happened.
A
Yet
five
days
ago,
22
commits
to
master
since
this
release,
and
that
would
be
this
okay,
so
crap
all
right.
So
we
haven't,
we
don't
have
a
next
one.
Yet
I
wonder
how
quickly
they
will
get
this
out,
though,
because
numpy
is
a
big
thing.
A
So,
okay
is
this
it?
Yes,
okay,
so
does
it
exist
and
it
came
out
today.
So
maybe
we
can
sidestep
this
whole
thing.
A
A
Okay,
here
it
is
two
hours
ago
tagged.
Okay,
so
they're
probably
doing
it
at
the
moment
all
right,
so
I
think
our
problem
will
be
so
soon.
That's
the
good
news
news.
So
in
that
case,
do
we
don't
need
this?
We
can
just
bump
the
version.
Let's
see
where
did
rpr
go.
A
A
B
B
B
Okay,
yeah,
you
can
still
do
it
with
the
dictionaries
through
yml.
B
C
A
B
So
currently
you
you
see
the
else
statement
here.
B
Changing
it's
currently
taking
in
dictionaries
layers
in
the
form
of
dictionaries
through
yml
files,
and
you
know,
reading
those
as
layers
converting
those
dictionaries
into
sequential
layers.
But
we
don't
have
a
way
to
directly.
You
know
read
layers
from
python
api
which
I'm
supposed
to
do
here
so.
B
All
I
check
if
the
the
base
class
name
is
module
tech,
sequential
module
list,
core
module,
and
if
that's
the
case,
I
just
you
know,
create
another
sequential
dictionary,
and
I
add
these
modules
directly
into
it
rather
than
you
know,
passing
these
layers
to
the
dictionary
function,
which
would
then
blow
up.
A
Yeah
yeah:
okay,
let's
see
so
okay,
so
we
overwrote
typing
typing.
B
No,
we
were,
we
were
able
to
use
it
from
the
command
line,
because
we
were
only
using
dictionaries
as
input.
We
were,
we
were
not,
you
know,
creating
layers,
the
form
of
pi
torch.
We
were
creating
layers
in
the
form
of
dictionaries
in
y,
a
dot
y
am.
B
And
we
have
this
this
functions
here
and
I
think
base
class
and
they
are
converting
these
dictionaries
into
pythagoras
form.
But
if
we're
using
python
api
we
are,
you
know,
expecting
them
to
be
directly
in
fire
torch
formed
on.
B
So
yeah
this
one's
just
you
know,
setting
layers
to
be
of
any
type
and
deciding
where
whether
they
are
dictionaries
and
they
need
to
be
processed
or
they
need
to
directly.
You
know,
be
a
sequential
layer
and
directly
go
to
the
model.
A
I'm
wondering
really
about
the
implications
of
changing
this
this
this
and
then
because
this
is
this
is
a
big,
bigger
change
right
or
more
wide
wide
range
change.
B
I
I
was
thinking
that
we
could
add
other
types
which
are
being
handled.
You
know
after
this.
After
you
know
this
condition,
we
could
add
those
types
as
well
to
any.
In
case
you
know
we
have
other
any
types
of
which
we
don't.
Currently,
we
are
not
really
having
fields
with
the
type
of
penny.
A
Yeah-
and
we
really
shouldn't
is
the
thing
because,
like
actually
there's
to
replace
the
last,
let
me
just
see:
okay,
I'm
moving.
A
A
Okay,
actually,
let
me
do
it
on
this
computer.
C
A
A
D
A
A
All
right,
okay,
I
think
this
looks
pretty
legible
great.
Let's
pull
out
of
the
web
ui.
A
A
A
A
A
I'm
looking
for
similar
things
here
to
see
if
we
handled
this,
you
know
to
see
how
else
we
handled
this
for
clues
here
as
to
how
just
just
to
make
sure
that
this
is
the
right
way
to
handle
this
similar.
Stick.
A
Just
so
long
as
we're
consistent
right
all
right
well,
it
looks
like
this.
The
only
place
right
now
so.
A
A
Is
there
a
way
to
make
wonder?
Is
there
a
way
to
make
the
I'm
trying
to
make.
A
A
A
A
D
A
A
I
want
to
understand
how
this
look
worked
before
before
we
just
to
make
sure,
because
I
think
your
patch
is
probably
correct,
but
I'm
sort
of
looking
for
what
I'm
looking
for
now
is
is
things
that
that
that
you
know
may
be
problematic
that
we
maybe
should
be
fixing
here
rather
than
in
the
future.
C
That
changes
it
to
dictionary
in
the
function
that
changes
to
create,
create.
A
A
B
We
are
currently
calling
this
function
whatsoever.
B
C
A
That
thing
that's
a
problem
right,
so
I
was
just
wondering
you
know:
how
is
he
defining
a
dictionary
on
the
command
line,
because
because
that
I
don't
think
we
didn't
have
support
for
that,
so
I
was
wondering
how
the
hell
that
was
happening,
so
we
do
have
support
from
pulling
from
the
style
yeah
that
was
sort
of
like
the
main
demand.
A
The
main
point
of
concern
here:
okay,
so
it
sort
of
half
works
right
now
it
works
from
python
and
it
works
from
the
command
line,
provided
you
involve
another
file,
so
it
doesn't
work
from
the
command
line
provide
like
if
you
didn't
use
a
file
and
that's
that's
sort
of
what
I
was
trying
to
get
out
here
is.
Is
it
seemed
like
there
was
a
missing
case.
B
A
A
A
A
C
A
A
A
C
C
A
And
we
see
that
type
equals
any
stuff
changes
into
convert
value
and
the
base
config
stuff,
and
we
see
this
type
any
change
here
and
so
now
we're
wondering
okay.
Well,
why
are
we?
Why
are
we
being
so
generic?
And
it's
because
we
want
to
support
taking
pi
torch
objects
directly
right,
yeah,
and
then
we
also
want
to
maintain
support
for
taking
dictionaries.
A
So
we
probably
want
to
do
a
typing
union
on
this.
A
Is
there
are
there
more
things
that
don't?
Because
this
essentially
is
you
know
a
type
like
you're
doing
a
typing.union
type
of
thing
here
right
I
mean
so
if
you're,
because,
because
looking
at
the
way,
your
code
is
implemented,
you're,
basically
saying
hey
any
of
these
types.
Otherwise,
if
it's
dictionary
goes
here
right,
so
we
should
probably
be
using
typing
union
on
this
right.
A
Okay,
so.
A
A
A
A
That
seems
like
a
pretty
generic
thing.
That
might
happen
a
lot
right,
we're
always
using
dictionaries
for
more
abstract
stuff.
So,
but
we
really
should
be.
A
A
Okay,
so
this
is
the
real
fix
here,
so
set
type
detective
layers
can
stick:
okay,
oh,
and
that
was
now
that
was
the
old
code.
So
this
is
really
the
end
change
anything
anything
that
happens
within
these
config
code.
Things
can
just
cascade
and
cause
massive
problems.
So
that's
why
I'm
like
looking
at
this
so
hard.
Sorry,
I
mean,
obviously,
all
the
tests
pass.
C
A
D
A
A
A
We
had
also
talked
about
this,
so
we're
also
we're
also
thinking
about
okay.
How
does
this
relate
to?
I
can't
remember
who
we
were
talking
about,
who
I
was
talking
about
with
this
okay,
but
that
we
want
to
be
able
to
create
config
objects
out
of
the
inputs
to
operations,
and
it
looks
like
a
few
of
these
use
any
so.
I'm
wondering
how
this
is
going
to
affect
that.
A
A
A
It
doesn't
really
tell
us
anything
right
so
when
we
it
would
be
best.
If
we
don't,
you
know
if
people
don't
type
things
as
any.
A
Now,
if
there's
now
like
with
when
you're
configuring
objects
and
stuff,
you
know,
usually,
ideally
you
would
never
type
them
as
any
now
when
you're
creating
up,
because,
for
example,
like
like
in
this
case,
you
know,
we
know
the
types
right:
it's
either
dict
module,
sequential
modulus
or
module.
So
let's
just
go.
Do
that.
D
D
A
A
So
that's
sort
of
what
this
is
the
same
same
thing
right
and
this
would
be
basically
no
change
now.
The
question
is
what
happens
in
the
typing
code
when
this
when
this
takes
place-
and
I
think
it
probably
grabs
the
first
one
here.
A
D
A
The
stuff,
so
if
it's
already
a
module,
deck,
sequential
list,
modulus
or
module
possible
types
equals
get
origin
annotation,
okay,
so
we
have
an
annotation
that
makes
arguments
so
just
type
in
union
check
if
the
value
is
already
in
any
of
the
types
listed
acceptable
types
in
union
by
calling
get
arcs
okay,
so
this
would
validate
this
code
here,
validates
that
it's
either
module
dict,
sequential
modulus
or
module
or
dict
okay.
So
that's
great!
So
it
looks
like
we'll
be
fine
there.
A
Okay,
the
one
place
where
that
might
break
down
is
something
like
the
operation
stuff,
but
that's
not
a
problem
for
now.
So
this
should
be
good.
A
Okay,
so
now
this
is
what
this
looks
like
and
where's
pi
torch
here.
These
are
modules,
dot
container
and
then
modules.
A
Not
modulate,
although
is
that
default
path,
or
how
do
you
import
them?
Can
we
import
these
top.
A
A
A
A
A
A
A
And
then
we'll
just
wait
because
it
looks
like
they'll
merge
it
today.
So.
A
A
This
is
what
I
do
to
check.
Okay,
if
I
can
ever
if
I
can't
remember
how
we've
formatted
commit
messages
before
for
things,
then
I
checked
this.
C
A
This
is
just
like:
do
we
have
a
qualifier
for
something
like
this?
I
don't
think
we
do
requirements.
A
A
All
right
there,
you
go
now
you're
happy
bump
pillow
to
8.3.1,
and
this
was
because
of
this
is
a
array
compatibility
fix.
D
D
A
Related
what
was
our
issue
number?
Do
you
remember
this
shoe
number.
A
A
A
A
A
All
right
so
we'll
bump
that,
and
then
that
means
that
we
want
to
change
that,
commit
there,
okay,
which
was
okay.
It
looks
like
we'll
drop
that
temp
temp
fix
ci
error
patch
and
we'll
change
it
to
this,
and
then
we'll
just
wait
for
them,
because
it'll
come
out
at
some
point
today.
A
B
Oh,
and
would
it
would
the
type
in
the
arts
still
be
any.
A
I'm
wondering
what
to
do
about
this,
because
this
is
obviously
there's
there's
obviously
like
an
open
issue
here.
It's
not
related
to
the
current
issue
that
you're
solving
right,
but
like
what
happens
if
somebody
type
type
casts
something
for
in
a
config
as
any
right
and
and
right
now
you
know
bad
things,
but
after
this
patch
it
works
right,
because
I
assume
it
just
blew
up
before
right.
A
Yeah,
okay,
so
it's
because
so
I'm
thinking
we
may
want
this
commit.
You
know
we
may
want
to
merge
this
commit
as
part
of
this
pr
here.
The
question
is:
what
are
the
implications
of
that
right?
So
the
benefit
is
that
it
doesn't
blow
up
anymore.
You
know,
should
somebody
do
that
the
downside
is,
then,
you
know
the
downside
is
well.
We
need
to
figure
out.
B
I'm
not
I'm
not
sure
why
it
would
still
blow
up,
because
what
this
condition
does
is
it
returns
if
the
value
second
yeah,
if
it
returns,
if
the
value
isn't
a
dictionary,
but
if
it's
otherwise,
it
would
still
blow
up
now,
if
we
don't
return,
it.
A
Yeah
so
well
I
mean,
if
you,
let's
see
you
do
you
basically,
if
it's
a
dictionary,
it'll
be
fine.
Otherwise
you
just
return
whatever
it
is
already
yeah.
So
this
anything
coming
down
from
the
python
api
that
comes
through
here
will
be
fine,
but
anything
coming
down
from
the
command
line
api.
A
A
A
B
Yeah,
but
does
the
argument
type
have
any
in
this
case
when
we're
using
union.
A
Well,
no,
so
this
this
won't,
I
guess
what
I'm
saying
here
is
this
code
won't
be
activated
within
the
context,
so
within
the
so
the
the
pr
that
you're
doing
really
is
about.
You
know
pie
torch
right
now.
We
ended
up
with
this
change
in
base
the
question
and
it
seems
like
we
may
want
to
keep
this
change
even
though
we're
going
to
put
the
pi
torch
stuff
over
to
be
union
based.
Just
as
like
you
know
what.
B
A
B
A
And
yeah,
so
I'm
thinking
is
so
okay.
Well,
we
can
just
run
through
another
scenario.
So
if
we
get
so
say
we
get
something
in
on
the
command
line,
and
it
was
a
let's
see
it's
a
value,
addict
mouse
return
value,
see
we
get
something
in
on
the
command
line.
A
Okay,
because
we
can
get
dictionaries.
We
definitely
know
what
we
might
end
up
with
dictionaries
because
of
because
of
you
know
the
the
at
and
then
the
file
name
for
the
ammo
files,
but
then
they
should
have
just
typed
this.
A
It's
good
to
support
this,
but
it
creates
problems
right
because
then,
all
of
a
sudden,
you
don't
know
certain
data
types,
and
you
really
should
know
them
right
and
then
you're
not
actually
using
the
config
infrastructure,
because
that's
actually
what
happened
here
this
stuff
is
not.
This
should
not
be
like
this.
We
should
be
using
the
config
infrastructure
to
do
this.
A
A
So
what
do
you
say
here
base
convert
value,
support
for
any
type
typing
dot
any
here.
A
A
Okay,
great,
oh,
hey,
wow,
much
faster
way
to
create
an
issue
so
now
we'll
track
this
here,
all
right
and
we'll
put
actually
we
should.
I
should
have
just
put
the
patch
in
the
body.
A
All
right,
so
what
are
our
notes
on
this?
So
essentially,
we
were
gonna
say
currently.
A
If
you,
if
you
put
it
as
any
without
this
patch,
does
it
throw
an.
B
Exception,
I
don't
think
I
tried
without
the.
A
Patch,
okay,
yeah,
you
just
try
you
you
did
the
patch
okay,
okay!
Well,
then,
I
won't
say
I
will
just
say
it
won't
work.
Do
not
support
one
of
properties
that
would
have
been
done
any.
We
need
to
figure
out
how
to
support
this
or
if
we
should
okay,
wait.
So
then
wait.
We
forgot
about
this.
D
B
So
after
after
putting
the
any
into
field,
you
mean
it
blows
up.
B
Yeah,
it
blows
up
because
you're
kind
of
trying
to
create
a
dick,
it's
expecting
a
dictionary
yeah
from
you
and
you're
sending
other
things
in
so
yeah.
It
blows
up.
A
Okay,
great,
so
we
need
to
figure
out
how
to
support
this.
Ideally
developers
would
put
the
correct
type.
A
A
A
A
Because
if
we
can
get
the,
if
we
can
create
config
objects
out
of
the
definitions
or
the
input
definitions
of
operations,
then
all
of
a
sudden
we
can
call
any
operation
from
the
command
line
very
easily
and
and
essentially
an
operation
becomes.
You
know
everything
becomes
the
same
thing.
So
all
our
command
line
classes
can
become
operations,
and
we
don't
have
to
do
that.
Whole
class
based
approach
and
this
everything
is
in
operation
and
if
all
the
command
line
things
are
operations.
A
So,
okay,
high
level
is
a
good
example
right.
So
we
have
all
this.
You
know
we're
doing
all
this
stuff
in
high
level
right
and
these
are
well.
These
are
pure
python
functions.
We
also
have
all
this
stuff
in
cli
right
ml,
and
this
is
basically
you
know.
These
are
calls
into
high
level,
but
really
they
just
wrap
it
at
such
a
minuscule
level.
Here
that,
like
it
should
just
be,
you
know
we
should
just
be
calling
this.
A
We
should
just
be
calling
high
level
accuracy,
and
this
is
just
a
python
function,
so
we
should
be
able
to
decorate
it
with
op
and
if
we
can
decorate
it
with
op
and
op
also
creates
config
for
the
inputs.
Then,
all
of
a
sudden,
we
don't
even
need
this
whole
file
right,
because
all
of
this
stuff
just
becomes
just
becomes
basically
a
pass
through
to
the
high
level
code.
A
A
A
B
A
A
All
the
even
if
it's,
if
it's
defined
through
a
python
api
if
it
comes
to
the
http
api,
if
it
comes
through
the
command
line,
it
all
comes
through
here.
So
when
it
comes
to
the
http
api,
it
could
possibly
be
you
know
a
list
or
a
dict
or
a
float
or
right.
It
could
be
any
of
the
standard
types
so
and
that's
fine
right,
so
type
class
equals
dict
and
then
it'll
do
the
right
thing.
A
So
if
okay,
if
n
or
so,
I
think
we
might
want
it
is
list
here-
I
think
we
might
want
to
do
is
instance
list
just
to
to
cover
that
case
in
case
anarchs
is
also
a
thing,
so.
A
A
God
this
is
such
it's
so
hard
to
do
all
this
config
stuff
and
that's
why
the
code's
such
a
mess.
So
I
think
you
know
I
think
it
may
be
fine
with
so
if
the
type
is
a
list.
So
if
we
happen
to
get
a,
this
is
the
only
place
where
this
you
know
this
could
get
tricky.
I
think,
is
if
it's
a
list
or
a
tuple.
A
Now,
through
the
http
api,
you
might
get
a
tuple
through
the
part,
http
api,
you
would
may
get
a
list.
You
may
get
a
list
through
the
python
api.
You
may
get
a
list
or
a
tuple
through
the
python.
Api
you'd
never
get
a
tuple
through
the
http
api
and
in
the
case
that
it
is
a
list
that
you're
getting
passed
in
then,
if
nrx.
A
So
we
skip
this,
I
think
we
may
just
need
to
add
the.
If
lister
is
if
tupple
case.
A
A
A
B
Okay,
which
function
is
this:
what's
the
name.
A
This
is
so
we're
in
convert
value
and
dfml
base
py
on
line
161..
A
So
because
you
we,
the
the
patch
you'd
added,
says
you
know
if
argh
type
is
any
then
check
if
we're
an
instance
of
check,
if
we're,
in
instance,
of
dict,
in
which
case
set
type
class,
equals
dict
and
it
looks
like
anarchs
only
gets
set
if
it's
a
list
or
a
tuple.
So
we're
not
going
to
hit
that
we're
not
going
to
hit
this
dot
load
because
it's.
A
A
A
B
Yeah,
so
this,
if
convert
value,
is
equal
to
type
class
value,
what
does
this
type
class
do?
Because
this
is
the
one
that's
that
gets
called
and
blows.
A
A
Well,
let's
see
so
where's
type
lookup,
so
I
think
this
does
okay.
So
I
think
this
just
does
a
python
lookup
that
just
that
just
casts
it
to.
For
example,
if
you
had.
A
It's
just
it's
it's
like
saying,
for
example,
so
if
type
class
was,
this
would
be
value,
equals
dict
value
right.
A
So,
in
which
case
you
know,
and
and
if
values
already
addict,
then
we
wouldn't
need
to
do
that.
So
I
think
we
can
probably
just
safely
return
value.
A
Okay,
okay!
Well,
I
guess
we'll
just
take
equal
signals.
Okay,
I
think
we
can
safely
return
value.
Okay,
great!
I
think
we
figured
this
out,
so
this
is
sort
of
a
separate
thing
now.
A
That
it's.
A
All
right
all
right,
so
this
is
sort
of
separate
thing.
So
let's
just
take
that
out.
A
A
A
A
A
D
A
Okay-
and
this
was
okay-
there's
that
fix
anyways
okay,
so
you
had
it
like
this
right
now.
Well
now,
I'm
kind
of
concerned-
because
you
had
it
like
this
and
then
you
changed
it,
so
maybe
it
didn't
work
out.
I
didn't
see
that
so.
A
A
A
A
Now,
let's
take
a
look
at
this,
so
we
have
the
pillow.
We
have
the
typing
in
e,
which
we're
going
to
change
in
a
second
here
and
okay.
This
is
not.
This
is
not
us,
so
all
right.
So
then
we
basically
just
determined
that
this
needs
to
be
the
union.
D
A
A
Oh-
and
this
is
our
union
fix
here,
but
okay,
we
just
need
to
import
union
all
right.
This
thing
is
there:
we
go
okay,
whatever.
A
B
After
changing
the
field
to
union,
we
also
need
to
check
if
the
type
in
arbs
is
union
like
before
we
were
checking
if
it
was
any
okay.
A
B
I
mean
if
it's
union,
we
still
want
to
do
the
same
thing
right
like
we
were
doing
with
any.
A
Well,
that's
what
I'm
wondering
here
is
is
well.
If
it's
union,
we
may
want
to
try
to
do
convert
value
for
each
kind
of
type
in
the
union,
so,
for
example,
right
so
if
we
so,
let's
see
type
field
type,.
A
A
So
I'll
run
we'll
run
through
the
code
and-
and
I
think
it'll
make
a
little
more
sense
here
so
of
what
I
was
saying
so
basically
so
you're
you're
saying:
okay,
well,
okay,
so
we
we
had
just.
We
just
had
this
right.
Essentially.
B
A
A
Yeah,
okay
operations,
love
this
stuff;
okay,
okay,
oh
here
way,
this
is
interesting,
probably
towards
neural
network
config
union
network
model,
so
it
is
being
used
here.
A
Okay,
so
we
have
a
config
class
that
uses
union
already,
which
is
interesting
so
now
the
question
is:
does
this
actually
get
called
anywhere?
There's
no.
B
A
Okay
and
what
was
that
this
is
the
neural
network
config
and
then
it's
network,
so
pi
torchnet
network,
okay,
so
we're
loading
from
a
dictionary
in
that
case,
so
a
network
network
loading
from
a
dictionary
loading
from
a
dictionary,
ruffles
ruffles
ruffles
we're
looking
for
a
dictionary.
A
So
once
again,
I
think
this
has
been
sidestepped
because
sidestep
simplified
because
we're
only
the
the
current
usage
is
just
a
dictionary,
and
so
we
haven't
run
into
anything.
A
So,
in
that
case,
current
usage
is
a
dictionary,
so
this
mk
arg
must
be
functioning
correctly,
because
obviously
we
make
an
argument.
We
make
you
this.
The
args
are
an
intermediate
representation
that
we
had
to
hold
over
from
the
old
code.
So
everything
goes
through
mkr
based
on
and
the
field
type
and
gets
these
oh
actually
yeah.
It
actually
creates
the
arg
parse
version
of
these
things
for
the
command
line.
A
A
Ruffles
ruffles:
oh,
let's
see.
A
Maybe
we
should
just
let's
just
throw
it
in
the
ci
and
see
what
happens,
because
if
it,
if
it,
if
it
doesn't
blow
up
right
now
with
a
union,
I
mean
I
don't
I
don't,
we
can
dig
deeper,
but
we
can
also
just
you
know,
throw
it
at
the
wall
and
see
if
it
sticks
so.
A
D,
where
was
it
you
were
doing?
What's
that
example
that
it's
related
to.
A
Okay,
do
you
have
code
for
it?
Can
you
make
a
pr.
A
As
usual,
we
always
forget
right.
So,
let's
see
related,
you
know
to
put
your.
A
Sticked
to
pie
torch.
A
A
Okay,
buddy,
okay-
and
this
is.
A
A
A
I'm
liking
this
command
line,
client,
okay,
so
great
all
right,
cool!
We're
done
here!
Sorry
that
took
so
long,
but
I
think
we
I
think
we
figured
out
a
lot
of
stuff
so.
A
C
A
A
All
right,
okay,
I
feel
like
there's
something
we're
forgetting
here,
so
basically
we
recorded
the
change
we
recorded.
We
documented
the
thing
that
we
just
found
in
an
issue
which
is
the
support
for
typing.
any.
No,
it's
open
still
open
right.
We,
oh,
I
remember
what
the
last
thing
was.
Okay,
so,
and
I
think
sakshan
may
have
an
open
issue
for
this
anyways,
so
this
should
not
be.
This
should
not
be
this
way.
A
A
A
Yes,
where
is
this,
so
the
config
infrastructure
doesn't
have
support
for
doing
nested
dictionaries
at
the
moment,
which
is,
I
think
why
this
is
the
way
it
is,
but
if
we
don't
have
an
issue
to
track
that,
we
definitely
need
an
issue
to
track
that.
A
Special
eye
command:
okay,
now
that's
not
the
one.
This
might
be
the
one
which
is
really
poorly.
A
A
A
In
we
have
something
that
does
this,
I
think
it
was
the.
A
What
do
you
guys
remember,
there's
an
issue,
this
issue
with
a
bunch
of
command
line
flags
in
it.
A
Okay,
well,
what
okay?
What
is
this
thing?
I
would
like
to
get
this
recorded
here
so
all
right,
it
doesn't
really
matter
for
the
context
of
this,
but
I
don't
want
to
forget
to
go
do
this,
so
where
is
let's
just
see
if
there's
something
for
model
pi
touch
to
track
this?
I
think
he
had
so
add
python
api
support
for
defining
last
layers
when
using
pre-trained
models.
So
is
this
what
you
just
did.
A
Yeah:
okay,
great
okay,
and
did
we
have
a
fixes
tag
on
that.
A
This
we
should
always
put
it
in
the
in
the
commit
message.
A
A
A
A
A
A
I'll
just
add:
let
you
amend
that
commit
when
you,
when
you're
finished
with
that,
I
just
thought.
If
you
just
done
it,
then
we're
good
great,
all
right.
Okay,
long
long
talk
there,
but.
A
All
right,
I'm
really
pissed
off
that.
I
can't
find
that
issue
but
whatever-
and
this
says
we've
mentioned
it
right-
great
okay-
so
now
it
references
there.
I
love
linking
all
these
things
together.
It
helps
us
figure
out
what
the
hell
happened.
Okay,
we're
good
right!
I
think
we're
good
on
that
one.
A
So
you'll
you
will.
A
A
A
So,
oh
sudhanshu,
do
you?
Do
you
remember
what
issue
do
you
remember
what
issue
tracks
the
there's,
an
issue
that
tracks
the
support
for
dictionary
data
types
in
the
config
code,
we
were
trying
to
find
that.
A
A
So
this
guy,
let's
see,
I
think
I
rebased
things
around
to
make
it
look.
The
commitment
message
is
a
little
prettier
or
something
yeah
yeah.
I
think
I
just
rebased
the
comment
message
is,
but
then
we
started
seeing
weird.
What
was
this?
A
A
A
A
A
The
title
is
okay,
so
this
is
like.
A
A
model
pie
torch
failure
to
numpy,
or
was
this
wait
a
minute?
Is
this
yeah?
Okay?
So
this
is
the
pillow
thing:
okay,
numpy.
A
C
A
What
else
do
we
have
in
a
bug
and
then
we'll
just
cat
the
body
in
so.
A
A
A
A
Wow,
that
was,
that
generated
a
lot
of
stuff.
Okay,
so
we'll
wait
on
the
ci
for
this
one
and
then
after
that,.
A
A
And
then
I
think
we
had
yeah
okay,
great
airbase,
perfect.
Oh
man,
I'm
loving
this
thing.
Okay!
So
then,
so
we
merged
this
great
job.
Nice
all
right
excited
to
see
that
on
the
doc
site,.
A
Great,
so
we
merged
that
add
support
for
additional
layers
in
python.
Oh
my
god,
something.
A
Merged
this
one
we're
waiting
on
ci.
A
Rating
on
ci
fixes
numpy,
okay
issue
with
pie,
torch
or
pillow,
and
then
this
is
waiting
actually
we're
waiting
on
release
from
pillow.
A
A
A
A
Cool
and
wait
a
minute.
A
A
B
Still
working
properly,
it's
quite
messy
and
I'm
I'm
still
waiting
on
some
a
few
queries
from.
B
A
A
Great
so
we'll
put
that
saksham
is
reviewing
that
so
transfer
learning.
A
B
D
Yep,
it's
going
late,
all
right
so
yeah,
so
I
just
wanted
to
let
you
know
that
there
are
two
pull
pull
requests
which
are
ready
for
review.
Okay,.
A
A
A
Crap
all
right
all
right.
A
A
C
B
D
C
A
Would
like
so
we
can't,
we
don't
want
to
have
the
add
score,
commits
and
add
score
scripts.
These
last
two
be
first,
because
if
we
applied
those
individually,
they
wouldn't
work
right.
So
because
this
is
like
what
you
did
here
is
first,
you
added
the
base
classes
right.
Yeah,
then
you
add
the
psychics
quarters
classes.
A
A
A
D
So,
actually,
when
I
was
like
writing
the
tests
for
them,
I
actually
found
that
the
default
pause
level
was
actually
one,
but
we
have
to
keep
it
either.
Two
or
four.
D
D
So
initially
I
was
thinking
like
to
make
it
a
config,
but
then
I
thought
it's
mostly
like
a
constant
value.
C
C
D
A
A
Okay,
so
let's
export
that
positive
and
actually
the
thing
is
so,
let's
also
so
we
talked
about
so
pause
label
set
to
2
for
some
scores.
A
Scorer
keyword
arguments
via
config,
so
let's.
A
A
A
Nothing
yep.
Okay,
it's
a
default
all
right.
So
so
we
have
the
functional
subpartial
method,
so
we
should
just
be
able
to
do.
A
D
A
So
it
did
work
yeah,
okay,
great!
Let's
give
it
a
shot
here
right,
so
making
fig
numpy.
A
A
A
Config,
so
if
so,
let's
just
move
this
stuff
over
here.
A
So-
and
this
I
think
will
be
will
be
a
little
cleaner
because
then
we
have
specific
class
names
in
in
the
in
the
part
where
we're
defining
specific
classes
and
then
we're
gonna
set
pause
label.
A
We
override
that
default
value
of
1,
two
two
two,
so
why
are
we
doing
that
again?
Two,
because
we're
doing
in
we
wanna.
A
A
D
And
yep
scorer
from
the
last
has
to
be
dropped.
If
you
are
doing.
A
A
A
So
properties,
vehicles,
none
property,
setups,
numpy
string,
all
right!
Well,
that's
kind
of
problematic,
so
properties
is
none,
numpy,
dot,
string,
arg!
So
right
now
it
looks
like
we're
overriding.
Well,
that's
annoying!
So
right
now
it
looks
like
we
are
taking
the
we
can
define
properties,
but
then
we
immediately
override
them
like
with
the
numpy
style,
doc
string,
arg
so
or
if
properties
equals
is.
A
Okay,
yes,
so
let's
not
do
that
right.
So
that
was
an
oversight,
so
let's
fix
we'll
fix
that,
but
so
properties,
okay,
so
basically
properties
equals
pause
label.
A
A
A
A
C
D
A
A
This
doesn't
tell
us
what
to
put
as
a
stir
value:
okay,
that's!
Okay!
We
have
the
source
code,
so
f,
beta
score.
D
I
don't
know
maybe
like
we
can
put
our
integer
value
in.
A
A
A
A
A
So,
if
possible,
that
looks
like
a
int
to
me
yeah,
I
don't
know
why
they
have
a
stir
annotation
on
that
all
right,
great
sorry,
just
want
to
make
sure.
A
A
D
A
Properties
equals
doc
rams
else.
Dark.
Darkparams.Updateproperties
now
will
override
right.
You
know
the
supplied
value
of
properties
will
override
the
defaults.
So
then,
okay,
we
call
make
config
and
doc.
Params
here
is
going
to
be
nice.
Great
helpful
error,
messenger
talk,
string,
john
dark
params
looks
like
it
is
such
a
chain
cleanup
description,
okay!
Well,
it
must
be
something
semi-helpful
to
okay
to
field
okay,
so
it
does
type
and
then
field.
A
So
we
just
do
this:
numpy
dot
string,
args
on
method.
We
can
override.
A
One
all
right,
so
this
is
the
field
dot
default.
A
A
Equals
what
was
it
two?
We
were
sitting
at
two?
Yes,
all
right,
great,
okay,
we'll
just
do
this
either
way
here.
A
Default
equals
two
and
then
we'll
just
pass
properties.
Okay,
so
in
this
way
I
mean
we'll
end
up
parsing
the
dock
string
twice,
but
whatever.
A
A
A
Okay,
when
we
do
parsing,
it's
going
to
it's
going
to
want.
Why
true
and
wipe
red
in
the
config,
and
obviously
we
don't
want
to
take
that
in
the
config,
so
we'll
probably
want
to
say
we'll
probably
want
to
do
to
make
dark
string
args.
A
So
we
may
just
want
to
like
do
this,
make
config
ourselves
here
so
make
yeah.
We
do
we
just
want
to
do
this
ourselves,
so
we
won't
call
actually
make
numpy
config
because
we're
doing
too
much
custom
stuff
here,
so
we
don't
want
to
require
what
are
they
always
they're
always?
Are
they
always?
Why
true,
and
why
pred.
A
Are
always
the
why
the
first
two
arguments
to
score
right
or
to
the
method
which
will
provide.
A
A
For
key
comment:
we
don't
care
about.
The
type
filled
in
properties
died.
A
A
Okay
and
produce
to
remove
if
field
dot
default
equals.
What
is
that
thing
missing.
A
A
Okay,
is
there
a
better
way
to
do
that?
I
wonder
deleting
a
bunch
of
things
clear,
deleting
a
bunch
of
things.
At
the
same
time,
move
specific
key
update
values.
D
A
All
right,
let's
see,
let's
see
what
happened
here.
So
we
moved
the
pause
label
stuff
into
the
psychic
scores.
We
are
creating
a
new
word.
We
create
a
config
for
each
score.
A
A
A
A
A
D
A
All
right,
whatever
my
current
environment
just
got
nuked,
I
think
so
all
right
so.
A
Okay,
so
these
are
all
our
scores
now
they
have
configs.
Let's
go
back
and
look
at
everything
else
here.
A
All
right,
so
we
moved
that
stuff
out
of
there.
This
all
looks
good
entry
points,
class
names,
testing.
A
A
Great
great
create
a
great
of
records,
class
model,
turn
name
and
model.
Dir
is
okay,
so
temp
file
temporary
directory
okay.
So
we
don't
want
to
do
that
so
test,
scikit
model,
okay,
so
also
this
needs
to
derive
from
so
this
needs
to
derive
from
you
know,
async
test
case
here.
A
Okay,
no
okay!
I
see
you're
deriving
from
async
test
case
there.
Okay,
it's
a
second
model;
okay,
so
you're
creating
a
mix.
Is
there
any
reason
why
we
wouldn't
just
derive
from
async
test
case
here.
D
A
A
A
We
don't
need
this
anymore,
let's
see
so
I
think
different
util
is
a
test
case.
I
think
that
we
probably
I
mean
those
psychic
plugins
are
old,
so
we
have
had
many
things
happen
between
now
and
then.
A
A
Oh
okay,
I
don't
like
I'm
not!
I
don't
like
this
all
right,
but
the
psychic
classes
put
themselves
in
okay.
This
means
the
way
that
this
is
done
right
now.
It
looks
like
oh
okay,
but
because
for
every
class,
okay,
yeah,
okay,
okay
right
because
we're
creating
new
instances
of
these
new
instances
of
these
classes,
all
right,
never
mind,
never
mind,
never
mind.
A
D
So
in
the
original
tests
it
was
electric
like
that.
A
So
did
we
should
we
should
we.
A
A
A
D
Yes,
we
can
do
that,
but
we
so
we
had
actually
a
discussion
on
whether
we
should
separate
the
models
from
these
scores.
A
C
D
We
have
to
like
separate
the
models
and
the
scooters
then.
A
Well,
I
think
I
think
yeah,
okay,
good
good
thinking,
good
thinking.
I
think
that
looking
at
the
code,
I
think
you
know
what
we,
if
I
remember
correctly,
we
talked
about
that
right
and
then
we
were
like
okay,
if
the
code,
because
because
you're
just
really,
I
mean
the
reason
for
doing
that
is
separating
out
dependencies
and
like
lines
of
code,
and
I
mean
the
scorer
implementation
is
what,
like
you
know,
model
sidekick
different
one.
A
second.
A
Model
score
score
base,
scores
251
lines,
so
I
don't
think
that
warrants
a
separate
package
at
this
point,
especially
I
I
think
I
think
we'll
just
leave
it
all
together.
So
I
mean,
unless
you
see,
does
do
you
see
more
work
happening
here
in
the
scikit
scores.
A
Okay,
so
in
this
case
I
would
say,
let's
just
start
importing
stuff
from
the
other
file,
because
I
don't
think
we'll
split
it
out
into
its
own.
It's
its
own
plug-in
yeah.
I
don't
think
we'll
split
it
out
until
it
sounds
plug-in,
so
I'm
gonna.
A
A
A
A
A
Okay,
we're
gonna
validate
that
the
score
stuff
works.
Okay,
that
the
config
stuff
works.
Add
tests
for
second
scores
all
right,
so
these
are
the
tests
and
okay.
So
that's
we're
gonna.
Just
you
know,
reuse
code,
where
appropriate
there.
That's
good!
That's
good!.
A
A
We
need
kind
of
like
one
of
these
ci
checks
that
we
keep
adding
so
ci
our
tests,
like
till
testing
test
or
validate
that
all
classes
or
plug-ins
of
a
given.
A
In
are
registered
as
entry
points
anyway,
because
we'll
need
this
we're
trying
to
think
about
ci
tests
that
will
go
across
all
the
plugins
and
making
sure
that
you
have
listed
it
in
the
entry
points
file
if
you've
defined
it
as
a
class
that
so,
if
you,
for
example,
inspect
everything
under
model
scikit
and
if
we
find
that
we
do
inspect
on
everything
under
model
scikit,
and
we
find
anything
that
subclasses
from
accuracy
score
that
doesn't
get
is
not
in
this
list,
then
we
would.
A
You
know
the
test
would
fail,
just
just
so
that,
because
I
see
this
is
I
mean
this
this?
This
is
gonna
become
a
problem.
We.
D
A
Will
we
will
inevitably
forget
to
create?
We
will
for
inevitably
forget
to
define
to
find
one
of
those
or
to
export
it,
but
we'll
have
written
it.
So
so,
okay,
so
inspect
all
objects
in
a
package
or
module
and
see.
If
and
the
entry
points.
A
For
that
package
to
verify
fail
test
if
and
entry
point
a
or
class
plus
was
defined,
but
not
exported
or
registered
all
right.
D
A
A
Okay
and
then
so.
A
Hello,
overwriting,
dark
stream,
parsed
config
parameters
with
fields
and
types
given
via
keyword,
argument.
D
D
A
A
A
A
A
A
A
Okay,
all
right
all
right,
all
right,
I
get
it,
I
get
it,
it
doesn't
work.
Okay,
let's
just
try
this
one.
A
A
A
All
right
great,
so
that
seems
to
be
happy,
except
for
something.
A
So
we
provide
config
objects.
A
Score
so
we've
provided
a
config
for
each
score.
We
will
have
the
test
case,
pull
variables
and
subclasses
in
the
original
test.
A
A
A
All
right,
okay,
we'll
do
that
because
we're
you
know
way
over
here,
so
all
right!
Anything
else.
Did
we
get
ci
passing
or
failing
on
anything.
We
just
did
yes,
you've,
you've
you're
using
the
test
book
here.
B
A
Okay,
cool,
I
think
it's
failing
to
master
too
based
on
that
release,
so
we
can
probably
just
safely.
Is
it
just?
Is
it
just
the
yeah?
Is
it
just
that
that's
failing
and
that
been
that
pr
like
it
can't
install?
Is
it
mad
because
it
can't
install
it.
A
A
All
right!
Every
single
thing
fails:
okay,
let's
not
merge!
This
like.
I
was
about
to
say,
let's
just
merge
this,
but
we
will
hold
off
on
that
until
the
later
in
today,
all
right
so
and
then
I'll
say
on
john
to
review
offline
great
and
we
need
to
say
that
we
updated
updated
or
make
config
numpy
to
did
it
make
config
numpy
to
allow
for
overriding
parsed
fields,
all
right
cool.
A
I
think
we're
good
here
anything
else
from
anyone.
Final
thoughts,
all
right.