►
From YouTube: 2020-01-14: Weekly Sync
Description
Meeting Minutes: https://docs.google.com/document/d/1vKYEPtqKiwsFwhVKPmPub5ebMqN9HteBcbdFAuTXalM/edit#heading=h.du0souabgio
A
B
B
B
B
Okay,
that's
what
I
thought
yeah,
I'm
sorry
about
that
yeah!
That's
that's!
No
fun!
I
ran
into
the
same
problem
at
my
on
my
work
machine,
one
of
my
work
machines.
Last
week
I
just
switched
to
using
a
different
machine.
B
That's
always
completely
hosed.
I
have
no
idea.
I
tried
reinstalling
everything
several
times.
I
just
I
have
no
idea
what's
wrong,
but
yeah
so
did
you?
Have
you
tried?
Let's
see
I
was
thinking.
Maybe
you
could
share
your
screen
and
we
could
go
through
it
here
to
see
what's
up
and
then
we'll
try
to
document
that
because
I'm
sure,
if
you
ran
into
it-
and
I
ran
into
it-
somebody.
D
A
B
C
A
Is
my
skin
visible?
Yes,.
A
Okay,
so
I
tried
this
with
like
a
docker
too
yeah.
E
B
Wow,
what
is
up
with
that?
Okay,
so.
B
Wow,
that's
really
weird
that
it
happens
in
docker
too
no
repos
to
train
on
okay.
A
Did
get
clean
hdfn
and
I
reinstalled
it.
It
still
didn't
work.
B
Okay,
so
because
okay
yeah,
so
then
I
guess
in
this
case,
let's
not
try
that
again,
how's
it
going
again,
oh
yeah,
so
we
got
ogen
and
we
got
himanshu
on
here
as
well,
we're
going
through
saksham's
issue
right
now
with
the
development
environment.
It's
like
something
something
seems
to
be
a
miss
here,
because
he
can't
train
can't
train
the
model.
It's
saying
that
there's
no
repos
in
the
source,
but
I
can
do
it
on
my
machine.
B
Oh
yeah,
well,
this
is
the.
I
think
this
is
a
classifier.
Actually
I
wrote
this
one,
but
so
it's
my
fault,
then.
B
B
Okay,
so
I
will
just
paste
this
into
getter
here:
let's
see,
okay,
actually
it
looks
like
the
source
test.
Oh,
the
source
test
is
failing
because
of
the
entry
point
stuff
that
got
updated.
A
Yeah
I
tried
they
didn't
get
bored.
The
entry
point
error
was
showing.
B
Okay,
okay,
you
know,
let's
see
so
you
tried
it
in
git
pod
and
it
didn't
work
too.
Actually,
let's
try
that's
interesting.
So.
B
B
You
know
we'll
have
that
entry
point
stuff
fixed,
so
cd
documents,
python
dffml.
B
All
right,
okay,
so
I'm
gonna
pull
it
down.
I'm
gonna
verify
it
works
here.
For
me,
let
me
share
with
you
guys
real
quick,
so
you
guys
can
see
what's
going
on
over
here.
B
B
B
C
B
B
B
What
could
be
wrong
with
the
docker
container?
Oh
there's
no
attribute
read,
write.
That's
right.
We
changed
that.
B
What
was
it
allow
allow
empty
empty.
E
A
B
Okay,
so
now,
if
you
guys
want
to
try
those
commands
to
pull
it
down,
we
can
see
if
that
works
on
everybody
else's
system,
so
I'll
post
up
this
guy.
Here
too,
then,
why
don't
you
also
try
to
pull
that
down
saksham
and
let's
see,
watch
your
screen
and
see
if
that
works
there
and
then
I'll
watch
the
ci
test
too.
Here,
let's
see
okay
and
then
I'm
gonna
think
about
what
could
be
going
wrong
with
the
docker
container.
C
It's
linux.
I
changed.
A
B
A
Thank
you,
okay,
have
you
pushed
it.
B
B
B
B
Oh
no
entry
point
not
found
just
because
of
the
classification
command
because
it
doesn't
have
colons
in
it.
So
it's
trying
to
load
it.
So
that's
which
is
we
need
to
get
rid
of
that,
but
I
think
I
can't
remember
what
his
name
is.
The
other
yash
is
is
gonna
work
on
that,
I
think
so,
let's
see
so
we
just
need
to
add
like
next
to
oh
and
actually
I
think
it
changed.
The
model
predict
anyways,
didn't
it
or
no.
B
B
So
if
you
change
the
model,
classification
parameter
to
be
label
colon
into
colon
one
and
actually
check
this
out.
Do
control
x,
control
e
control,
x
or
ctrl
x
and
then
e
with
the
up
arrow.
First,
let's
see,
I
wonder
if
it'll
work
for
you
yeah
yeah,
so
that
opens
your
thing
in
the
editor.
Sometimes
it's
easier
to
fly
around
there.
I
just
somebody
just
told
me
this,
so
I
thought
I
would
share
so
label
model.
Wait.
You
went
too
far,
I
think
so
yeah
so
label
and
then.
B
What's
yeah,
it
should
be
model
classification
label,
then
colon
into
colon
one.
B
B
Okay,
so
far
so
good
over
here,
yeah
things
look
okay
for
me
over
here.
Let's
see
what
happens
and
I'll
update
those
docs
pages
here
yeah
and
then,
if
you
just
exit
the
editor,
it'll
run
the
command.
B
What
the
hell,
what
the
okay,
the
only
thing
I
can
think
of
at
this
point
is
the
hash
values
of
those
files
are
incorrect.
What
happens
if
you
do
sha
384
some
space
train
images,
you
byte,
can
you
repeat
that
so
do
sha,
384
sum
all
one
word
and
then
space
and
then
one
of
those
file
names
or
both
of
the
file
names.
Both
the
train
file
names.
You
just
use,
train
hyphen,
star,
wait.
They
start
with
train,
so
just
do
train
hyphen
star
and
then
hit
enter.
B
B
Oh
no,
those
are
correct.
Let's
see,
yeah
check
out
the
test
case
there
under
tests
yeah
up
up
up
yeah
there.
We
go
so
yeah
open
that
guy.
So
these
scroll
down
a
bit.
B
And
then
okay,
so
these
hashes
should
match
if
it's
the
right
file
data.
Well,
I
guess
it
is.
I
mean
it
yeah
just
make
sure
those
hashes
match.
Basically
so.
B
What
the
hell
could
be
the
wrong
here.
I
guess
what
what
happens
if
you
look
in
the
okay,
what
happens
if
you
look
in
dffml
sources
or
source
source
py.
B
B
Yeah,
so
what
happened?
Yeah?
What's
up?
Okay?
What
is
up
with
this
all
right?
Let's
do
let's
do
like
right
above
that
make
a
new
line
right
above
that,
if
statement
at
the
same
indentation
level
yeah
and
do
just
like
print
repo
dot
features
and
call
the
features
method,
so
dot
features
open,
paren,
close
paren
and
then
let's
see
yeah
so
parentheses,
sorry
not
brackets.
B
There
you
go!
So,
let's
just
just
look
at
that
and
see
what
happens.
We're
probably
gonna
get
a
whole
lot
of
crap,
so
you're
gonna
have
to
control,
see
it
or
something,
but
just
like
wait
until
it
starts
printing
massive
stuff.
D
B
B
B
D
B
B
Sorry
I
was
trying
to
copy
paste
something
it
didn't
work
all
right.
Let's
see,
yeah
console.
B
B
E
B
B
B
A
B
B
Mine
is
combining
because
it's
getting
through
here
and
it
wouldn't
you
know
it,
wouldn't
it
wouldn't
train
unless
it's
combining.
So
I
think
maybe
just
try.
B
B
All
right
so
we'll
just
we'll
just
wait
for
this
to
start
up
and
let's
jump
to
somebody
else
for
now
and
see
if
we
can
loop
back
on
this
well,
you
tried
in
git,
pod
you're
gonna
need
to
figure
out
how
to
like
add
your
branch,
and
you
know
you'll
need
to
do
like
get
remote
ad
and
stuff
to
add
your
branch
and
then
pull
it
down,
but
yeah,
okay.
So,
let's
see
okay,
so
who
else?
Who
else
is
on
the
call
here?
I
think
we
have
somebody
new.
B
B
Is
that
it
might
have
been
this
guy?
Oh
you
guys
can't
see
my
screen.
Can
you?
Oh?
No,
you
can't
okay
anyways
yeah.
So
how
are
you
guys
doing
again
and
I
have
been
working
furiously
on
that
on
the
dataflow
stuff
over
there?
So
I
think.
F
Like
how
do
you
know
what
function
runs
when
you
enter
a
cli
command.
B
Document
stuff
about
the
code
based,
and
so
if
you
guys,
let's
see
if
you
guys
want
to
add
things
to
this
area
here.
Basically,
so
I
changed
the
the
contributing.md
to.
I
put
it
on
the
website
to
make
it
a
little
more
consistent
since
the
rest
of
the
documentation.
Is
there-
and
I
added
this
section
about
like
code
based
layout
and
notes
so,
and
this
is
like
a
lot
of
stuff.
B
So,
basically,
what
I
did
is
I
just
outlined
the
fact
that
you
know
we're
plug
everything's
plugin
based
plugins
are
in
their
respective
directories,
and
then
I
outlined
that
I
put
a
note
about
the
double
context:
entry
pattern,
because
I
know
that
that's
like
a
common
thing.
B
That
is
it's
that's
new,
so
and
then
the
other
thing
was
what
I
wanted
to
do
was
write
some
more
notes
about
the
various
subsystems,
and
it's
like
okay,
like
this,
is
what
we're
doing
with
sources
and
and
here's
how
the
here's,
how
you
you
know,
here's
some
tips
and
tricks
right
and
here's
some
things
you
got
to
watch
out
for
and
here's
you
know
the
way
you
run
some
test
cases
and
stuff
like
that
for
their
respective
plug-ins,
as
as
you
run
into
them,
so
basically
like,
if
you
think
of
things
where
it's
like.
B
Oh,
I
wish
I
wish.
I
I
keep
copy
pasting
this
or
something
like
you
might
want
to
write
it
in
here.
We'll
just
add
it
to
the
main
documentation,
but
so
on
this
note
of
how
do
cli
commands
work.
This
is
something
that
would
go
here
right
like
how
does
the
cli
work
and
the
answer
is
so
the
answer
on
this
one
check
out:
oh
well,
this
is
up
to
date
with
master
now.
So
let
me
make
the
text
a
little
bigger
here.
B
B
Sorry,
all
right
and
so.
B
B
Okay-
and
I'm
gonna,
preface
this
by
saying
this
is
so
that
we
can
tie
in
with
like
the
arc
parsed
argument,
parsing
and
stuff
and
then
like
it,
does
some
output
formatting
so
that,
basically
I
don't
know
if,
if
you
guys,
if
you
return
anything
from
a
cli
command,
it'll
just
get
json
dumped
and
that's
because
so
many
things
understand
json
that,
like
if
you're
running
this
as
a
sub
process-
or
you
know
you
might
be
running
it
from
php
or
something
you
can
just
json-
load
the
input
back
in
and
then
you
know
you
don't
have
to
do
a
bunch
of
parsing
so
but
the
way,
the
way
that
it
works
is
we've
got
this
abstraction
layer
around
around
argpers,
where
similar
to
like
how
there's
all
these
classes
with
their
config
objects.
B
This
is
from
before
that.
So
it's
all
it's
eventually.
What
we
should
do
is
is
con
consolidate
that
code,
those
code
buses,
so
that
we
use
we
we
can
make.
We
just
like
assign
a
config
object,
and
that
would
be
like
all
of
the
arguments
that
go
to
this
thing.
But
this
is
all
holdovers
because
it
hasn't
gotten.
You
know
it
hasn't,
got
an
update
to
use
that
more
clean
way
of
doing
it.
B
So
the
way
that
it
works
is
you
basically
say
you
derive
from
this
command
class
and
you
implement
the
run
method
so,
and
what
this
allows
us
to
do
is
it
makes
it
really
easy
to
write,
commands
that
have
similar
arguments
without
rewriting
a
bunch
of
code
and
that's
why
it
works
this
way.
So
if
we
look
at
like
dffml
cli
ml
dot
py,
so
let
me
open
diff,
no
utils
commands.
B
So
there's
like
a
couple
commands
in
this
commands
file
that
basically
like
make
it
easier
for
anytime.
You
do
something
with
sources
you
would
just
like
derive
from
this
subclass.
That
has
the
sources,
and
then
it's
already
filled
out
this
arg
and
if
you
guys
have
seen,
I
can't
remember
how
long
everybody's
been
around,
but
if
you
saw
the
args
and
config
methods,
you
would
know
that
you
have
to
redefine
all
these
every
time
you
use
a
class
that
has
the
same
type.
B
You
would
have
to
like
put
in
this
arg
structure
and
say:
okay
like
I
want
to
use
this
argument,
that's
of
type
source-
and
you
know
basically
do
all
the
art
stuff
for
it
so
that
it
could
come
in
from
the
command
line
or
it
could
come
in
from
somewhere
else
and
get
parsed
correctly,
and
so
once
again,
this
is
all
kind
of
fold
over
stuff
and
the
reason
why
it's
all
like
this
subclass
based
little
bit
of
a
mess
here
is
because
then
you
could
just
say:
okay,
I
need
something.
B
That's
uses
source
and
it
uses
a
model,
and
all
you
have
to
do
is
basically
just
subclass
use
two
base
classes
right
and
now
anytime,
you
need
a
a
command
that
does
machine
learning
stuff.
You
can
now
just
derive
from
ml
command,
which
is
just
a
combination
of
something
that
uses
a
model
and
something
that
uses
a
source.
B
I
think
something
that
uses
a
model:
okay,
yeah
it
used
to
have
more.
You
have
a
lot
more
arguments,
but
as
we've
consolidated
this
config
stuff,
it's
all
kind
of,
like
you
know,
become
under
the
the
object.
The
main
configurable
objects
themselves
and
not
the
cli
commands.
B
Does
that
make
any
sense
so
far,
okay,
cool
and
so,
and
if
you
want
to
see
what
it
I
mean,
this
is
what
they
actually
look
like
right.
They
basically
just
say:
okay,
I'm
a
train
command
right.
This
is
the
train
command
and
then
where's
the
run
method.
It's
kind
of
it's.
This
is
the
same
way
that
the
operations
work
is
they
implement
a
run
method
and
that
op
decorator
actually
just
creates
a
class
with
a
run
method.
That
is
the
function
that
we're
decorating.
B
If
that
makes
sense,
so
it
basically
just
inserts
the
the
method,
the
function
that
you're
decorating
into
the
run
method
of
the
operation,
and
so
then,
within
the
run
method,
you
just
do
whatever
you
need
to
do
right.
B
So
in
this
case
we
basically
say:
okay
enter
the
sources,
then,
to
the
main
context
of
the
source
enter
the
main
context
of
the
model
check
out
the
source
context,
check
out
the
model
context
and
then
run
the
train
right
accuracy
very
similar
predict
gets
a
little
weird,
but
basically
the
way
that
you
tell
how
what
what
command
gets
run
is
by.
So,
if
you
look
in
cli.cli,
this
is
the
main
class
that
does.
B
This
is
the
main
entry
point
for
this
command
line
interface
and
if
you
define
like
you've
seen
over
here,
if
we
define
like
an
arg
like
arg
underscore
log,
and
then
you
say:
okay,
what
is
this
argument?
That'll
end
up
in
the
self.log
property
of
the
class,
so
that
was
another
thing
to
make
it
easier
so
that
you
basically
you
say
what
are
my
arguments
and
then
within
the
run
method
you
just
can
access
them
with
self.whatever
and
it
simplified
writing
these
commands
very
greatly
right.
B
So
then,
if
you
want
something,
that's
not
an
argument,
but
is
like
a
sub
command.
You
just
say:
what
is
the
property
equals,
something
that
derives
from
command
right,
and
so
then,
if
I
say
so,
if
I
say
dffml,
it's
going
to
just
print
out.
These
are
all
the
sub
commands
right,
which
is
what
we're
seeing
over
here
and
then.
B
So,
if
I
say
like
dfwml
train
help
it'll
it
knows
about
the
model
and
it
knows
about
the
source
which
is
above
here
because
they're
both
you
know,
it's
a
double.
F
B
Yeah
polymorphic
or
whatever
that
is
right,
and
so
the
issue
with
this
becomes
that
all
these
very
nested
config
options
aren't
really
captured.
We
could
capture
them,
but
the
problem
becomes
like
okay,
so
we
have
all
these
different
models.
Then
this
was
originally
one
of
the
things
that
happened.
It's
basically
like
okay.
I
wanted
to
list
out
all
the
different
command
line
options
for
all
the
different
models.
F
B
B
Become
this
massive
massive
thing-
and
I
don't
know
if
you
guys
have
ever
used
like
docker
docker
machine,
but
they
hit
that
problem
and
you
would
dump
their
help
message
and
it's
like.
Oh,
it's
like
400
lines,
long,
it's
insane
because
they
did
the
same
sort
of
thing
where
they
have
these
plugins
and
the
plugins
have
different
options.
It's
a
mess!
So
that's
why
it's
all
on
the
docs
website
so
but
yeah!
This
is
how
you
figure
out
what
cli
command
is
getting
run.
B
So
if
you
do
dfml
train
that
becomes
the
train
command
and
like
if
we're
looking
at
the
data
flow
stuff,
dfml
data
flow,
then
the
data
flow
is
actually
imported
from.
Let's
see,
where's
I'll
open
it
here,
so
dfml
cli
data
flow,
so
the
data
flow
commands
are
all
defined
above
here
right.
Okay,
that's
the
display
one
so
run
repos
is
this:
is
the
run
command
and
then
you
can
do
it
on
the
set
or
you
can
do
it
on
all
of
them?
B
And
if
you,
the
reason
why
these
have
an
underscore
in
front
of
them
is
because
set
and
all
are
keywords
in
python,
right,
they're,
built-in
functions
and.
B
To
define
a
sub-command
with
a
built-in
function
as
a
name,
you
can
just
preface
it
with
an
underscore,
and
so
obviously
this
is
all
the
type
of
thing
that
needs
to
be
documented,
which
is
why
I
started
writing
that
little
code
notes
thing
and
actually
this
might
go
under
the
notes
for
subsystems,
but
this
is
all
actually
stuff
that
needs
to
change
long
term,
because
we
have
that
existing
config
infrastructure
and
those
like
auto
args
and
config
methods
that
can
extrapolate
from
any
class.
That's
decorated
with
that
config
method.
B
So
we
really
should
try
to
reduce
the
amount
of
like
double
parsing
code
of
arguments.
That's
going
on
here,
which
should
be
should
be
pretty
doable
just
need
to
go.
Do
it
right,
but
yeah,
let's
see!
Does
that
answer
your
questions
enough
on
how
this
works.
B
It
was
that's
right,
I
think
we
might
have
lost
him.
F
B
F
Like
now,
we
are
planning
to
merge
both
of
those
right
across
right
yeah.
We
won't
be
using
the
run
battery
flow
method.
B
Yeah,
so,
okay,
so
here's
the
thing
actually
and
you
just
made
me-
realize
something
having
outlined
the
stuff
about
the
command
line.
So
if
we
go
to
the
docs
and
we
go
to
the
usage
examples,
data
flow
deployment
combining
operations.
B
So
this
is
the
example
of
right.
We
took
one
data
flow.
We
tagged
on
this
other
data
flow
here
to
the
side.
We
made
them
one,
we
merged
them
right
and
so
the
way
that
this
worked
is
we
basically
just
said
dfml
dataflow
merge
and
then
we
piped
that
to
the
dataflow
diagram
command
and
we
said:
okay,
your
data
flow
is
coming
from
standard
in
what
I
had
been
trying
to
say
was
that
we
could
implement.
B
B
Yeah,
okay,
so
let's
do
that?
I'm
glad
we
realized
that
so
that's
good,
so,
okay
yeah!
So
I
think
then,
in
that
case,
but
this
is
the
thing
right,
so
we
were
talking
about
and-
and
we
were
talking
about
how
you
did
you
understand-
I
think
it
I
feel,
like
I
remember
seeing
that
I
was
able
to
communicate
successfully
the
thing
about
the
repo
deaf
right.
B
F
B
So
for
everybody
else
that
hasn't
been
deep
in
this
area.
Basically,
let's
see
when
we
run
these
data
flows
with
with
okay,
so
with
a
data
source
right,
we
iterate
over
the
data
source
and
we
say
four
repos
in
source
dot.
You
know
with
features,
and
I
give
all
you
all
the
repos
with
those
features
right
and
that
internally
calls
that
source.repos
method
which
just
basically
dumps
out
everything
that
could
be
in
that
data
source.
B
So
with
that,
what
we
do
with
the
data
flows
is
the
data
flow
basically
takes
inputs
of
different,
like
a
meta
definition,
type
of
things.
So
this
thing
is
a
url,
and
then
you
create
a
basically
a
global
variable
that
says
this.
This
is
a.
F
B
You
know
the
string
representation
so
url
and
a
url
is
of
data
type
string.
That's
like
it's
primitive
right,
and
so
all
of
these
different.
All
of
these
all
of
these
little
blocks
here,
except
for
the
the
the
purple
or,
let's
see
the
more
pink
ones,
are
these
input
pieces
of
data
and
the
more
purple
ones.
E
B
B
Purple
ones,
basically,
they
have
you
say
what
the
inputs
are
and
the
inputs
are
of
specific
data
types
right,
and
so
the
package
is
one
of
these.
B
B
So
when
the
way
that
this
works,
when
we're
combining
it
with
the
sources
is
that
if
we
want
to
run
a
data
flow
on
each
entry
in
a.
E
Source
right
repo,
what
we
do
is
we
basically,
we
iterate
over.
B
B
We
still
need
to
rename
that
it's
I'm
going
to
assign
the
the
key
will
be
the
value
for
some
definition
right,
because
any
input
to
one
of
these
operations
is
the
value
right
which
would
be.
You
know,
the
string
representation
of.
D
B
All
the
things
that
take
an
input
of
definition
package,
let
me
go
run
those
right,
and
so
that
would
be
this
pie,
pie,
package,
json
and
this
safety
check,
and
so
basically
what's
happening
now
is
that
we
have
this
concept
of
subflows,
so
you
could
run.
B
Flow,
but
to
do
that,
you
can't
really
pass
currently.
You
can't
just
pass
the
certain
inputs
matching
definitions
directly
into
that
subflow,
which
would
be
helpful
because
then
you
could
just
say:
okay,
like
you
know,
to
kick
start
this
thing.
The
first
thing
I
do
is
I
run
this
subflow,
which
is
the
current
case
with
what
we're
working
on,
but
so.
B
B
B
Right
what
what
needs
to
happen
is
basically
just
is
as
long
as
you
remove
that
I
guess
what
I'm
trying
to
ask
here
is:
do
you
know
how
to
test
that
right?
You
would
you
could
yeah.
B
B
D
Sure
that
was
kind
of
confusing,
because
this
stuff
is
wacky.
A
B
Because
the
nice,
the
reason
why
we're
doing
all
of
this
data
flow
stuff
is
because
you
you
can
take
these
models
that
we're
using
and
then
directly
combine
the
data
set
generation
with
the
model.
Predictions
and
stuff
like.
D
B
And
you
could
you
can
create
arbitrary,
like
okay,
go,
do
this
right.
E
B
And
then
you're
gonna
actually
use
that
feature
data
to
train
it
through
the
model
right.
So
when
anybody,
if
somebody
just
gives
you
a
package
name
now,
you
can
go
generate
all
the
numbers
that
are
associated
with
that.
That
would
actually
let
you
go
take
that
package
and
run
it
through
a
trained
model.
So
that's
the
point
of
this
and
yeah.
Okay.
So
did
what?
What
else
do
we
need
to
talk
about?
On
that
front?
There.
B
B
Sort
of
like
a
dive
into
the
the
data.
B
They
usually
use
like
sort
of
the
standard.
You
know
they
use
like
mapreduce
type
of
ways
of
doing
things
right
and
what's.
B
Is
more
of
it's
it's
an
event
loop
right?
It's
not!
It's
not
like
a
defined
sequence
of
things
like
this
is
a
event.
B
B
B
It
is
pretty,
I
feel
like
it's
pretty
commented,
though:
okay,
okay,
all
right.
Okay,
that's
a
lie.
Yeah,
there's
some
comments.
There
needs
to
be
more.
There
needs
to
be
really
like
a
whole
write-up
on
this,
because,
what's.
B
To
write
that
up,
but
you're
definitely
the
one
who's
approaching
understanding
that
very
quickly
and
the
other
person
who
really
understood
that
was
arvin,
but
he
hasn't
been
around
for
a
while.
I
think
he
had
to.
He
had
to
go
finish
up
school
and
get
a
job,
but
he
had
done
a
lot
of
stuff
on
the
data
flow
side
of
things.
B
D
C
B
B
D
B
A
it
would
be
subscribed
to
some
kind
of
you
know,
distributed
queue
where
it's
waiting
for
its
input
arguments
and
then
it's
publishing.
D
F
B
B
E
B
B
Distributed
setting
as
well
as,
let's
see
as
well
as
the
other
thing
that
this
lets.
B
B
B
Network
and
that's
like
a
network
of
operation,
implementations.
B
What
are
the
outputs,
so
what
that
allows
us
to
do
is
there's
an
open.
B
Are
accessing
all
the.
B
Right
so.
B
B
Definition
act
as
a
network
of
implementations,
so
you
could
just
like
point
it
at
this
web
server.
That
knows
how
to
do
you
know
xyz
things.
You
mapped
that
onto
the
operation,
equivalence
and
now
you
you
have
access
to
all
of
those
things
without
having
to
you
know,
rewrite
the
python
api
code
to
call
those
right
so
to
let
us
integrate
with
other
things
as
well.
B
The
longer
term
goals
for
the
data
flow
side
of
things.
I
don't
think
I
told
you
guys
that
yet
I
I
know
yash
had
been
on
a
while
ago
when
I
talked
to
him
and
I
let
him
and
sudarson
know
all
about
that
stuff.
But
I
don't
and
I
know
arvin
knew,
but
I
don't
think
I
had
told
you
guys
yet
and
I
haven't
gotten
the
chance
to
write
it
all
down
so
anyways.
So
let's
talk
about
himanshu.
Did
you
want
to
talk
about?
B
So,
let's
see,
how
did
that
you
were
working,
you
were
working
on
the
repo
dot,
merge
right.
A
A
What
I
think
that
so
what
I
think
for
now,
given
my
current
level
of
understanding
for
the
code,
let's
open
the
community
input
for
it,
okay,.
A
B
B
Yeah,
let's
let's
jump
on
that
and
then,
if
you
could
post
your
you
know
you
had
that
idea.
If
you.
B
They'll
have
all
the
ideas,
but
yeah,
let's
see
we
could
just
look
at
that
service.
Http
code,
real
quick,
so
the
thing
is-
and
this
is
why
this
is
kind
of
like
what
I
was
talking
about
with
the
cli
commands
is.
This
is
why
it's
really
important
that
at
some.
B
B
Config
parsing
code
that
that
handles
those
config
objects
for
that
base.
Configurable
class
in
base.ey
because,
like.
B
B
Some
config
class
that's
been
decorated
with
that
config
we've
figured.
B
Options
into
there
and
then
the
other
thing
is,
it
should
be
able
to
successfully
load
things
from
files.
At
this
point,
I
don't
know
again.
B
B
Option
or
the
config
flask
class
for
this
plug-in
type.
That
is,
you
know
the
database.
So
we've
got
we're
to
the
point
where
we
can
like
serialize
and
deserialize.
B
B
Is
right
where
you
do
that
you
define
you,
create
a
list
like
on
the
fly,
with
like
a
little
for
loop
or
whatever,
so
for
some
reason
it
was
leaving
a
certain
property
as
a
generator,
but
it
wasn't
just
leaving
the
property
as
a
generator.
It
had
like
set
the
property
to
a
list
with
the
first
element,
as
the
actual
generator
for
that
list,
which
was
really
weird.
So
we
have
this
like
a
little
hack
around
that
now,
but
but
yeah.
So
we
need
to.
We
should.
B
We
can
kind
of
serialize
and
deserialize
any
of
the
classes,
and
so,
if
we
can
figure
out
at
some
point
how
to
unify
that
that
config
code,
that
would
make
things
like
you
know
it
would
make
things
like
this
easier
right,
because
the
thing
is
right
with
the
cli
commands.
They
don't
use
that
config
that
auto
config
stuff
right.
C
B
B
Simplify
things
and
I'm
not.
B
Timeline,
we
do
that,
but
it's
sort
of
just
something
where
somebody.
B
On
like
okay,
how
does
this
work
and
you
know
it's
it's-
it's
gonna,
be
a.
B
B
We
won't
be
hitting
you
know,
weird
it'll,
be
it's
easier
to
change
things
and
add
things
when
it's
all
just
a
property
of
some
class
right.
It
becomes
way
easier
to
modify
stuff.
B
Without
doing
massive
refactors
but
yeah
yeah
cool,
I
guess
and
then
anything
else.
A
The
other
thing
I'm
thinking
of
like
I'm,
not
sure
what
should
I
do
first,
so
one
thing
is,
we
can
add
more
models
in
scikit-learn,
okay,.
D
A
There
are
a
few
things
like
we
are
not
opening
giving
all
the
we
are
not
opening
up
our
api
fully.
I
am
thinking
because
yeah,
what
we
have
is
train,
predict
and
fit
right.
A
Functions
yeah
we
need,
especially
if
you
go
for
unsupervised
thing,
okay,
and
so
maybe
I
was
thinking
we
can
open
it
up
so
like
we
can
access
all
the
functions,
if
not
all,
for
now
at
least
few
more
okay-
and
that
is
one
point
then
the
other
point
was
for
accuracy
thing
so.
A
A
That
second
point
and
I'm
not
sure
what
should
I
do
first
and
the
third
point:
is
I've
been
working
on
side
by
side
on
the
tensorflow
pre-trained
models?
Okay,
so
oh.
A
So
for
now
I
cannot
do
for
images,
because
obviously
suction
is
doing
that
image
thing.
Okay,
so
for
text!
I
can
do
that,
okay.
So
what.
B
Should
I
do
first,
you
mean
like
natural
language,
processing
and
stuff.
A
Yeah,
because
if
you
see
on
the
tensorflow
hub,
then
you
we
have
three
options
for
pretend
models
for
now.
D
B
E
A
B
Really
cool,
I
think,
yeah.
I
think
that
would
be
super
helpful,
and
so
I
think
one
of
the
things
is
with
this
with
the
pre-trained.
B
A
D
A
B
This
sounds
yeah.
This
sounds
awesome.
I
guess
I
don't
really
know
much
of
what
direction
to
give
you
here.
I
think
you
kind
of
probably
have
a
good
idea.
It's
you
know
it's
same
type
of
stuff
that
was
in
scikit,
especially
since
you
actually
had
to
go
implement
the
other.
D
B
Context,
you're,
probably
pretty
familiar
with
you
know
how
this
is
gonna
go
now,
but
basically
you
know
you
do
the
same
sort
of
format
right
and
if
you
want
something,
that's
a
little
more
slimmed
down
to
use
it
as
a
reference.
B
If
you
are
in
model
tensorflow
itself,
you'll
see
you.
B
Functions
and
stuff,
the
one
thing
is
that
the
one
thing
is
that,
obviously,
that
is
using
the
one
point.
B
Api
and
if
you
wanted
to
move
to
the
2.0
api
you'd,
have
to
change
the
rest
of
tensorflow
code
right
and
I
know.
D
B
B
It
was
like
over
the
winter
here
or
like
slightly
before
that,
but
that
would
be
no.
B
Yeah
sweet
so
yeah,
if
you,
if
you're,
I
think
and
I
think
the
ap
apis
are.
You
know,
hopefully,
when.
B
Like
it
was
a
while
ago,
when
I
first
implemented
this
code,
I
think
it
was
like
almost
three
years
ago
now
so
and
as
you'll
notice,
like
all
the
apis,
are
it's
kind
of
weird
the
way.
B
Yeah,
if
all
of
that
stuff
is
very
2.0
focus,
then
yeah
definitely
try
to
figure
out
how
to
switch
the
existing
code
to
2.0
as
the
first
thing
and
then
once
you
switch
the
existing.
D
B
The
functions
for
training
input
data
is
derived
for
they're
the
same
so
hopefully
that
that
goes
smoothly.
If
not,
you
know
me
and
definitely
give
some
pointers
on
that,
but
yeah
so
yeah.
I
think
that's.
B
And
I
think
the
other
thing
is
that
you
know
if
you're
gonna
do
those
applicable
features,
methods
those
right
now,
you're
just
gonna,
need
to
make
sure
to
override
that
there
was
it
with
the
existing
api.
We
could
have
easily
added
the
string
data
type
in
there
and
done
one
of
those
like
what
was
that,
like
hash
bins
or
something
for
the
column
to
basically
add
string
all
they
were
doing.
D
B
Your
text,
presence
and
stuff
we'll
be
doing
because
I'm
not
super
familiar.
F
B
Will
work
on
mlp
for
for
tensorflow
and
then
yeah.
It
tends
to
look
pre-trained
models
and
then
you
were
also
talking
about
giving
the
user
the
option
to
select
what
accuracy
means
with
psychic
models:
yeah,
okay-
and
so
let's
open
an
issue
for
this.
So
that's
that's!
You
know.
We
want
to
track
that.
B
All
right
so
model
scikit,
oh
and
I
was
able
to
successfully
demo
the
whole
project
to
somebody
over
the
weekend
is
the
first
time
I
actually
got
to
give
a
live
demo.
I've
given
a
presentation
before,
but
but
this
is
the
first
time
somebody
said
hey.
Can
you
use
your
thing
for
that
and
I
was
able
to
you
know,
pull
it
down
and
be
like
pip
install
everything
worked,
and
then
we
ran
it
on
their
data,
and
you
know
it
actually
got
90
accuracy.
So
we're
like
great.
B
You
know
we
were
using.
I
think
one
of
the
scikit
one
of
the
sidekick
models,
so
that
was
really
awesome
right.
It
was
like,
oh,
you
know
success.
We
can
pull
down
this
thing
and
people
can
use
it
for
machine
learning.
That
really
don't
understand
much
other
than
knowing
how
to
use
the
command
line,
and
you
know
export
their
cs
excel
files
to
csv.
So
that
was
a
big
win
right
there.
It
was
exciting
to
see
yeah
and
I
was
really
I
was
also
you
know.
B
You
know
what
really
made
my
weekend
was
when
you
said
ogen
when
you
said
that
it
was
exciting
to
see
all
the
stuff
come
together
in
the
data
flow
that
was
like.
That
was
really
fun
to
hear,
because
that
is
like
exactly
what
we're
going
for.
You
know
it's
like
you,
throw.
D
F
B
Yeah
it
can
be,
it
can
be
a
little
nerve-wracking,
but
it's
really
cool
when
it
runs,
and
it's
just
like.
Okay,
I
did
all
of
this
stuff
at
the
same
time,
for
you
enjoy
like
it's
great,
let's
see
so
give
the
user
the
option
to
select
what
they
want.
I'm
going
to
add
an
issue
for
this
and
then
all
right,
great
yeah,
and
we
got
those
clustering
models
in.
I
want
to
go
use
those
myself.
I
want
to
see
how
that
works.
That's
going
to
be
fun.
Let's
see!
B
B
Okay
and
what
happens?
Oh
okay,
here's
a
test.
What
happens
if
we
take
that?
I
wonder
if
this
is
documented
in
the
docs
stuff,
so
check
out
id
actually
says
docs
usage
did
I
document?
Did
we
document
the
list
command
here?
It
doesn't
look
like
we
did
yeah.
B
Let's
try
that
list
command
and
then
you
can,
you
know,
do
the
head
and
whatever
to
cut
it
off,
see
here
it's
doing
the
same
thing.
It
looked
like
that
was
a
list
yeah
that
one
list,
repos
yeah
grab
that
and
let's
see
and
you're
gonna
need
to
obviously
cut
off
those
little
carrots
at
the
front.
There.
B
Here
do
that
control
x,
e
yeah
and
then
you
can
get
rid
of
those
guys
easier.
B
That
reminds
me
I
need
to
set
the
editor
to
vim
very
well.
I
guess
not
everybody
likes
vm.
I
shouldn't
set
the
editor
to
bam.
F
So
I
was
just
trying
to
run
the
code.
I
got
an
entry
point
not
found
you
made
some
changes
right.
Did
you
push
that.
B
Okay,
let's
see
we're.
B
Okay,
it's
not
merging
wait!
So,
okay
yeah,
it's
still
not
merging
what
the
hell
is
up
with
that.
So
are
again
sorry.
What
branch
are
you
talking
about.
F
B
Sure
can
you
try
pulling
the
branch
again
because
I
think
or
well,
oh,
if
you're
getting
entry
point
not
found,
you
have
to
install
the
source.
Sorry!
So
that's
that
is.
A
B
B
Yeah
you
can
do
that
yeah
that
one
will
work
or
if
you
just
want
to
install
it
one
off.
You
can
do
the
pip
install
dash
e
source
idx
and
then
obviously
with
the
prefix
or
not
it
goes
yeah.
I
started
just
saying
everybody
should
install
to.local,
because
the
user
flag
sometimes
doesn't
work.
I
noticed
and
then
and
then
also
yeah.
Installing
globally
is
a
mess.
The
user
flag
sometimes
doesn't
work
so
that
prefix
seems
to
work
more
than
not
so
I
just
put
that
everywhere
in
the
docs.
B
E
B
I'm
going
to
run
it
in
my
github
too,
and
then,
if
anybody
else,
let's
see
yeah
it's
here
I'll
share
mine
too.
So
you
can
see
what
I'm
doing.
B
B
D
B
It
in
docker
I
was
like
oh,
this
will
have
to
isolate
it,
but
I
guess
not
okay,
but
this
is
really
weird
right,
because.
A
B
B
Give
me
a
show
all
right,
so
vam
dffml
source,
you
guys,
can
see
my
screen
right.
Yes,
okay,
cool!
B
All
right,
I
put
t-mux.
B
A
B
B
F
B
Okay
expected
most
one
argument:
scott
two:
why
why
would
you
expect
most
one
arguments,
oh
giving
you?
Okay,
whatever.
B
Maybe
just
because
of
that
just
just
tell
us
what's
going
on,
okay
and
and
everybody,
if
you
guys
don't
want
to,
if
you
guys
have
other
things
to
do
or
sleep
to
get,
don't
feel
the
need
to
stay
on.
The
call
oh
and
I.
C
B
Yeah
that
yeah
that
could
be,
that
could
be,
let's
see
so,
let's
try
this
here,
because
it's
we're
sane,
okay,.
A
B
B
We
could
try
printing
it
out
to
see
what
the
hell
is
going
on
there
like.
If
it's
parsing
that
stuff
correctly,
I
mean
I
kind
of
assume
it
is.
I
think
it's
a
problem
with
this
method
here,
because
I
bet
you
if
you
listed
the
one
source
and
then
you
listed
the
other
source,
it
would
show
up
correctly
if
you
listed
them
one
by
one,
but
I
guess
we
should
try
that
okay,
oh,
this
is
gonna,
probably
create
a
problem
for
my
machine:
oh
jesus,
oh
god,
okay,
yeah!
B
Okay,
geez
all
right.
Well,
I'm
gonna
have
60
000.
Okay.
There
we
go
all
right,
so
this
self
one
is
zero,
all
of
a
sudden.
So
what
happens?
If
we
do?
Okay-
let's
do
this
here
and
then
let's
do
I
swear.
I
tested
this.
It's
probably
self-zero
is
what
it's
supposed
to
be,
but
I
swear
I
tested
it
and
it
was
supposed
to
be
self
one.
C
But
it
should
be
one,
it
starts
from
the
the
number.
B
B
B
E
B
17.6
yeah
and
I
have
3.7.3
either
I
was
relying
on
a
bug
or
there
is
a
bug
now,
because
this
is
obviously
working
on
my
machine
and
not
working
on
3.7.6
so,
and
this
is
also
but
the
other
thing
is
that
this
async
context
manager
lists
context,
god
that
is
a
long
class
name.
I
apologize
it's
what
it
is
async
helper,
so.
B
B
Why
can't
we
access
it
like
that?
What
the
hell
for
item
himself.parent
does
the
user
list
have
a
property?
I
think
user
list
has
some
property.
B
Okay
check
this
out.
You
see
this
here
and
then
this
user
list.init,
okay,
self.pen,
oh
no,
okay,
so
that
should
be
fine
too
damn
it
or
maybe
we
should
be
doing
like
self.data.append.
B
B
Still
not
correct
what
is
up
with
that,
because
what
happens
if
we
just
do
like.
B
D
A
And
before
that,
let's.
B
B
B
B
Will
work
for
everything,
because
this
is
this
is
basically
right.
This
is
where
the
generic
abstraction
stuff
comes
in,
because
this
is
this
is
the
part
right
here.
This
sources
class
is
what's
abstracting
all
of
the
the
background
data
sources,
whether
it
be
csv
or
json
or
whatever.
So
all
sources
will
now
be
fixed
right.
If
you,
if
you
combined
like
a
mysql
database
and
a
csv
source,
then
it'll,
it
should
all
be
fixed
at
this
point,
unless
I
interpreted
your
question
wrong.
B
B
Maybe
this
is
where
they
screwed
it
up,
because
this
I
feel
like
this
is.
This
is
around
the
time
that
I
got.
I
pulled
that
new
version
of
python
anyways.
It's
not
super
important,
we'll
see
if
it's
still
in
correct
return,
type
for
slicing
operation.
B
Yeah,
so
it's
updated
that
I
I
don't
know
we
should
see
if
it's
in
3.8.
I
guess.
Let's
just
make
a
note,
and
then
this
is
the
new
issue
for
the
models.
A
Now
that
the
this
one
is
fixed,
the
you
you
talked,
you
wanted
to
talk
about.
The
ideas
like
on
the
mnist.rst5u.
C
That's
what
is
next,
what
student
right.
B
Okay,
let's
see
yeah
so
what's
next,
here
is
okay
and
let
let's
just
let's
just
see
what
just
happened
here.
Let's
before
we
call
it
declare
success:
let's
do
git
diff
and
actually
apply
this
so
because
I'm
working
in
the
okay,
let
me
go.
Let
me
go
copy
paste
this
into.
B
Okay-
okay,
great
so.
B
Fix
for
context
manager
list
slice,
okay,
now,
hopefully
that
works
I
mean
if
it
works
in
getpod.
I
would
assume
it
works
everywhere
else.
So
docs
usage,
amnest
all
right
so.
B
And
doing
pre-processing
to
scale
you
know
to
to
do
normalization
right
yeah,
you
did
the
normalization,
but
then
you
removed
it
yeah
I
removed
it.
Actually
that
may
have
been
when
the
accuracy
was
good.
I
may
be
confused
so
because
when
I
got
it
without
the
normalization,
I
remember
I
was
getting.
B
B
To
that,
but
that
may
not
be
true
right,
so
the
thing
is
what
we
really
need.
Long
term
is:
we
need
to
create
a
source
that
is
like
the
it
will
be
under
the
entry
point
of
pre-process
or
something,
and
then
we
would
say
like
that
that
source
itself
would
take
a
source
and
then,
when
it's.
B
On
the
feature,
data-
and
actually
this.
E
B
A
that's
an
open
question
right
now,
but
so
yeah.
What
we'd
want
to
do
is
we'd
want
to
create
this
source
where,
as
it's
iterating
through
all
the
entries
in
it,
it
does
some
pre-processing
right
and
to
start
with,
the
initial
idea
would
be
just
do
like
you
could
specify.
Okay,
for
this
feature
that
you're
getting
out
of
this
source.
I
want.
B
Apply
this
pre-processing,
which
would
be
something
like
run
this
specific
operation
right
and
then
you
would
say,
use
the
use
the
featured
data,
as
you
know,
whatever
the
input
to
so,
for
example,
the
array
divide
operation
might
take
two
inputs
right.
It
might
be
like
you
know
the
array
and
then
the.
B
C
D
D
B
A
B
Actually,
I
think
I
figured
out
what
the
syntax
for
this
would
be,
so
it
would
be
like
source
feature,
so
then
images
and
then
something
like
this
is
where.
B
User
dict
mapping
so
or
sorry.
B
Think
I
put
that
in
a
comment
somewhere,
but
basically
what
we
need
to
do
is
add
to
config
person
code
away
to
do
dict,
stir
some
class
mapping
and
what
that'll
let
us
do
is
basically.
B
And
then
we
do
like
so
image
operation
will
be
array
norm
normalize
and
then
you
do
like
feature
feature
to
input
and
then
this
would
be
like
the
input
name,
input,
name
array
or
feature
data.
B
B
A
B
D
B
B
And
then
there's
probably
like
there's
probably
another
one
of
these
dict
mappings,
where
we're
starting
mapping
feature
input
names
so,
like
future
inputs.
B
E
B
Okay,
great,
let's
do
that
so
it'll
probably
be
tomorrow
for
me,
but
so
let's
see
so
let
me
push
this
up
docks.
Okay,
all
right!
I
will
see
you
guys
have
a
great
night
and
thanks
for
working
this
all
out.
This
is
great
and
let
me
know
if
there's
anything
we
missed
on
twitter,
then,
okay,
all
right
thanks.