►
From YouTube: ONNX Training working group meeting on September 3, 2019
Description
Webex meeting of ONNX Training working group on September 3, 2019.
A
A
B
B
A
B
Back
right,
no
I
am
basing
a
study
that
I
need
someone
to
tell
me
what
they
need,
for
example,
if
the
proposal
can
can
can
really
capture
the
snare.
Oh,
they
are
doing.
Of
course,
I've
been
a
list.
That
proposal
can
already
capture
all
the
targets
in
a
row
in
my
mind,
and
it
can
also
support
high
order
differentiation,
not
sure
if
you
guys
have
familiar
with
it
so
I
mean
you
guys
need
to
tell
me
what
you
want
from
me,
because
that
propose
is
already
stable.
A
B
A
A
B
A
A
B
D
B
E
E
D
F
A
E
E
So,
for
instance,
you
know
you
have
this
simple
model
to
start
with.
Maybe
later
I'm
like
you
know,
we
talked
earlier.
We
want
to
have
more
difficult
or
you
know
complex
one.
So
maybe
we
can
have
two
example
files.
So
for
the
back
end
they
can
always
take
these
two
and
try
to
make
it
to
Train
right
in
different
platforms.
E
I
guess
so.
So
that's
the
one
part
I
just
wonder
how
we
can
make
this
whole
of
this
or
experimental,
easier
I'm.
So
wishing
I'm
going
to
ask
you
to
help
because
you
already
open
the
pupae
ours,
maybe
you
can
merge
them
into
one,
and
maybe
you
can
even
put
some
pre-created
the
onyx
files
or
just
include
the
scripts,
try
to
create
a
new
onyx
files,
so
that
is
part
I
think
we
should
be
working
on
right.
E
Okay,
yeah
I'm,
sorry
I'll
call
it
makes
sense
right,
so
the
second
part
step
thrown
in
the
front
end
I,
believe,
is
going
to
be
coming
from
some.
You
know
of
frameworks
I
like
pie,
torch.
We
have
this
model.
I
want
to
see
again
some
way
to
generate
the
target
onyx
file
from
them
exercise
arm.
Maybe
we
can
ask
spendin
someone
to
kind
of
come
up
with
a
proposal,
a
generic
way
once
he's
done
that
once
we
can
ask,
maybe
the
second
front
should
try
to
go
through
that
right.
E
G
Should
separate
from
there
so
I
think
wait
back
I
think
it
would
be
a
good
idea
if
we
write
down
a
proposal
and
then
share
it
and
agree
with
the
rest
of
the
community.
What
we
don't
want
to
do
is
do
some
testing
think
we're
happy
with
it.
Try
and
then
propose
it
to
the
rest
of
community,
and
they
say:
oh
actually,
we
wanted
x,
y&z
also
done
so
it's
probably
worth
agreeing
what
the
steps
are
to
validate
that
the
training
proposal
is
good
enough
with
everyone.
Before
we
even
start
doing
it.
G
G
That,
but
at
the
time
isn't
it
you
know
each
framework
will
take
a
certain
amount
of
time.
The
more
frameworks
which
I
imagine
people
will
want
the
more
effort
that
we
will
be
required
to
prove
that
so
as
long
as,
if
we
can
get
an
agreement
to
then
there's
more
Chancellor's
getting
it
in.
If
we
have
to
do
five,
then
it's
a
lot
of
effort
on
our
part,
but
I
think
we
should
try
and
agree
with
them
what
this
extra
scope
is.
A
F
A
E
E
F
E
E
A
E
A
C
C
G
A
G
E
E
A
A
G
B
G
B
C
B
C
A
C
A
H
So
before
we
commit
to
the
day's
like
III,
prefer
that
it's
it's
it's
it's
the
it's
probably
something
that
we
we
should
skip.
I
see
the
gradient
operator
as
a
placeholder
and
it's
up
to
each
end,
our
back
end
to
decide
how
to
implement
it.
So
the
cotton
is
of
the
growing
up.
It's
it's
on
back-end,
but.
G
H
H
Like
I've
been
tuned
in
or
the
numerical
crack
knows,
I
have
a
proposal.
I
have
a
way
to
test
it.
That
is
we
we
can
numerically
approximate
the
gradient.
You
take
an
F.
Are
you
perturbative
by
a
delta
x,
like
you
have
an
effect,
and
you
have
a
delta
f
x,
plus
delta
x,
and
then
you
take
the
difference
and
divide
by
Delta
X.
That's
a
proclamation
of
the
gradient.
H
A
A
H
H
C
B
Yes,
so
one
point
Shahrukh,
which
I
said
sometimes
or
having
open
times,
Radian
is
at
the
unique
value,
is
a
wrench
and
it
can
vary
from
0
to
1.
That's
that
happen.
That
happens
when
the
function
is
not
differentiable,
for
example,
in
the
real
it
has
artists
at
point
of
the
euro.
It's
not
differentiable
right.
D
C
We're
talking
about
unit
test
cases
right
so
I
mean
raising.
Your
point
is,
of
course
true,
but
it's
slightly
different
like
that.
That
derivative
may
not
be
defined
at
certain
points.
There
may
be
points
of
discontinuity,
etc.
That's
fine,
but
when
we
are
designing
unit
test
cases
you
can
choose
the
points
where
the
gradients
are
being
computed
and
okay,
you.
C
F
H
Consider
in
both
tensorflow
unit
test
and
on
excellent
I'm,
like
the
way
that
we
are
currently
how
we
implement
it.
It
is
the
way
that
we
test
the
correctness
of
their
gradient
ops,
so
the
back
ends
have
to
have
some
way
to
cart
the
correctness
on
the
implementation
of
these
kernels
or
some
gradient
ops.
H
C
H
C
C
A
H
A
G
A
H
H
A
H
D
C
F
I
H
H
C
H
G
F
H
A
G
H
A
E
E
A
G
There
was
another
point
that
someone
raised,
and
it
very
worth
coming
to
an
agreement
on
is
whether
the
operation,
the
operated
definition,
needs
to
have
a
description
of
what
the
gradient
operation
should
do.
Do
we
feel
that
the
operators
need
to
be
updated
with
some
text
regarding
the
function
of
the
gradient
operator.
G
G
G
A
D
D
C
So
so
sorry
did
this.
To
mention
anything
I
mean
I.
Guess
one
of
the
proposal
is
to
see
whether
for
ups,
we
can
have
a
way
of
attaching
it's
derivative
as
a
symbolic
function
right
they
express
the
gradient
of
each
annex
up
in
embolic
form
as
a
function
and
as
raising
was
saying,
I
mean
I.
Think
it's
doable
for
at
least
for
some
of
the
ups,
and
in
the
long
run
it
may
be
useful
for
multiple
purposes,
so
it
could
be
used
to
help
the
testing
we
were
just
discussing.
C
For
example,
the
gradient
computation
is
being
done
correctly
and
so
on.
It
could
even
be
used
to
automatically
produce
the
gradient.
Maybe,
but
so
it
can
be
done,
seems
useful,
but
aspiration
was
asked.
Maybe
I
mean
I'm,
not
clear.
Well,
it
has
to
be
coupled
completely
with
the
other
changes.
I
mean
it
could
be
something
that
is
done
separately,
written.
D
Yeah
I
agree
and
I.
Yes,
my
question
is
for
a
lot
of
the
ops.
The
definition
of
the
backwards
up
is
implicit
in
the
definition
of
the
forwards
up
and
for
those
ups
will
be.
Are
we
going
to
eventually
explicitly
state
what
the
backwards
op
is?
Even
though
it's
definition
is
implicit
in
the
definition
of
the
forward
up.
C
D
B
C
Yeah
yeah,
but
I
guess
what
I
meant
was.
It
is
not
captured
in
a
machine
usable
form
today
in
onyx,
the
the
semantics
of
a
convolution
operator
is
not
captured
in
a
machine,
readable
format
today
and
so
cannot
be
used
as
the
basis
for
automatically
reducing
a
gradient
today.
Right
so
I
was
answer
this
question
of
whether
there
is
some
mechanism
for
automatically
completing
the
gradient
from
some
formal
specification
of
the
op
zoom
antics
and.
C
C
B
D
A
A
Okay,
let's
move
on
so
so
Sherlock
will
write
some
document
and
if
I
I
just
have
strong
opinions,
please
also
maybe
write
some.
If
you
could
information
that
you
have
into
the
Gator
captains
and
we
could
discuss
orsa
okay
now.
My
next
question
was
okay,
so
here
we
were
talking
about
unit
tests,
but
do
I
have
any
criteria
for
the
like
overall
test
of
this.
A
E
E
A
A
A
F
B
E
E
We
do
interaction
with
thing
on
expect:
I'm,
just
sayin
intensive
flow.
We
don't
do
this
sort
of
interaction
right,
there's,
no
additional
wrapping
for
them
to
graph
wrapper.
Rather
have
this
proxy
thing
from
the
main
graph.
You
can
have
reference
to
a
function.
Okay,
so
I'm
just
saying
that
kind
of
notation
is
not
applicable
intensity.
E
E
E
A
H
H
H
E
Agree,
it's
good
to
have
others
to
you,
know,
validate
and
but
I
also
think
we
should
at
least
put
clearly
what
we
are
trying
to
get
a
review
or
more
official
way
Thanks.
So
if
you
can
have
a
BA
video
document
somewhere,
including
all
the
details
we
are
intending
to
do,
including
how
we
uni
tasks
right,
I
think
that's
better
present.
We
go
through
the
first
round
which
I
wrong
idea,
the
wings,
so
it
is
on
a
later
on.
That's
my
opinion.
E
H
E
Also
so
can
we
risk
the
document
exactly
what
is
needed
and
how
we
are
going
to
approach
if
we
want
this
to
be
approved
right,
what
is
included
in
the
whole
package?
Do
we
have
clear
definition
on
that
to
present
I
don't
want
to
be
leading
a
lot
of
your
roomful.
There
are
ice
spirit
actuation,
so,
let's
be
very
specific
because
it's
a
PR.
It
isn't
that
who
will
be
part
about
the
future.
Thank
you.
We
can
know
exactly
what's
coming
with
that
package
right.