►
From YouTube: SimPEG Meeting July 7th
Description
Weekly SimPEG Meeting from July 7th, 2021
A
A
Let's
see
so,
I
can
just
start
here:
I've
I've
been
basically
working
through
editing,
some
of
the
older
pr's
that
have
been
kind
of
that
have
been
sitting
there
for
a
while.
So
like
so
long,
I've
been
looking
through
your
cross
gradient
one
again.
What
I,
what
I'm
doing
basically
is.
I
made
a
branch
of
it
of
your
brand
and
then
I'll
suggest
changes
to
your
branch,
and
then
you
can
see
what
changed
then
we
can
merge
them
together,
hopefully,
and
get
it
all
in
properly
sounds
good.
Thank
you,
yeah!
A
Sorry,
it's
taking
a
while,
it's
just
a
bunch
of
stuff
going
on
and
as
well
as
some
of
the
other
stuff
like
I've
been
catching
up
on
the
mt
2d
stuff
pulling
it
in
and
as
we
talked
about
last
week,
we
I
think
I
made
this.
We
made
a
decision
to
just
keep
that
site
or
that
our
diesel
wrappers
separate
john.
Did
you
get
a
chance
to
do
anything
else
with
it?
Besides,
compile
it
and
run
it?
Oh
yeah.
B
I
ran
a
few
simple,
like
forward
models
with
the
tiling
and
yeah.
It
seems
to
seems
to
be
good.
I
did
find
like
all
the
ones
or
I
dug
up
a
few
old
projects
to
just
see
what
some
of
those
error
messages
would
be
and
yeah
they're
all
related
to
too
big
or
too
much
memory.
They
couldn't
do
it.
It's
just
it's
nice
to
see
that
that's
what
it
was
now.
So
I
had
a
suspicion
but
yeah,
it's
just
nice
to
have
the
outputs.
So
that's
cool
yeah,
that's
good!
B
I
was
wondering
I
might
add
some
other
changes
that
I
did
previously
when
I
was
playing
around.
I
sent
you
that
one
script
where
I
was
using
the
multi-processing
toolbox
in
there
I
had
some
options
or
I
was
also
playing
around
with
parties.
So
writing
the
fact
factorizations
to
the
disc
and
then
importing
them
back
in.
I
had
a
few
options
that
I
put
into
highmat
solver
that
I
probably
bring
it
in
easier
into
piety
so
I'll
I'll
dig
that
up.
A
A
A
D
D
So
the
first
one
is
adding
publications
to
the
website,
and
so,
as
I've
been
kind
of
thinking
through
this,
it's
raised.
A
couple
of
questions.
First
thing
is:
is
bibtex,
okay
with
folks
here
for,
like
that,
being
what
you
make
the
pull
request
on
for
for
getting
a
publication
posted
on
the
website?
D
A
D
D
D
I
messed
around
with
that
for
my
personal
site
and
got
something
I
was
moderately
happy
with.
The
thing
is,
though,
the
template
that
I
use
for
my
site
relies
on
jinja,
which
I
don't
think
works
with
github
pages,
because
it's
dynamic,
but
we
need
a
static
site
with
github
pages,
so
we
need
to
build
something
intermediate
or
like
basically
build
build.
We
need
to
build
the
whole
page
ourselves.
D
D
D
So
it
basically
lets
you
like
template,
a
page
which
is
actually
kind
of
nice.
Is
that,
like
you,
can
do
for
loops,
and
things
like
that,
so,
instead
of
actually
writing
out
the
html
for
every
single
one
of
my
publications
like,
I
can
just
do
a
for
loop
on
a
python
object
that
has
them
loaded
in
a
list
which
is
nice.
But
that
means
you
actually
have
to
run
python
code
to
like
when,
when
the
page
is
hit,
it
actually
runs
some
code
which
you
can't
do
with
github
pages.
D
D
D
D
Cool,
so
we'll
do
that,
and
so
what
I'll
include
in
this
newsletter-
and
this
is
where
it'd
be
great-
to
get
ideas
on
how
we
want
to
actually
pull
these
together
each
month
is
right.
Now
I
have
three
sections
in
mind
and
if
there's
other
things
you
guys
have
in
mind,
that
would
be
great
to
hear,
but
the
three
that
I
have
in
mind
are:
first
just
starting
with
some
quick,
sim
peg
updates,
and
I
would
imagine
these
being
like
three
ish,
maybe
well.
D
And
then
the
last
piece
being
community
updates,
and
so,
if
there's
anything
like,
if
anybody's
won
an
award
or
if
you've
published
something
that
you
want
to
share
or
done
something
cool
that
you
just
want
people
to
know
about,
we
can
we
can
include
those
there.
So
that's
kind
of
what
I
was
thinking
for
a
structure.
I
don't
know
what
what
do
people
think
of
that?
Is
there
anything
else
you'd
like
to
see
in
there.
D
We
can
figure
that
out.
I
imagine
it
might
end
up
changing
depending
on
who
who
we
invite,
because
there
are
folks
different
time
zones
that
you
know
morning
may
or
may
not
work,
but
we
can
maybe
try
and
stick
with
thursdays
and
see
if
thoughts
if
that
works,
but
yeah.
If
you
would
like
to
give
one
of
the
next
ones,
that'd
be
awesome,
xiaolong
and
anybody
else.
We
can
maybe
come
up
with
a
bit
of
a
list
and
just
sort
of
have
that
open
and
invite.
D
Then
the
the
last
item
with
the
newsletters
is,
if
you
have
ideas
of
who
we
should
send
them
to
and
who
we
should
invite
to
be
on.
That
list
feel
free
to
ping
me
with
with
those
emails,
and
we
can
either
add
them
right
away,
or
we
can
send
them
an
invite
to
subscribe
to
the
to
subscribe
to
the
list,
which
I
think
might
be
a
better
way
to
go.
So.
D
E
E
A
E
E
I'm
not
actually
even
sure
how
to
how
to
describe,
but
because,
like
you
have
some
expectation,
okay,
we're
storing
certain
number
of
arrays
and
here's
the
x
amount
of
memory
that
we're
going
to
consume,
but
somehow
it
kind
of
when
I'm
just
seriously
running
like
jvec,
jtvac,
deep
red,
that's
kind
of
fine,
but
once
it
gets
into
the
inversion
and
then
after
a
couple
of
iterations,
it
kind
of
adds
quite
a
few
memory
which
it's
not
skyrocketing.
So
it
actually
at
one
point
but
yeah.
It's
it's
much
more
than
expected.
E
A
E
Garbage
collections,
and
so
we
thought
sort
of
figured
out,
but
that's
keep
happening,
and
there
are
many
things
happening.
It
could
be
related
to
solvers
or
could
be
just
related
to.
I
don't
know
stack
of
some
sort
of
memory
leakage,
so
I
was
wondering,
like
a:
is
there
a
good
way
to
profile
that
memory
and
just
kind
of
a
good
way
to
get
kind
of
you
know,
search
for
what
are
the
problem
yeah?
I
was
just
kind
of
curious
what
other
people
are
doing
and
there's
a
good
way.
A
A
B
Yeah,
I
I
can
send
you,
I
mean
I'm
using
something
similar
to
what
joe's
explaining
I
just
can't
find
what
the
name
of
it
is,
but
I
can
ping
you
on
slack
there
soggy,
if
you
want
to
know
which
one
or
which
one
I've
been
using
but
yeah.
It
goes
by
line
by
line
and
tells
you
everything
it's
really
cool.
E
B
Of
curious,
like
whether
you
have
that
kind
of
in
this
same
boat
right
now,
I'm
trying
to
do
that
with
desk
and
doing
a
whole
bunch
of
jt,
vex
and
jvex,
and
I'm
getting
an
error
message
saying
that
exceeding
memory.
However,
there's
not
any
memory
to
dump
to
disk
it's
saying
so
it's
like
it's
done
like
a
pre-calculation,
it
thinks
it
has
all
this
data
and
it's
stored
or
it
thinks
it
needs
it
and
then
it's
dumping
and
usually
it
would
just
dump
it
if
it
had
too
much,
but
there's
nothing
to
dump.
B
So
I'm
thinking
it's
something
of
I'm
past
passing
around
like
a
wrong
kind
of
object
or
an
object.
That's
just
not,
I
guess
agreeing
with
the
parallelization
yeah.
I
don't
know
joe.
I
sent
you
that
mpi
that's
kind
of
what
I'm
trying
to
do
with
desk
and
I'm
trying
to
pass
that
object
around
the
the
misfit
object
or
the
objective
function
object
is
like.
Is
it
just
something
that
that
object?
Maybe,
like
I
know,
did
you
have
any
enlightenment
when
you
looked
at
that.
A
E
No,
so
the,
although
we
were
using
multi-processing
in
the
simulation
level,
so
we
are
paralyzing
over
source
and
frequency.
So
that's
actually,
where
also
another
complication
is
coming
because
we're
passing
the
object
to
each
worker.
So
it's
actually
doubling
the
memory,
memory,
footprint
and
yeah
yeah.
So
that's
actually
what's
happening.
The
memory
consumption
at
each
worker
is
increased
a
lot.
So
that
was
that's
actually
what
was
happening,
so
it
was
not
necessarily
related
to
perdisso.
Only
point
pardesa
was
used
at
the
the
preconditional
level.
A
B
Field,
it
should
keep
its
own
fields
yeah
right
now
on
the
dash,
we're
just
keeping
the
deep
red
and
then
passing
deep
red
around.
But
then,
when
you
get
into
the
the
derivatives,
you
need
the
fields
again
and
then
that's
when
you
need
to
calculate
it
again.
Deep
and
storing
d
print
just
doesn't
work
for
that
part,
but
it
does
for
the
forward
simulation
yeah,
I'm
going
to
find
a
way
to
pass
a
little
bit
less
around
or
because
yeah
you
need.
You
need
so
much
information.
B
E
C
B
Yeah
I
like
the
storing
idea
just
because
they
are
competing
on
the
fly
some
when
the
problems
get
big
it
just
yeah.
It
takes
up
the
majority
of
the
time.
So
I
don't
know
I
got
my
boss
is
really
opposed
to
pulling
it
like
saving
to
disc
and
pulling
in
yeah.
I
don't
know,
I
think
it's
at
this
stage.
If
you've
got
like
a
good
enough,
nvme
drive
you're
not
hit,
but
you're
still
not
there
like
with
ram,
but
I
think
it
might
do
for
the
situation
right.
E
E
So
this
is
actually
what
I
got
today.
That's
the
l,
typical
l
model
and
that's
what
we
have
inverted.
I
just
we
just
use
the
electrical
field,
so
the
setup
is
the
source
is
here
like
far
away
and
then
I
was
only
measuring
the
electrical
field
in
y
direction.
So
that's
reason
why
we're
not
seeing
this
part
but
you're
nicely
seeing
that
part
and
a
little
bit
of
that
part.
E
I
think
if
I
use
the
magnetic
field
or
other
excitation,
I
could
have
obtained
better.
So
this
actually
takes
about
an
hour.
I
think
it's
10
frequency
ish
and
it's
about
60
cube.
So
it's
not
a
big
model,
but
I
think,
what's
actually
cool
about
theater's
code,
it's
actually
linearly
scaled
with
the
the
size
of
the
domain.
So
even
if
I
use
like
millions
of
signals,
it
wouldn't
take
that
much
longer
so,
but
the
problem
we're
having
the
the
memory
is
sucking
up.
E
E
How
that's
gonna
make
an
impact,
but
it's
pretty
easy,
like
it's
almost
embarrassingly
parallel
problem,
so
only
thing
that
I
need
to
do
is
just
frequency
and
source
and
then
only
the
tricky
part
was
passing
the
fields
when
I'm
computing,
the
jvac
and
jt
vac.
So
basically
the
same
problem
with
what
john
is
facing.
I
guess,
but
anyway,
that's
where
we
are.
A
B
A
B
A
Like
unless
you
like
specifically
want
to
look
at
a
field
right,
then
you
don't
need
to
care
about
it
as
long
as
each
mpi
process
has
access
to
the
fields
it
needs
to
use,
and
that
should
be
the
case
because
it's
like
usually
with
mpi,
you
break
it
up
the
same
way.
Every
time
you
go
into
a
parallel
region
like
the
same
process,
gets
the
same
chunk
of
things
every
time
right
so.
C
Sounds
good,
I
got
a
maybe
a
question
for
you,
john,
because
are
you
you're
working
on
stuff
related
to
frequency
domain
em?
I
was
talking
with
dom
the
other
day
because
I
was
doing
a
really
basic
example,
or
I
just
I
was
just
inverting
a
really
small
airborne.
Fem
example,
and
I
noticed
that
you
couldn't
apply
a
pre-conditioner
for
that
that
problem.
C
So
as
it
stands
right
now,
just
the
fem
simulations
they
don't
seem
to
have
a
like
a
get
diag
j
transpose
j,
and
I
know
on
the
tiling
stuff.
You
guys
are
working
on
you
could
you
could
play
around
with
that
code
and
you
could
define
a
single
tile
and
you
could
do
it,
but
I
was
wondering
if
it's
that
much
trouble
to
have
a
simple
implementation
like
a
non-tiled
and
being
able
to
apply
the
preconditioner.
B
Yeah
that
might
be
a
little
bit
deeper
yeah
dom
might
be
able
to
feel
that
a
little
bit
more,
I'm
I've
dabbled
into
the
fem
stuff,
but
mostly
I'm
just
merging
it
in
so
that
we
could
use
that
as
the
base
for
the
mt
but
yeah
I
thought
there
was.
There
would
have
been
there's
not
oh
okay,
you're
saying
that
there
is
a
get
jt
j
dieg,
but
it's
the
tiled
one,
not
yeah.
B
C
A
Yeah,
as
I
said
before,
I've
gotten
that
soggy
brought
it
up
that
we're
gonna
bring
in
what
we
had
at
least
the
simulation
parts
in
right
now
and
then
keep
working
on
the
rest
of
it.
So
it
would
be
the
like
I
mentioned
before.
The
things
that
need
to
be
worked
out
are
bringing
in
the
stitched
regularizers
bringing
those
into
like
the
center
regularization,
as
well
as
figuring
out
what
to
do
with
all
like
the
mesh
I
o
stuff
or
the
all
the
input
output
stuff.
That
was
there.
A
E
And
speaking
of
devin's
comment,
while,
like
I
keep
tracking
where
the
memory
is
leaking,
one
of
my
suspicion
was
on
the
party
so
and
that
use
of
because
it
we
are
using
that
as
a
pre
like
when,
when
we're
generating
the
preconditioning,
so
the
default
preconditioning
is
different
from
what
devin
said
like.
But
devin
was
mentioning
it's
a
jacobi,
so
you
form
j,
compute
the
jtj
diag
and
use
the
diagonal
term
as
a
pre-conditioner,
but
the
actual
sim
perks.
The
syntax
default
is
not
that
one.
E
So
what
simpac
does
like
it
uses
the
regularization
as
a
pre-conditioner
and
then
actually
that's
the
initial
gas
and
then
use
the
bfgs
to
precondition
the
h10
vector
system.
So
that's
what
it
does
and
it
actually
amazingly
working
well
to
me.
I'm
not
sure
how
other
people
think,
especially
it
works.
Well
yeah.
So
actually
I
I
was
curious.
How
like
how
important
it
is.
E
Anyway,
so
that's,
but
that's
also,
it's
I'm
not
sure
that
could
also
be
a
part
that
could
lick
the
memory
so
now
partisan
comes
in
as
a
pre-conditioner
and
that
that
could
actually
kind
of
cause
that
memory
leakage
anyway.
That's
a
sort
of
a
point
that
I
was
thinking
and
I
always
keep
tracking
just
one.
Thank
you.
A
The
new
scythe
unwrapper
automatically
does
like
it'll,
do
that
I
have
it
set
up
to
do
that
properly,
but,
like
so
I'd
be
curious.
When
last
time
you
updated
finance
solver,
because
that
was
like
this
was
the
most
recent
version
that
I
added
a
thing
that
just
cleared
it
whenever
it
got
garbage
collected
or
it
should
have
cleared
it.
Oh,
I.
A
G
A
Other
than
that,
that's
I
can't
I
kept
meaning
to
put
out
a
post,
but
let's
look
at
maybe
the
next,
maybe
in
two
weeks
having
it
outside,
get
together
with
people.
I
know
everything
are
kind
of
so
tough
to
do
it
like
right
away,
because
people
usually
have
plans,
but
we'll
put
some.
I
can't
wait
to
put
a
thing
out.
Let's
look
for
about
two
weeks
out,
if
that's
okay
like
a
saturday
or
something
so
those
people
in
vancouver.
If
you
want
to
meet
up.