►
From YouTube: Day 2 Steady Flow Estimation Challenge
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
A
A
What
was
it
like?
A
car
company
like
toyota,
they
run
a
cfd
simulation
for
like
two
weeks
and
get
a
result
and
have
to
tweak
something
and
do
it
again,
so
you
can
see
how
doing
something
like
ai
we're
hoping
to
get
the
same,
fidelity
as
numerical
models
but,
more
importantly,
just
speed
this
up
right
because
the
faster
we
can
do
research,
the
more
research
we
can
do.
I
mean
that's
just
common
sense.
A
A
A
Okay,
so
we're
going
to
be
doing
this
with
neural
networks,
obviously-
and
our
aim
is
to
predict
this
2d
flow
around
an
object-
the
inputs,
the
boundary
around,
which
we
want
to
calculate
the
flow,
so
you
can
think
of
it
similar
set
up
to
a
cfd
problem
where
you're
building
a
grid
out
and
in
this
example,
our
input
data
and
the
corresponding
flow
was
calculated
using
the
lattice
boltzmann
method.
So
you
can
click
the
link
in
the
notebook
to
that.
If
you
really
want
to
read
up
on
that,
I
don't.
A
About
ladies
boltzmann,
I
my
undergrad
is
in
aerospace
engineering.
We
did
some
cfd,
but
not
enough
for
me
to
be
anywhere.
D
A
A
E
A
Which
will
make
this
large
data
set
a
lot
easier
to
be
digested
in
to
to
the
memory
for
the
computer,
but
again.
A
Notebook,
please
please,
please
shut
down
your
kernel
and
we're
going
to
try
to
predict
the
velocities
of
both
x
and
y.
So
it's
like
a
you
know,
a
2d
regression
problem
in
a
sense
right.
A
C
A
B
A
A
And
then
something
even
crazier
right
is
this
unit.
I
mentioned
it.
You
know
I
think
steven
had
it
up
on
his
screen.
Yesterday
he
was
talking
about
some
of
the
groundbreaking
stuff
they're
doing
there.
Units
are
super
valuable.
Super
important
and
they're
used
a
lot
in
medical
image
segmentation,
but
many
visual
tasks.
A
So
that
idea
that
you
know
that
segmentation,
so
it
will
be
interesting
if
we
get
to
this
right
on
this-
you
gated
network,
because
it
would
be
cool
to
see
how
this
does
considering
it's,
not
a
segmentation
problem
right.
So
I
hope
we
do
get
to
get
there
and
don't
be
surprised
when
you
see
this
right.
It's
just
just
something
something
to
you
know,
add
on
to
your
repertoire
of
networks
to
try
so.
B
C
A
So
you
should
kill
that
job.
I
don't
know
why.
I
didn't
kill
your
clothes,
so
kill
the
job
that
you
have
that
we
just
did
climb
it
in
if
you're
still
working
on
that-
and
you
don't
feel
like
working
on
cfd
because
maybe
climate's
more
your
type,
there's
no
pressure,
you
don't
have
to
do
the
cfv
lab
right
now,
but.
A
A
B
A
A
A
A
B
A
A
Sorry
for
that
confusion
on
my
part,
and
then
here
we'll
start
here,
just
like
we
did
before,
and
we've
got
notebook
2
and
you'll
start
actually
going
through
on
the
approach,
the
problem,
the
data
and
task.
You
know
what
we're
going
through,
what
we're
actually
doing
again
a
lot
more
plots
in
this
one,
so
it's
kind
of
cool
to
look
at
our
models
and
what
we're
going
to
be
looking
at
and
how
we're
going
to
measure
the
loss
right
and
yeah
I'll.
Let
you
all
go
too.
We
got
an
hour
left.
A
If
anyone
has
questions,
please
ask
in
the
let's
see
what,
if
asking
the
cluster
support,
if
you're
having
issues
getting
that
poor
forwarding,
I'm
sorry
I
made
it
look
so
complicated.
It
was
pretty
simple.
I
should
have
just
checked
localhost
8888,
where.
C
A
G
Awesome
and
we
actually
have
about
an
hour
and
15
minutes
until
12
15
pacific
time.
I
know
I'm
in
central
time
so
trying
to
do
math
today,
it's
just
not
working,
so
you
should
be
thrown
back
into
your
breakout
rooms
from
just
a
little
bit
ago
and
we'll
meet
back
here
in
about
an
hour
and
15.
A
G
All
right
welcome
back
everyone
we're
going
to
go
ahead
and
kind
of
briefly
talk
about
what
we
just
went
through
and
then
we'll
open
it
up
for
a
little
bit
of
q
a
before
we
close
out.
So
I'm
going
to
hand
it
back
to
caleb.
A
Awesome
thanks,
troy
yeah,
so
that
was
I
really.
A
Lab,
I
think
it's
pretty
cool,
get
to
go
through
a
lot
of
like
real
world
stuff.
Now,
there's
not
as
many
questions
in
the
lab
section
as
before.
There
was
a
few
one
of
the
big
ones
that
we
had
in
my
section.
Was
you
know?
Why
did
we
go?
I
think
it's
in
part,
three
yeah.
Why
don't
we
go
from
a
2d,
convolutional
network
and
down
to
a
fully
connected
layer
and
then
scaled
back
up
to
a
convolutional
network,
two
different
ones
right
to
get
the
x
and
y
component?
A
It's
pretty
interesting
for
sure,
but
that
going
down
to
that
fully
connected
layer
from
a
2d
convolutional
network
happens
a
ton
right.
It's
almost
like
an
embedding,
you
think
of
it
as
embedding,
where
you're
just
trying
to
put
together
everything
you
learned
from
that
2d
sense
into
something.
That's
going
to
make
rationality
for
you,
you
know,
learn
what
the
good
embedding
is.
A
So
a
lot
of
good
talk,
let's
see,
there's
a
couple
there's
one
in
the
challenge
that
finished
through
this
that's
great,
I'm
trying
to
think
there's
not
really
much
to
go
over.
Hopefully,
everyone
got
to
see
the
more
advanced
networks
too.
B
A
G
A
G
G
Yeah,
thank
you
to
all
the
tas
and
to
you
caleb
as
well
and
jeremy.
I
can't
do
my
job
without
you
guys.
So,
if
anyone's
got
questions
I've,
given
you
the
ability
to
unmute
yourselves,
we've
got
about
10
minutes
that
we
can
answer
some
questions.
C
I
just
have
general
questions:
is
cnn
some
sort
of
a
black
magic,
or
is
it
some
some
logic
to
to
that?
To
that
madness,
it
seems
that
people
just
do
stuff.
It's
not
always
clear
to
me
why
people
do
what
they
do.
A
B
A
And
it's
not
just
you,
that's
not
me
picking
on
you,
but
a
lot
of
people
say
that
higher
ups
and
everything
industry,
research
and
all
that
I
mean
the
principles
there
right
so
instead
of
having
someone
sit
there
and
try
to
find
these
features
for
hours.
D
A
The
papers
that
these
different
layers
learn
features
that
we
could
never
do
a
handmade
crafted
kernel
to
extract
that
from
an
image
right.
So
it's
pretty
powerful
and
then
you
know
one
other
thing
you
could
do
say
you
know
we
always
connect
these
to
a
fully
connected
layer
at
the
end,
which
is
just
an
mlp.
So
you
have
a
cnn
attached
to
mlp.
You
can
extract
the
features
from
different
layers
right
and
then
send
that
into
your
favorite
machine
learning,
algorithm
random
forest
svm.
A
C
How
do
you
build
a
model
if
you
start
out
with
not
knowing
how
to
extract
the
features
on
the
different
layers?
How
do
you
know
how
many
layers
to
put
in
there
and
what
kind
of
layers
I
want
to
put
in
there?
That
kind
of
things.
A
A
This
to
make
my
application
better
more
than
likely
someone
else
has
thought
of
it
too
or
thought
of
something
similar
and
you
can
use
and
harness
what
they
did
right
extracting
layers.
That's
pretty
simple
and
keras
and
pie
torch
through
the
api
and
the
frameworks
we
have
and
make
it
again
they
try
to
make
it
as
simple
as
possible,
and
then,
on
top
of
that,
you
know
we
talked
about
it
yesterday
in
cnn,
layers
closest
to
the
input
is
going
to
be
really
those
low-level
features
like
you
know,
edges
and
blobs
and
textures.
A
That's
something
that
if
all
natural
images
share
similar
lines
right,
textures
and
edges,
you
probably
wouldn't
use
those
for
a
classification
feature
right,
but
the
higher
you
go
towards
the
output.
Those
features
and
those
convolutional
layers
get
more
distinct.
More
representative
of
you
know
higher
objects
too.
So
that's
something
you
would
use.
I
I
always
say
in
the
papers.
Vgg
is
a
good
one
right
where
they
take
that.
A
E
E
Wanted
to
ask
a
question:
I
asked
in
the
breakout
of
breakout
room
so
here
we
use
the
data
generated
with
this
numerical
solver,
which
is
very
expensive
and
the
users
use
it
as
input
right.
So
now,
let's
say
we
built
a
model,
that's
good
enough,
but
now
we
ask
another
question:
maybe
we
can
so
we
can?
Maybe
we
can
generate
some
other
inputs
to
improve
our
model.
So
now
the
question
would
be
based
on
our
results.
E
A
Yeah,
so
that's
a
good
question,
sorry
about
that.
We
we
definitely
don't
have
enough
time
now,
considering.
I
just
talked
for
a
minute
to
myself
to
really
go
in
depth
with
that
yvonne,
so
the
slack
channel
is
going
to
be
open.
I'd
really
appreciate
if
you
asked
that
in
the
slack
and
then
that
would
give
everyone
a
little
bit
more
time
to
digest
it,
because
it
was
a
pretty
long
question
and
then
you
know
reflect
on
it.
You'll
get
a
lot
more
opinions
than
just
mine
on
that.
That's
okay,
because.
F
You
this
is,
I
take,
and
I
have
a
quick
question
also
about
general
workshop
or
the
bootcamp.
Can
I
ask
yeah
so
when
I
go
to
the
github
repository
that
was
shared
for
the
gpu
bootcamp
official
training
materials,
I
do
see
in
addition
to
the
cfd
and
the
climate,
there's
the
pins
physics
informed
neural
network
related
material.
Is
there
a
separate
like
a
workshop?
You
are
organizing,
for
you
know,
looking
into
the
pins
which
will
be
of
interest.
So
that's
why
I
want
to
check.
Oh
you
try.
G
Yeah,
I
can
take
that
so
earlier
this
year
we
actually
ran
an
ai
for
science
with
all
of
the
content
you
went
through
yesterday
and
today
and
the
pin.
We
call
modulus
content
and
it
was
just
way
too
way
too
much
so
we
are
looking
at.
I
don't
know
when
this
will
happen,
but
we
will
be
running
a
modulus
boot
camp
at
some
point
in
the
future.
F
G
We
don't
know
we
do.
We
do
work
with
a
lot
of
labs
and
and
partners
around
the
world.
So
that's
yet
to
be
determined.
F
G
All
right
lots
of
great
questions,
so
we
will
go
ahead
and
wrap
up
for
today.
Again,
I
want
to
thank
caleb
all
of
my
tas
jeremy
and
all
of
you
for
attending.
I
think
this
is
a
great
boot
camp.
I
just
sent
a
slack
message
and
a
follow-up
email
that
links
to
the
recordings.
G
G
I
don't
have
a
specific
time,
so
just
do
whatever
you
want
to
do
until
sunday
and
then,
whenever
jeremy
logs
on
on
monday
is
when
we'll
we
have
to
boot
you
off
and
prepare
for
the
next
event.
So
thank
you
all
so
much
for
joining
and
have
a
great
weekend.