►
From YouTube: Day 2 Tropical Cycle Detection Lab & 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
All
right
so
today
is
going
to
be
a
lot
of
hands-on
components
for
you,
I'm
going
to
go
ahead
and
let
caleb
kick
things
off
and
we're
going
to
talk
about
this
tropical
cyclone
detection.
A
So
this
is
where
the
challenges
channel
within
our
slack
workspace
comes
in
so
as
you're
getting
results,
feel
free
to
post
in
there
and
it'll
just
be
nice
bragging
rights
at
the
at
the
end
of
the
day
and
we
may
or
may
not
have
you
share
your
solution
with
the
group
depends
on
if
you
want
to
do
that
or
not.
So
that's
enough
out
out
of
me
for
the
day,
so
I'm
gonna
go
ahead
and
turn
things
over
to
caleb.
B
Awesome
thanks,
troy
all
right,
so
we
all
have
to
get
on
the
cluster
again.
Hopefully,
everybody
remembers
how
to
do
that.
If
not
the
basically
just
go
through
the
same
thing
you
did
yesterday
get
on
to
curiosity.
B
C
B
C
B
C
B
Let
that
connect
yeah,
so
all
I'm
doing
right
now
for
anyone
that's
strolling
in
like
we're
just
connecting
to
the
cluster
again,
it's
exactly
the
same
steps
you
did
yesterday
from
the
slide.
That's
the
slides
that
are
under
the
announcement
that
jeremy
posted
yesterday.
So
we
wait
for
that
to
kick
up,
you
can
see.
Mine
did
so
that's
good.
B
And
the
reason
we're
doing
this
again,
the
same
way
is:
you
have
to
get
a
new
key
for
everyone
who's
coming
in,
so
you
can't.
B
B
So
I
know
they
have
a
new
windows
terminal,
but
I
still
use
command
prompt
and
then.
C
B
B
B
B
B
B
There
we
go
all
right
so
that
took
a
little
bit
of
time.
For
me,
I'm
going
to
give
everybody
a
minute
or
more.
I
think
it's
still
loading.
B
And
then,
what
we'll
do
is,
instead
of
intro
to
deep
learning,
we'll
be
in
tropical
cyclone
estimation,
intensity,
estimation
right
and
there's
a
there's,
a
series
of
notebooks
in
that,
but
before
we
get
into
all
that,
I
have
some
slides
that
go
over
the
really
high
level,
what
we're
doing
right
kind
of
like
ai
for
science.
Let's
look
at
this
tropical
cyclone
detection,
so
I'm
going
gonna
give
it
maybe
a
couple
more
minutes,
probably
two
more
minutes
and
I'm
going
to
go
ahead
and
start
presenting.
B
All
right
so
yeah
a
couple
more
minutes,
everyone
and
then
I'm
going
to
start
just
presenting,
as
you
can
see
over
here
to
the
side.
It's
not
that
many
slides!
It's
like,
I
don't
know
eight
slides,
so
it's
not
gonna
take
long
and
then
we'll
just
jump
right
in
we'll
get
into
our
breakout
rooms,
and
everyone
will
start.
You
know
working
on
the
the
climate
lab.
B
Cool
and
I'm
looking
around
so
anyone
that
came
in
late,
I
didn't
check
participants
before
I
started
everything
all
we
did
was
you
have
to
log
back
into
curiosity
again,
so
you'll
follow
the
same
procedure.
You
did
yesterday
from
the
slides
jeremy
posted
in
the
announcements
and
the
whole
reason
we
had
to
do.
That
is
we
got
to
generate
a
new
key
for
today.
B
So
let
me
actually,
if
I
push
this
over
here,.
B
An
issue
or
anything
like
that:
will
you
go
ahead
and
raise
your
hand,
if
not
I'm
going
to
start
in
the
next
minute,
just
because
we're
short
we're
not
really
short
on
time,
but
there's
a
lot
of
material
to
go
through.
B
C
B
D
Yep,
you
guys
can
only
run
one
job
at
a
time.
So,
if
you're
seeing
your
job,
like
somebody
mentioned
that
they
saw
their
saw
a
job
that
was
listed
for
max.
C
A
B
And
thank
you
I'll
go
ahead
and
so
jeremy
is
saying:
if
you
hit
sq
when
you're
on
curiosity,
you
can
see
here
that
here's
your
name,
I
don't
oh
yeah.
These
are
these
individuals.
You
have
to
cancel
so
just
hit
s
cancel
and
you
can
use
your
name.
So
you
can
do
something
like.
B
I
think
mine's,
kate
smith
right
might
have
to
be
dash.
You
I'm
not
the
best.
Slurm
guy
should
be,
but
you
can
cancel
your
jobs
and
start
over
or
you
can
cancel
the
individual
jobs
outright
by
hitting
the
number
here.
That's
canceled,
you
know
12
1,
6
4,
that's
yours,
but
there's
not
too
many
people
that
raise
their
hand.
So
we're
going
to
go
ahead
and
begin
that
way.
We
can
get
into
our
breakout
rooms
and
then
we
can
continue
getting.
A
B
We
had
cat
five
hit
four
years
ago
when
I
was
here
and
I
remodeled
my
outside
bathroom
when
when
it
came
so
it
was
a
great
use
of
time,
but
it
wasn't
too
crazy.
B
B
B
For
the
space
coast
here
but
anyways,
this
is
near
and
dear
right
for
floridians
or
anyone
interested
in
climate.
So
we're
going
to
look
at
some
satellite
images
and
from
that
and
some
wind
data
that
we
actually
have
we're
going
to
do
a
multimodal
approach
to
categorize
these
these
cyclones
and
that
and
that's
our
goal
for
today
right.
B
So
it's
a
little
bit
about
everything.
So
this
is
our
input.
That's
how
our
input
data
is
going
to
look
we're
in
the
232
by
232
pixels,
a
lot
larger
than
our
28
by
28
that
we
were
dealing
with
yesterday.
We're
also,
I
think,
going
to
yeah.
So
this
is
the
the
paper
that
they're
going
off
of,
and
I
mentioned
yesterday
you
know
you
typically
see-
and
I
think
all
the
tas
mentioned
too
power's
a
two
right.
You
can
see
here.
B
We
have
some
very
wonky
numbers
as
a
number
of
filters
in
a
layer,
288
and
272..
So
there
is
actually
a
paper
that
this
was
published
from
this
exact
title:
tropical
cycle
intensity
estimation
using
a
deep
cnn.
They
can.
A
B
In
a
10x10
kernel,
it's
pretty
pretty
interesting,
so
estimating
a
little
background
right
this.
Oh,
I
still
can't
pronounce
that
name.
Durvrock
durbrock
technique.
Meteorologists
have
been
attempting
to
estimate.
You
know
tropical
cyclones
from
satellite
imaging
imagery
for
a
long
time
decades
now,
and
when
this
individual,
this
vernon,
I
think
vernon
drovoc,
he
submitted
a
paper
describing
how
manual
pattern
matching
can
actually
give
an
idea
from
satellite
imagery
what
the
category
of
the
cyclone
is
going
to
be
right.
B
It's
crazy
to
think
that's
still,
2000
that
it
was
actually
automated
right
and
the
algorithm
has
been
repeatedly
improved
upon
and
up
till
the
present
day.
This
is
like
the
tried
and
true
for
noah
they're
on
version
nine.
It
looks
like,
and
it's
still
in
operation
in
noaa.
However,
these
scientists
were
like
hey,
we
can.
We
can
use
deep
learning
right.
B
We
know
deep
learning
is
great
on
images,
why
not
use
deep
learning
and
see
if
we
can
make
a
better
prediction
and
a
faster
prediction,
so
the
goal
this
lab
is
to
see
if
we
can
indeed
do
that
if
this
algorithm
can
achieve
the
same
accuracy
or
better
than
this
benchmark
technique
that
noah
uses
today-
and
this
is
the
cnn
model
they're
using
it's
based
off
that
paper
and
it's
kind
of
it's
very
similar
to
lynnette
other
than
the
like
this
288
and
272.
I
believe
in
this
third
3584
dense
layer.
B
So
this
is
a
cool
picture,
though,
because
yesterday
I
don't
know,
if
you
remember,
we
all
were
looking
at
classical
machine
learning
versus
deep
learning
and
we
have
this
input
right.
So
we
have
our
input
and
you
can
think
of
all
of
these
3d
blocks
that
they're
showing
in
this
diagram
as
the
feature
extraction
right.
So
that's
the
convolutional
neural
network
part
and
then,
when
you
get
that
last
layer
in
the
cnn,
you
pass
it
through
to
a
fully
connected
layer.
That's
your
classifier
right!
B
B
Then
we're
going
to
train
and
we're
going
to
evaluate
and
what's
really
interesting
about
this
is
this-
isn't
just
an
easy
data
set
right,
so
we
have
one
two,
three,
four:
five:
five
categories
for
hurricanes
and
two
categories
for
tropical
storm
and
depression
and
then
a
no
category
which
is
really
interesting
right.
There's
no
category.
B
It's
really
good
to
have
things
like
this.
You
know
an
unknown
just
for
outliers,
right
things
that
just
don't
make
sense
and
it
can
throw
off
precision,
recall
or
accuracy
for
like
real
world
deployment,
and
you
can
see
that
the
wind
speed
associated
with
these
is
also
part
of
the
category
right.
So
we're
actually
going
to
look
into
using
this
downstream,
no
pun
intended.
B
B
Too,
it
looks
like
in
these
satellite
images,
the
cat
2
and
the
cat
4
looks
similar
sizes
as
the
cat
3
and
the
cat
5
look
similar
sizes,
but
we
all
know
wind
speed
has
a
huge
impact
to
do
with
that
right.
You
can't
go
just
off
of
size
of
the
hurricane
alone.
I
think
one
of
the
largest
and
widest
hurricanes
to
hit
since
I've
been
here
was.
B
B
B
So
looking
at
it
again,
we
mentioned
here's,
here's
our
network
and
the
loss
is
going
to
be.
I
mean
this
is
just
going
through.
What
we're
doing
the
loss
is
going
to
be
multi-class
cross-entropy
we're
going
to
use
this
sgd
subgradient
descent,
and
this
is
off
on
the
numbers.
I
just
don't
know
why
I
said
they're
fixed,
so
we're
going
to
72
for
training,
we're
going
to
use
18
for
testing,
which
gets
us
to
90
and
our
validation
set's
going
to
be
10,
and
this
is
the
summary
of
the
approach
right.
B
So
we
are
going
to
basically
interpolate
some
text
data
to
find
the
velocity
at
every
point
that
we
have
a
satellite
image.
So
we
compare
the
two
and
then
we're
going
to
do
a
multimodal
approach
to
feed
into
the
network
and
get
a
predicted
class
at
the
end
and
again,
they'll
talk
more
about
this
in
the
notebook,
so
I'm
kind
of
glancing
through
it,
but
I
want
to
make
sure
everyone
has
enough
time
so
we're
doing
an
image.
That's
the
original
image
is
1024x1024x3
right,
so
it's
rgb
too.
B
Now
from
there,
that's
just
to
get
a
handle
on
it
right
we're
going
to
choose
random
232
by
232
by
three
patches
from
that
256
by
256
by
three
resizing
and
we're
doing
that
right.
So
our
algorithm,
our
model
doesn't
know
where
this
hurricane
is
going
to
be
in
the
vicinity
of
the
patch
right,
because
if
we
give
it
to
256
by
256
every
time
and
every
time
the
satellite
image
has
a
shot
with
the
hurricane
or
the
cyclone.
B
B
And
again
in
the
notebook
just
make
sure
we
shut
it
down.
That's
the
most
important
thing
there
is
and
we'll
go
ahead
and
get
on
this
so
again
before
we
even
begin
right
check
your
kernels
make
sure
you
don't
have
any
kernels
running.
B
B
And
yeah
just
take
your
time,
read
this
through
and
really
get
an
idea
of
what
we're
doing,
and
I
mean
that
from
all
the
all
the
slides,
all
the
slides,
all
the
notebooks
right.
So
the
notebooks
are
extremely
well
written
approaching
the
problem.
Here's
the
data
we
have.
You
could
look
at
the
data
on
your
on
your.
C
B
B
All
right
so
from
there,
I
think
I'm
just
gonna.
Let
you
all,
I
think,
we're
gonna
start
to
break
out
rooms,
troy
and
get
going.
It's
12
almost
12
30.,
9
30
right
now,
and
you
have
an
hour
and
15
minutes
to
work
on
this.
Is
that
right?
Let's
see
we
just
had
to
slide
up.
When
are
we
stopping
this
10
30?
B
Yes,
10
30.!
You
have
an
hour,
so
that's
that's
good
enough
time,
but
really
dig
in
and
then,
when
you
get
to
the
challenge
the
competition
here,
it's
basically
going
to
try
to
utilize
everything
from
the
previous
notebooks
for
you
to
get
the
best
training
performance
for.
B
Testing
performance
and
it's
a
great
data
set
because
there's
imbalance
right
so
there's
a
notebook
about
imbalanced
data
set,
and
we
have
that
because
you
know
how
many
cat5
hurricanes
are
there
in
the
world
right.
There's,
not
many.
I
think
in
the
last
you
know
several
decades.
Only
six
has
made
landfall
on
this
side
of
the
united
states
alone.
So
there's.
B
Of
data
on
cat5,
so
that's
something
we
got
to
keep
in
mind
right.
So
it
goes
through
that
how
to
work
on
imbalanced
data
sets
how
you
manipulate
and
pre-process
the
text
data
to
the
image
data,
but
the
competition
you
know
feel
free
in
your
breakout
room
to
just
make
it
open
right,
open
collaboration.
B
If
you
want
to
work
by
yourself,
that's
totally
fine,
you
don't
have
to
talk.
Obviously
you
don't
take
your
camera
off
anything
like
that.
You
can
just
you
know,
nose
to
the
grindstone
and
get
to
it.
B
But
if
you,
if
you,
if
you
want
to
collaborate
with
your
teammates
in
your
lab
in
your
breakout
room,
please
do
there's.
There's
no
issues
in
that
and
then
I'll
check
in
periodically
and
see
how
everyone's
doing
is
there
a
slack
room
for
the
challenge,
troy.
B
The
very
last
accuracy
this
right
here,
the
test
model
against
validation
set.
So
your
accuracy
and
score
your
loss.
Please
post
that
in
there
and
you
know
we'll,
give
come
up
instant
and
praise
to
the
highest,
the
highest
accuracy,
lowest
loss
right,
but
yeah.
So
good
luck
and
troy,
it's
all
you.
A
Awesome
thanks
so
much
caleb
all
right.
I'm
gonna
go
ahead
and
open
up
the
breakout
rooms.
The
configuration
of
the
breakout
rooms
wouldn't
save
for
me
last
night,
so
you'll
probably
be
in
a
different
room,
but
again
each
room
has
a
ta,
so
the
ta
will
just
introduce
themselves
briefly
again
and
you
are
free
to
move
around.
A
So
if
there's,
if
you
were
involved
in
conversations
yesterday,
I
know
so
a
few
of
you
have
slacked
me
feel
free
to
jump
around
to
a
different
room
if
you've
already
been
working
with
other
people,
I
don't
mind
and
with
that
I
will
whisk
you
off
into
breakout
rooms.
A
All
right
welcome
back
everybody.
I
hope
you
all
got
a
chance
for
a
break
in
we'll
go
ahead
and
turn
this
back
over
to
caleb.
So
you
can
talk
about
what
we
just
went
through
and
then
introduce
the
next
session.
B
B
We
had
some
people
in
the
challenge.
Get
done
looks
like.
B
That's
really
good
all
right!
So
exactly
what's
up
troy,
I'm
gonna!
Let
you
quit
sharing
your
screen.
B
All
right,
hopefully,
I
can
see
that
I
think
the
the
best
part
of
this
lab
right.
We
had
some
good
discussions
so
yeah,
but
the
text
data
was
used
to
label
those
15
minute
increments
right
for
the
satellite
to
help
label
it.
So
we
could
get
something
going
with
training
and
the
imbalance
part
that's
one
of
my
favorite
ones
right.
So
we
use
something
called
data
augmentation
and
you
know,
there's
a
lot
of
scrutiny
behind
the
augmentation
of
the
satellite
images,
for
instance.
B
Right,
I'm
not
sure
if
that's
true
or
not,
I
had
somebody
bring
that
up
in
a
the
nasa
boot
camp
that
you
couldn't
have
rotate
a
certain
way
because
hurricanes
don't
spin
that
way.
So.
C
B
Right
so
we
we
just
flipped
the
image
and
it
seems
to
work.
C
B
Augmentation
techniques-
you
could
do
too
and
that's
something
you
would
have
tried
in
the
competition.
If
you
were
to
go
to
the
competition
and
it's
okay,
you
can
work
on
this
afterwards,
too
right.
If
you
download
it
to
your
local
system,
you
can
do
it
and
I
think
they
have
access
to
curiosity
throughout
the
weekend.
B
D
B
Yeah
really
really
good
lab
the
approach.
The
problem
was
great.
This
alternative
approach
was
something
kind
of
skeptical
someone
brought
up,
so
I
was
trying
to
read
through
it
and
I
don't
quite
get
it
other
than
we're.
Gonna
try
to
take
the
six
hour
increments
and
break
them
down
to
15,
which
seems
kind
of
impossible.
B
B
This
it'll
take
you
to
their
paper
where
they
go
through
what
they
were
doing,
why
they
did
what
why
they
picked
the
sizes
they
picked,
and
things
like
that
so
check
that
out.
If
you're
curious
on
anything,
you
did
in
the
notebook-
and
this
is
pretty
cool
because
they
actually
go
through
each.
B
B
So
I'm
doing
that,
oh
there's
something
I
have
to
bring
up
everyone,
please,
when
you
finish
a
tab
when
you
finish
a
notebook,
please
shut
down
your
kernel
right
and
when
we
leave
them
up,
you'll
get
ooms
and
other
than
that
it
will
drain
the
node,
we'll
get
a
bunch
of
errors,
and
it's
just
not
good
right.
So
we
desperately
need
you
all
to
close
your
kernels
shut
them
down
when
you're,
not
working
on
that
notebook.
B
As
you
can
see,
you
know,
I
ran
the
approach.
It
saves
everything
out
when
you
close
down
the
kernel.
It
doesn't
like
delete
it
so
you'll
have
it
there
to
look
at
if
you
want
to
go
back
so
yeah.
So
just
just
try
to
do
try
to
remember
to
do
that
without
me.
Having
to
constantly
tell
you
like,
I
do
my
kids
all
the
time
all
right
so
now
we're
going
in
this.