►
From YouTube: ONNX20210324 V20 VisualizingONNXmodels
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
So
the
zetain
viewer
is
used
to
open
the
ai
block
black
box
and
just
like
microscopes
and
telescopes
and
other
like
brain
imaging
techniques
that
open
the
new
way
to
look
at
complex
systems
with
the
zetaine
viewer.
We
we
want
to
look
at
internal
tensors
of
the
network
and
internal
structure
of
the
networks
and
neural
networks
and
to
get
insights,
how
to
improve
their
models
and
how
to
improve
the
training
of
the
models
and
the
freesetting
viewer
can
be
used
to
visualize
the
architecture
and
tensors
of
the
onyx
models.
A
So
you
can
easily
load
your
onyx
model
and
see
the
structures
and
internal
tensors
and
to
get
the
the
time
viewer
you
can.
You
can
have
a
look
at
the
our
github
page
or
our
website
page,
and
you
can
download
it
for
windows,
linux
or
mac.
So
in
this
presentation,
I'm
gonna
present
some
examples.
How
visualization
helped
help
us
in
a
few
projects,
and
I
categorized
the
different
aspects
into
three
different
groups
of
data
network
architectures
and
internal
tensor
inspection
and
because
of
the
limited
time
I
only
focus
on
the
segmentation
problem.
A
A
So
in
semantic
segmentation
we
don't
care
about
different
objects
from
the
same
type
and
we
will
consider
them
as
one
class
as
you
can
see
in
this
example,
but
in
the
instance
segmentation
we
care
about
different
instances
of
each
category.
So
in
this
case
we
have
six
labels
or
six
classes,
five,
four
different
people
and
one
for
the
background.
While
in
this
in
semantic
segmentation,
we
have
only
two
labels,
one
for
the
people
and
one
for
the
background
and
to
evaluate
segmentation.
There
are
different
measures
used.
A
The
first
one
is
pixel
accuracy
and
the
other
more
complex,
complicated
indexes
are
record
index
and
dice
index,
and
I
don't
go
into
the
details.
But
as
but
as
you
can
see
they,
they
can
be
some
values
between
0
and
1
and
where
0
means
there's
no
overlap
between
the
ground,
truth
mask
and
the
segmented
mask,
and
one
means
a
perfect
and
complete
overlap
between
the
ground
through
mask
and
segmented,
mask
and
the
perfect
segmentation.
A
So
let's
go
into
the
something
to
see
some
examples
of
how
visualization
helped
us
improve
our
neural
network
architectures.
As
you
all
know,
data
is
really
important
in
deep
learning.
We
need
to
make
sure
that
we
are
feeding
the
incorrect
and
the
good
data
to
the
network,
because
the
network
will
learn
what
we
have
what
we
have
said
to
him.
A
A
But
just
out
of
curiosity,
we
wanted
to
see
what
the
model
sees
inside
the
network
and
we
noticed
some
strong
activation
at
the
top
corner
of
the
images
in
different
layers,
and
we
realize
that
the
network
is
paying
more
attention
like
more
attention
to
the
tag
compared
to
other
parts,
and
this
tag
was
used
to
make
the
data
anonymous,
and
we
realized
that,
maybe
by
cleaning
the
data,
we
could
improve
the
segmentation.
A
The
other
aspect
in
neural
network
rather
than
the
data,
is
the
network
architecture.
When
we
are
designing
a
network
we
want,
we
want
to
make
sure
that
all
layers
are
connected
correctly
and
we
usually
suffer
from
the
copy-paste
issues
when
we
copy
the
layer
from
when
we
are
implementing
a
network,
we
copy
the
layer
and
then
paste
it,
and
we
forget
to
change
all
the
all
the
necessary
inputs
or
outputs,
and
especially
when
we
have
a
smooth.
A
We
have
a
big
network
with
lots
of
skip
connections
and
with
like
with
different
activation
functions
with
batch,
normalizations
and
concatenations,
etc.
We
it's
it's
usually
difficult
to
make
sure
that
the
network
architecture
is
as
the
shape
as
we
wanted.
A
In
this
example,
you
you
are,
you
are
seeing
a
unit
and
you
can
easily
see
the
u
shape
and
all
the
connections
between
the
encoder
and
the
chordae
parts
and
if
we
zoom
more-
and
this
is
another
network-
but
we
can
see
all
the
convolutional
layers
and
the
activation
functions
and
the
batch
normalization
and
how
they
are
connected
to
each
other.
And
if
we
zoom
more,
we
can
see
the
tensors
for
each
layers
and
all
the
statistical
parameters
of
the
sensors,
and
we
can
see
visual
weights.
A
We
usually
use
a
piece
of
paper
and
draw
the
and
the
architecture
on
that,
but
by
using
visualization
in
the
same
viewer,
we
can
easily
see
the
network
architecture
and
make
sure
that
the
design
is
correct,
and
the
last
part,
which
is
the
most
interesting
part,
is
how
visualization
you
can
improve,
can
improve
the
model
performance
when
we,
when
we
look
at
more
carefully
at
the
internal
tensors
and
the
future
maps
filters,
biases
and
the
histogram
and
parameter
ranges,
shapes
and
number
of
tensors.
A
So
this
is
one
example
that
we
were
trying
to
segment
some
covet
19
lesions
in
the
city,
images
of
patients
and
we
didn't
have
good
results.
But
we
didn't
know
why.
So
when
we
visualize
the
network
in
the
viewer,
we
realized.
We
realized
that
one
branch
of
the
network
doesn't
pass
the
signal
and
we
all
the
neurons
are
dead.
A
So
we
try
to
add
some
batch
normalization
and
also
we
could
use
another
activation
function,
such
as
leaky
relu,
and
after
that,
we
we
could
confirm
that
the
signal
is
passing
through
this
branch
and
we
could
improve
the
results.
A
This
is
a
another
example
when
we
were
trying
to
segment
the
breast
tumors
in
ultrasound
data,
and
we
didn't
have
good
results
for
some
patients
and
we
maybe
we
we
could
have
used
error
analysis
but
like
the
usual
way,
is
to
analyze
the
errors
and
try
to
find
a
way
to
improve
the
results.
A
But
what
we
did
we?
We.
We
noticed
that
the
histogram
of
the
last
layer
is
different,
is
different
when
we
have
a
good
results
compared
to
when
we
have
a
bad
results
or
poor
results.
So
at
the
top
you
can
see
some
examples
of
good
results
and
you
can
see
it's
it's
so
sharp,
but
for
for
poor
results,
we
have
a
more
proud
histogram.
A
So,
based
on
this
observation,
we
try
to
to
design
it
an
adaptive
thresholding
and
we
could
improve
the
dice
score
in
four
segmented
images
by
by
a
good
amount.
So
this
is
like
this
is
just
one
example
when
this
is
the
ground
rules.
This
is
before
applying
the
adaptive
thresholding
and
the
dice
was
0.74,
and
this
is
after
applying
adaptive,
thresholding
and
the
dice
is
0.82.