Convert your model to tensorflow lite Video Lecture | Machine Learning with Java - AI & ML

hello explorers and welcome to another
video today I'm gonna talk a little bit
about tensorflow lite and how to
take your model and convert it into
tensorflow lite model and also I'm
gonna show a little demo on it how it
works in practice and so first off we
look into this little file here we have
this training that we had before with
tensorflow data and I haven't changed
that much in this file but I wanted to
highlight one thing that is changed here
I've actually added an extra layer and I
also added a drop out layer down here so
one convolution more and a drop out
because how much ever trained this model
I didn't get it better than 74% but with
the next convolutional layer and the
drop out layer I got it out up to 84%
and I think I need a little bit to
actually be able to predict things and I
also added some identities here so I had
had an output prediction and output
class and this I will not use I just
added some extra to know which things in
my model was which so I can
actually access them when I convert
later on so output prediction is
important and I also in at the start
here set an identifier for the input
tensor so actually know where my input
data comes and where my predictions
comes out and I ran this for I think
four or five hours so an afternoon and I
got it up to 84% somewhere there it was
actually up at 84 very early and so I
have added some extra text here into the
tensorflow data repositories you can look
there if you like to see how to
changed your graph so first of you when
you are trying to convert it as a tensorflow
lite model and you have just trained
the models so you have a directory with
your model in it you need to freeze it
so it actually has a specific point in
time that you have saved if you are
using the saved model you can use that
right off the bat and convert that but
if you are training a model and it just
does an output to a folder then you need
to freeze it first so you run this
command and so I have my input graph
here I have a specific check point that
I want to freeze my graph at I say as
this is a PB text up here then I say
that it's not binary so that's false I
have an output graph so that's the
frozen graph that I put out and here we
have my output node names and
this is actually wrong because we said
that we had input tensor and outputs pred
so that's the actual node names in your
graph that you are interested in so you
are giving the command freeze graph some
information about which names you really
need to keep in your graph so when it
freezes it it doesn't remove those names
and then you need a command called toco
it's the tensorflow lite optimizing
converter and it is available in the PIP
package when you install it but I was
not able to run it from there it just
gave me an error that it wasn't
available in the PIP package yet so I
had to build it with bazel and that
works fine on a Linux machine so
everything here freeze graph and so on I
had to actually run my model in Windows
and give it to my little VirtualBox and
convert it in there and then put it back
into my development environment
again and that was a little bit of a
hassle back and forth and then we
have this command down here when we run
toco we say where our input file is we
tell it that this in this case it's a
tensorflow graph def and we can also tell
it that it's a tensorflow saved model
for instance if we just save the model
so you can look into that there's a lot
of different parameters that you can put
into this and the output format should
be tensorflow lite model and we say
where this tensorflow lite model should
be at so this is cats and dogs tensorflow
lite and then we have my input here
it's not reshape its input tensor that's
my input arrays and then output pred
in my case that's my output arrays and the
input shape this is the size of my image
so it's 224 times 224 and three
colors and this one up here is just so
you get the matrix correct and yeah so
that's pretty much it and then I use the
demo available by a Google tensorflow or
in the tensorflow repository to try
this out on my device we can go over to
that demo and one of the things I've
done first off I went in here to the on
activity create and the camera to basics
and here I added my classifier so this
is image classifier cats versus dogs and
then inside of my new classifiers I
copied one of the classifiers that I
already had there I added the model path
so I put in my little tensor flow model
here and then I put in some labels and
then I changed the size over here so
this
they tried the new model that's very
interesting and then we have this little
thing down here
where they have actually taken each
pixel my value and fetch the mean of
that each pixel values they have the
rounded the pixel a bit and removed some
noise I guess and for my case it didn't
work as well if I remove that kind of
noise so I just take the pixel values
straight off so I get each channel into
my image data and that works better for
this model because it's trained on real
images I guess and then I keep the rest
of the file intact
so this labels file that could be
interested in looking at so this is just
cats and dogs and dummy and in their
model I created now it's just cat and
dog and so cat is zero and dog is one so
this is just to tell the model which
labels should be used for each
prediction in the output so let's
connect my device here and try some
inference so now you see here that I
have actually a dog over here so this is
running an inference on dog let's see if
we can figure out that this is a dog
yeah it's turning around it's turning
around and now it's pretty sure that
this is a dog so let's switch pages here
and see should these be a cats can it be
a cat it's not really sure yet do you
see it yeah it's starting around and now
it's pretty much sure that that's a cat
so you can see here even though I didn't
have that a well-trained model it's
actually able to do
between cats and dogs and with some
certainty it's not perfect but it's at
least something that you can start with
and play around with your own model to
check the difference between cat and dog
and this model I'm actually running now
it was before I actually retrained the
model and this afternoon so that was a
model I ran this morning and it didn't
have the drop out layer and I actually
got an error when I tried to convert
that model into tensorflow lite it
said that was some operations that was
not supported and when I tried to run it
I also got an error so maybe it's not
possible to actually run your models
with the drop out layer which i think is
very sad because drop out layers
actually makes your model more robust it
might be something else that I put into
the model that actually did that but I
think it was a drop out because that's
the only new layer that I added and I
hope that you found this video
interesting I hope that you learned
something today if you have any
questions about transforming your model
back and forth converting your model into
tensorflow lite please leave them in the
comment section below if you like this
video give it a like share it with your
friends and colleagues and if you
haven't subscribed yet please do that
and I really hope to see you in the next
video