Object detection Part 3 - Local training [Tensorflow] Video Lecture | Machine Learning with Java - AI & ML

hello explorers and welcome to another
video today is the third installment in
the tensorflow object detection series
that we are doing here and this part 3
is gonna be about training and it's
gonna be about local training we start
there so if you have a machine at home
that is pretty good and you can do some
inference and training on it
then you can do this locally but in
order to do that you need to install
some stuff so we're gonna start with
that to install the dependencies we need
to actually run this so here we have
some dependencies listed we have
protobuf and that's Google's library to
make a configuration and read
configuration so it's a standardized way
of writing and reading configurations
then we need Python TK and that's also a
language extension to Python
pillow I think that's a image library I
don't know LXML is pretty big much
reading XML files I guess TF slim hmm
and it should be included here so it's
not something that we need, jupiter
notebook I'm not really sure if you need
that because I'm not using any notebooks
in my examples here maybe they have some
internal code that requires it I'm not
sure if you need it for training
matplotlib that's a Lib from MATLAB that
you actually can do some calculations
and so on so that's required tensorflow
of course cython you have a connection
to c libraries context Lib don't know
where that is and the cocoa API and
that's to read the coke or training
files and do evaluation on those and so
first off we install tensorflow you can
either install the pure tensorflow for
CPUs or the one for GPUs if you install
the one for GPUs via pip package
it will tell you to actually set your
path so you have the Cuda libraries in
your path because tensorflow GPU at
least in my installation uses CUDA and
in the latest a 1.8.0
they are requiring CUDA 9.0 and
then you also need to install cudnn
7.0 so I install those in a directory
and then I pointed my path to that cuda
directory slash bin so it could find all
the libraries and then you can use this
GPU I have a long-running series where
I'm trying to build tensorflow so you
can build it from the latest source with
some CUDA library to get a good
compilation there and I haven't been
able to do that yet but you can actually
set it up just by installing the people
packages so that's interesting and I'm
using Windows with msys2 so this is
everything is for Windows at the moment
but I have run this in Linux so I've
used at home my Debian installation that
a multiple boot and I also used that
work in my gentoo installation and it
works fine in both those and if I
believe it's easier to install on Linux
and get it running on Linux than on
windows so this should apply to
Linux as well if you are using ubuntu
you have a quick come on here to install
protobuf and everything like that and in
Windows I used all the PIP packages so I
installed all of these like these
commands here then I needed the coca API
installation and this was a little bit
interesting because first of I got git
repository and I tried to do make in that
git repository but it wasn't able to
actually do make because it sets some
extra flags that is not available in
Windows I will show you that so here
inside of the setup.py inside of this
Python API library you have some
extra compile arguments here and if you
comment those out they will build in
Windows and you will get a directory
called cocoa tooling as it says in this
installation here pycocotools and
that directory you then copy over to
models research and in that directory
this directory should be located in
order to actually do this evaluation I'm
not doing any evaluation on this coco
metrics so I haven't actually tried it
but I think it should work fine it
installs it builds finds which should
work and then you need to compile their
protobuf files
configuration files and this command is
the one that you run but in Windows
you have one little headache that this
protobuf or protoc binary is
actually not installed in the protobuf
package if you install protobuf via the
pacman system you use in msys2 so in
order to get protobuf library or the
protobuf executable I actually just went
to the release page of protobuf and here
under the reached release protobuf
3.6.0 I found this protobuf
windows binary so I downloaded that zip
package open it up put it and then used that binary to compile my
protobufs or convert my protobufs
and then in msys2 you export
pythonpath and add the current directory
this research directory and also slim so
Python knows about the extra libraries
in this models research directory and
after you've done that you run this
python script and it will actually
output some interesting information can
actually run it it's quite fast and when
you run this you will get some status
information I will tell you if you have
were able to actually install everything
and get it up and running and here we
have it
you see run this tooling and it gives an
error up here that something is
deprecated but that doesn't matter and
then when it's done with that it
actually tells it ran 15 tests and all
15 tests were ok so now you know that
you have system up and running and as
you see up here I have actually been
running my set training for a bit so
this is the command that I ran for
training
I run Python in the object detection
I run the training Python script I want
my errors to log out to standard error
so log to standard error of this flag
says and then I need this pipeline
configuration path so I tell it where is
my configuration for my model and this
is the path in my home directory so it's
in c: github rcnn model test
Waldo model
Waldo model config and then I also tell
it where my training director is and
that's everything that is needed to
actually run training so it took a while
to get it started so in Linux
I believe it's actually quite fast to
get it up and running
but in Windows it takes some time for it
to actually accelerate up and actually
get to doing the actual training I'm not
sure if it's a lot of libraries it needs
to read or if I have a slow disk on my
computer or what it is but it took a
while let's say three or four minutes
for it's actually start training but as
as you see here the training steps is
around half a second and I believe in
what Linux
it was around 0.3 or something so it's a
bit slower in Windows but it's not that
much slower so you can use this in
Windows but a dedicated Linux machine
with all the drivers and so on will be a
bit faster I believe still and so that
was the command let's look at the
training configuration we looked at that
last time and the only thing to note in
the training configuration today is when
you are training in Windows you use the
msys paths in your command line but in
your config you need to actually put in
C colon slash because it will not find
your files if you are using the msys
notation and when you are running in
running this training you can also get
to a point where you it tells you that
this shape of your training is a wrong
shape it should be a list of things but
you actually get just a number and
that's a little problem you have in
Python 3 that you don't have in Python 2
and because they are a little bit
different
in this case they have actually not
built this training model in a Python 3
way it works fine in Python 2 but if you
are running into this problem you can go
into this learning schedules Python so
that's inside of your models directory
research and then you go into the object
detection and utils and in this utils you
have learning scheduled dot py and at
the bottom of this file it actually is
just like this reduce max were and so on
and here where you see range you add
lists so you force that to be a list of
numbers and that will work in Python 3
as well so that's a little gotcha and I
hope that they fix it in the actually
source shall git repository but there is
a lot of people talking about this issue
on their issue board so they are aware
of it but I haven't seen they actually
updating the the files so that was how
to run training you can also evaluate
your model that's pretty much the same
you will tell it you want
to standard error the errors to the
standard error and you pipeline
configuration checkpoint there that's
your training directory so that's
directory we has told it to put your
output data into when your word training
and then you want to give it an
evaluation output dir where it will put
the logs from the evaluation so that's
how you run evaluation on your model
next up we have the tensorboard so this
is how you look into your module when
you are training it and that's just
telling it where your output dir is so
your training output dir and it will
give you a view into that and tell
you things about your model and if you
are running it during training you can
also see the value change and you can
get some histograms and so on and that's
quite interesting to look at when you
are running your training of your
model so this was what I wanted to cover
today about training your model locally
I hope that you found this interesting
and I hope that I covered some of
the gotchas that you might run into if
you are trying to train this model
yourself if you have any questions at
all about this approach or anything else
then leave it in the comment
section below if you liked this episode
give it a like share it with your
friends and colleagues if you haven't
subscribed yet and want to know when the
next episode comes out then please
subscribe and I really hope to see you
in the next video