Building tensorflow on windows (Update) Video Lecture | Machine Learning with Java - AI & ML

hello explorers and welcome back to
another Java video and today I'm going
to talk a little bit more about
tensorflow windows and java and last
video I said that for the moment running
Java on windows or building tensorflow
on windows should not even be considered
because there were too many hurdles and
you weren't able to actually create
something in Windows and that was pretty
much because of the bazel version that
were was out then 0.4.4
didn't work in Windows and didn't
build a new version of bazil so there
were a lot of hurt there they have
actually been a release of the bazil in
the middle of mars where they released
0.4.5 and this
version is actually able to build bazil
with the same structure that I tried out
earlier this works and you get a bazil
install that you can later use with
tensorflow
but you need to download bazil the
binary in order to build bazil so I'm
I'm not sure if that helps us really and
because we want then the new version of
bazil and the new version of bazil is
the one that is out and when it comes to
building tensorflow from source I was
able to actually build test flow from
source and everything went fine with GPU
support and everything the sad part is
that when tensorflow
has built from source you should create
a pip package to install it on your
system and that pip package is built
from an output directory where the build
source will be put and that build
directory is empty
so the PIP package becomes pretty sparse
I added the things that it was looking
for into this build directory and it was
able to build the PIP package and
install the PIP package
sadly the bindings between the dll's and
the library was not sufficient to run
and actually train the data but the
version of tensorflow 1.0.1
is actually out as a pip package now so
you can install that directly from the
repository online and that will actually
have this save module function that we
were using in Java so we can run with
that and train your model with give you
support and I ran it for a couple of
minutes and I got a result of
99.15 which is a little
bit worse than the first time around it
but it can differ a little bit so I
don't think that it is actually because
of the installation it could be
something else and the Java bindings
were I was actually able to build those
from source without any hiccups without
any problems the only thing there was
that you needed to change the name of
them so we can go over here to my little
work log here everything in this work
log is up on github so you can read it
there and here in under experimental I
put some of the documentation of what
I've done I will flush this out even
more but here I have all the things
needed to set in order to run an install
compile bazil and then I checked out
tensorflow revision 1.1 and
and download numpy set some
variables configure it with a lot of
parameters I think in my case I only
need the feature 6.1 because that's what
my computer graphics card is actually
supporting but your miles may vary so if
5.2 or 6.1 then you
actually cover most of the newest cards
that could be a good way but there is
actually a page where you can look this
up I will put that into this tutorial as
well and then you just build this
pip packager which will create a pip
package for you and these I found out
these build options are quite
interesting as you actually need to set
your architecture your CPU and host CPU
in order for it to build
so these build options are required to
build this package in Windows and that I
found in one of the issues on github so
I think there might be some some more
issues before you get a smooth build of
tensorflow but then again if you have a
package already working then you don't
really need to build it by your own but
I wanted to try it out and this is how
you later build the pip package and this
is the command that didn't work for me
and here we have how to build the jni
Java bindings and the thing to note
there is it will actually create a lib
tensorflow
underscore JNI dll so and you need to
change the name of that to tensorflow
underscore JNI dot DLL in order and put
that in your jni map in order for
tensorflow to pick it up and
actually get it working so if you build
your jni files by your own
don't forget to change the name of your
jni bindings dll to the new name and
then you also get the jar file that you
need to install into your maven in order
to get that working and I have the
instructions how to do that in the Linux
tutorial so those are not different from
the windows tutorial and I think
everything in works when it comes to
building the jni is not different from
the Linux tutorial so then I've talked
about everything now I just had to run
it in Windows and see that my Java my
own built Java bindings worked and I
actually got the result of
99.15 which is the same result as I
got inside of tensorflow
Python bindings when I train the network
so I'm quite happy that even though I
running on my own build so this is 1.1.0 rc1 which is that revision
that I built and I'm quite happy that I
can build my own Java bindings get those
to work
sadly the Java bindings do not support
GPUs at the moment so it will be a bit
slower but I still think if you just use
these to verify or use these to actually
find results or predictions I think they
are fast enough so yeah that was my Java
update on how to build tensorflow on
Windows last time it was not possible to
do and much has changed since last time
I talked about this I hope that you
found this information hopeful and
interesting and please leave me comments
if you try to run
tensorflow for your own on Windows and
what your result was and if you have
tried to build it on Windows please give
me a hint on how to do that and I really
hope to see you in the next video