Build Tensorflow on windows (June-07 update) Video Lecture | Machine Learning with Java - AI & ML

hello explorers and welcome to another
video and today I'm going to talk a
little bit more about how to build
tensorflow
on Windows so this will be somewhat of a
June update i've spent a couple of weeks
trying to build and tensorflow on windows
and I learned a lot of things and I
figured out a lot of stuff that could be
good for you to know and I actually got
tensorflow for Windows built once and
with the release 1.1 sadly I have not
been able to build any of the other
newer builds and in Windows you're still
able to install tensorflow as a pip
package without building it yourself and
the GPU version might not work for you
and that's why we are trying to build it
ourselves
and also to get a newer version than the
actual release version but using this
method I we're going to talk about today
will work fine to build your jni
libraries for Java if you want to run
those on Windows so first off one thing
that I found is this github project made
by Microsoft which is quite nice which
will remove everything that is installed
by Visual Studio because the Visual
Studio will actually leave a lot of
stuff behind when we uninstall it so
this Visual Studio uninstaller will
actually remove all the registry keys
and all the other things in your system
that you won't be able to remove
yourself you might be required to go in
and delete the left-behind directories
and so on but no keys will actually
point anything in your system and that's
really good if you want to do with clean
reinstall of stuff on your system
and to start this off we need to
download a lot of stuff
to build tensorflow first off we need a
compiler so you can actually either use
the Visual Studio 2017 the MS build
package or the C++ build tools from 2015
so this the last one which is what I
have done because it will actually be
less impact on your system will install
less and bazel likes it more it will if
you install it in the standard
directories bazel it will pick it up just
like that
and just work with it and the other ones
you have to do a little bit more
configuration to actually get them to
run but it's possible and other things
that you need to download is Python very
important to know about both Python and
msys 2 which is another requirement is
that you need to download them for your
system so in my case I run on a 64-bit
system so I need the 64-bit version of
both Python and msys 2 or else things
will be really strange it will work
somewhat but it will break in some case
so if you are in a 64-bit system keep
those packages and if you're on a 32 bit
system then keep those packages you
really need to separate yourself you
can't have some 64 and some 32 and
you can't actually use the wrong
packages on the wrong platform so keep
that in mind and next up we need to go
to the bazel build git repository and
fetch one of the releases the newest
release that I found there was
0.5.1 but I saw also that
built 0.5.2 but I
haven't actually uploaded it to the
github so that will probably get there
soon next up you need to go and fetch
the
CUDA libraries from Nvidia so you get
the CUDA download and the cudnn
download and that package is required
you to log in and actually make an
account with them it's not really
something that will spam your email or
anything like that they just want to
keep track on which people are using
their graphic drivers for Tensorflow or
others such machine learning purposes so
that's required it's behind a login wall
but you just need to log in and tell
them why you want this and next up we
build bazel so after we have downloaded
a distribution from the releases a page we
will unpack that into a directory and
then we will run pacman to get git gcc
curl zip unzip and zlib development
we will export a bunch of variables into
our environment to set the work dir and
SH command and your Visual Studio install and
also Python so you will put those paths
into your system and depending on where
you install them then you need to change
these paths and then you run compile to
actually do the first run of the dist
package where you build bazel the simple
version and after that you run compile
with the compile and then the the actual
built version of bazel and then when you
do that you will build bazel with bazel
in order to get a fully packaged bazel
version that you can actually build
tensorflow on and if you do that it
also will make your builds faster so
it's a good way good thing to do
and next up we clone the Tensorflow
repository and you can also check
out the release 1.1 as I did and I
haven't been able to test all the other
ones fully I had a lot of work just to get
one build working and I really wanted to
get this video out so I
stopped there and after you have checked
out this version of Tensorflow you'll
enter that directory you do pacman to
install patch and protobuf stuff that
you need to work with I'm not sure that
you actually need protobuf but patch you
need a protobuf is something that when
you build the wheel package it might ask
you for that command and but that that
error should not be any any problem to
you and then you install the pip package
so here you will actually get the
protobuf the python package you have
that binary and binding and after you
have installed six numpy wheel
and protobuf and you will also set up a
lot of paths where your Python binary is
where your Python library path is and
then we have these bazel work dir Sh
visual studio and Python again so those
are the ones that you set up earlier
when you build bazel but if you run this
over and over and turn take down your
console it's good it's good to copy
paste these in and start running again
and then you point out your cuda path
and cudnn install path and
this is just you don't have to enter
those when you actually configure the
tensorflow build and in my case i also
set some paths to
the cuda bin directory so it can
find the nvidia compiler and nvcc I think
it's named and also where Python is
installed so it finds python.exe and
also where bazel is outputted in my case
it was bazel 0.5.1-1 was my directory and the output
so it actually finds the bazel exe so it
can run those so you extend your path
with a few extra and then we will
configure our tensorflow build we will
answer a lot of questions in my case i
also no for everything except for cuda
where else yes and then I just press
ENTER for all the next choices because
those will be set up for me but when I
come down to this additional compute
capability significantly increases your
build time and binary size I actually
just tell it to use 6.1 because my
graphics card is using the feature 6.1
and you can go to this developer Nvidia
CUDA GPUs to know which features your
graphics card is actually supporting so
you can put the right number in and also
get less build time and also have a
smaller binary and then I just says no
to everything else so I don't get any
extras and I set some build
parameters these are required in order
to tell tensorflow
which cpu architecture are building on
and what is the host CPU so these are
two ways to tell what kind of system are
you running on so if you set these and
then run the build pip package you will
get this build pip package executable
that you can use to build your pip
package sadly on windows some of the
linking isn't done during this process
so you get nothing in your build pip
package run files directory so in my
case I actually copy all the files using
this command copy preserved links from
bazel tensorflow in the tensorflow
directory all the way to bazel bin tensorflow tools pip package build pip package
run files that will copy all those files
over there so I have the actual files
that I need inside of my wheel package
and then I will run this command that
tensorflow has built for me in order to
package my pip package and then I install
that wheel package in order to get
tensorflow
on my system and last but not least you
have the command for building your jni
files for Java bindings so you get a jar
file that you can bind into your java
application and also an so file that
you will rename to a dll file in order
to have it include the tensorflow
so you have all the bindings to your
tensorflow installation so in my
case you run this build again the
previous options will still be required
and then tensorflow and lib tensorflow
JNI as the two packages you want to
build and that will create stuff inside
of your bazel build it tensorflow
Java directory and from there you copy
the Lib tensorflow Java file and that is
something that you bind into your
application as I said and then you copy
this Lib tensorflow JNI so to tensorflow
JNI in the directory where you have your
jni files and then you need to include
those when you run your java application
in order to get tensorflow capabilities
in your java application so this is
pretty much what I wanted to go through
today and the big things that I have
actually struggled with is that I had
different versions of both my MSYS2
and my Python installation another thing
that I had a really hard time was that I
installed Python 3.5.1 or 3.6.1 it was a three
installation but that didn't work for me
because there are a lot of places in the
tensorflow code where it actually looks
if your Python installation is three or
higher and then it will run some extra
steps but on Windows that flag isn't set
so it will not know what the major
version of your Python installation is
on Windows which means that those
function that is required for Python 3
are not run and will break your build so
when I found that out I just uninstalled
Python 3 and installed Python 2.7 point
one and everything built and worked fine
after that so that was a lot of headache
just to fight figure that one out and
during all of these installation and
using different build tools I also found
that if you install either the Visual
Studio or any of your other compilers in
a custom directory then you actually
need to
go into the tensorflow installation
and find where it looks for your VS vars
all dot bat file and edit that file so it
actually will find your bat file in your
installation because it has a lot of pre
precomputed things that it actually
looks at in the path in order to figure
out how it should find that file and if
some of those triggers aren't there it
will not find the file and so my simple
workaround was actually just go into
this configuration file and add my
direct path to this bat file because
that bat file will actually set up the
full environment for Visual Studio so
the rest of the build process will work
so that's what I've been working on for
a couple of weeks and yeah I'm really
happy that I got it built finally and I
will continue to look into this and have
another update probably the next month
if I figure out how to build newer
releases and also if I find an easier
way to both install and build tensorflow package because I think this way to
install a test flow package is still a
bit convoluted and a bit strange and
hard I've been looking into the cmake
way of building tensorflow as well but
it both have a lot of restrictions for
now you have a lot of options that are
operators that is not available in the
cmakes build and you also can't build
jar implementation through the cmake
build and I haven't actually got
cmake working on
my machine either so there's a lot of
hurdles before you actually get a good
cmake build but hopefully in the
future if cmake is a good way to
build it then I'm very hopeful that I
will create a good implementation that
you can build with cmake the really
bad thing about how you build it today
is actually bazel because bazel as I've
said before is at the moment on 0.5.1 and so much in the bazel
build is experimental and badly written
for Windows so you have to do a lot of
manual work and and figuring out to
actually run basil so hopefully when
bazel stabilizes and gets to 1.0 we will have a lot easier build
process for Tensorflow so I hope you like
this video I hope you learned something
today and I hope to hear from you in the
comments section subscribe if you
haven't done that yet and I really hope
to see you in the next video
you