[Bioc-devel] Tensorflow support for bioconductor packages

FWIW the tensorflow authors didn't opt for automatic lazy installation:

  > run_example("hello.R")

  Error: Installation of TensorFlow not found.

  Python environments searched for 'tensorflow' package:
   /usr/bin/python2.7
   /usr/bin/python3.5

  You can install TensorFlow using the install_tensorflow() function.

Would be interesting to know why.

install_tensorflow() has various arguments and the chances that it
will just work and do the right thing when called with no argument
are low. There is also this 'restart_session' argument that is TRUE
by default and will only work within RStudio. This suggests that after
successful completion R needs to be restarted before the tensorflow
package becomes operational. I didn't test that but that's something
you might want to investigate before opting for lazy installation.

Also it might help to look at how the handful of CRAN packages that
depend on tensorflow handle this. These packages are listed in the
reverse dependencies section of the tensorflow landing page:

  https://cran.r-project.org/web/packages/tensorflow/index.html

We'll install the tensorflow Python module on the build machines when
you submit your package.

Cheers,
H.



On 03/29/2018 10:08 AM, Michael Lawrence wrote:

The problem with requiring explicit tensor flow installation is that
it is tantamount to a system dependency in many ways, and those are
annoying. Herve points out the problems with installing at load time.
My suggestion was to install the package the first time someone tries
to e.g. load an R matrix into a tensor. That way, you know that
examples and vignettes will always just work (if the installation
works) on any build machine, without any admin intervention. And, the
last thing a user wants when running an example is an error, even if
that error is easily remedied. One downside is that the user could
have just forgotten to point the package to a system installation of
tensorflow, in which case they will be cursing themselves for
forgetting while watching the installation process. You could check
for interactive() and then prompt the user to avoid that case.

On Thu, Mar 29, 2018 at 9:44 AM, Kieran Campbell
<kieranrcampbell at gmail.com> wrote:

Hi Herv?, Michael,

Thanks for your feedback. I will add in the reticulate check to ensure
tensorflow is installed prior to running and appropriate sections in
the vignettes. We have one package essentially ready for submission to
bioc, so is the best route forward to submit now or wait until
tensorflow is installed on the build servers?

Many thanks

Kieran


On 28 March 2018 at 15:10, Herv? Pag?s <hpages at fredhutch.org> wrote:

On 03/28/2018 02:41 PM, Herv? Pag?s wrote:

Hi Kieran,

Note that you can execute arbitrary code at load time by defining
an .onLoad() hook in your package. So you *could* put something
like this in your package:

    .onUnload <- function(libpath)
    {
      if (!reticulate::py_module_available("tensorflow"))
          tensorflow::install_tensorflow()
    }

should be .onLoad() in the above code

more below...

However, having things being automatically downloaded/installed
on the user machine at package load-time is not a good idea. There
are just too many things that can go wrong.

For example, I just tried to run tensorflow::install_tensorflow()
on my laptop (Ubuntu 16.04) and was successful only after the 3rd
attempt (I had to make some changes/adjustments to my system between
each attempt). And Debian Linux is probably the easiest target!

Also note that install.packages() tries to load the package at the
end of the installation when installing from source so if the
.onUnload() hook fails, install.packages() considers that

   ^^^^^^^^^^^
    .onLoad()

same here, sorry

H.

the installation of the package failed and it removes it.

Finally note that this installation needs to download hundreds of
Mb of Python stuff.

So this is probably the reasons why the authors of the tensorflow
CRAN package chose to separate installation of the tensorflow Python
module from the installation of the package itself. There are plenty
of good reasons for doing that.

What I would suggest instead is that you start your vignette with a
note reminding the user to run tensorflow::install_tensorflow() if
s/he didn't already do it. As a side note: I couldn't find a way to
programmatically figure out whether the tensorflow Python module is
already installed in the man page for tensorflow::install_tensorflow(),
I had to dig in the source code of the unit tests to find
reticulate::py_module_available("tensorflow")).

In addition, you could also start each of your functions that rely on
the tensorflow Python module with a check to see whether the module is
available, and fail gracefully (with an informative error message) if
it's not.

We'll figure out a way to install the tensorflow Python module on our
build machines.

Hope this helps,
H.


On 03/28/2018 09:23 AM, Kieran Campbell wrote:

Hi all,

Rstudio have released the Tensorflow package for R -


https://urldefense.proofpoint.com/v2/url?u=https-3A__tensorflow.rstudio.com_tensorflow_&d=DwICAg&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=w2p-VnxwECq9u90RNv_B6yCOpXxDkcIPAjcgcpbEeBE&s=AchAIWmKzcnyw9VXJ7eH5M4dqnTAS0SACVMigCPusHk&e=
- and we have started
incorporating it into some of our genomics packages for the heavy
numerical computation.

We would ideally like these to be submitted to Bioconductor, but
there's a custom line required for Tensorflow installation in that
after calling

install.packages("tensorflow")

then Tensorflow must be installed via

tensorflow::install_tensorflow()

which would break package testing if tensorflow was simply imported
into the R package and wasn't already installed. Is there any way to
customise a package installation within Bioconductor to trigger the
tensorflow::install_tensorflow() ?

As more people use tensorflow / deep learning in genomics I can see
this being a problem so it would be good to have a solution in place.

Many thanks,

Kieran Campbell

--
Herv? Pag?s

Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024

E-mail: hpages at fredhutch.org
Phone:  (206) 667-5791
Fax:    (206) 667-1319

--
Herv? Pag?s

Program in Computational Biology
Division of Public Health Sciences
Fred Hutchinson Cancer Research Center
1100 Fairview Ave. N, M1-B514
P.O. Box 19024
Seattle, WA 98109-1024

E-mail: hpages at fredhutch.org
Phone:  (206) 667-5791
Fax:    (206) 667-1319