18:37 < ShikharJ> zoq: I'm having trouble setting up tensorflow on my system for computing the upsampling, what method would you suggest for the forward pass test?

18:46 < rcurtin> ShikharJ: maybe you could use TensorFlow via docker?

18:47 < rcurtin> that's how I will typically use it... e.g. 'docker run -it --rm -p 8888:8888 jupyter/tensorflow-notebook'

18:47 < rcurtin> (if you want a notebook interface)

19:08 < ShikharJ> Thanks rcurtin. I'll be sure to give that a try :)

19:09 < ShikharJ> I really made a mistake updating to ubuntu 18.04, a lot of drivers (Nvidia included) are giving me pain, and there's no bumblebee support :(

19:13 < rcurtin> ack, that is frustrating

19:14 < rcurtin> I maintain an internal machine learning service inside of Symantec that's build in a container on TensorFlow

19:14 < rcurtin> and I recently rebuilt that against 18.04... it was a little bit of a disaster

19:14 < rcurtin> I needed 18.04 so that the nvidia CUDA libraries were available in apt (instead of needing to be installed by hand)

19:14 < rcurtin> but making all the versions of everything match correctly (including the underlying system that I was running nvidia-docker on to make it all work) was a bit of a nightmare

19:15 < rcurtin> to make it worse, it's a container that runs systemd internally to manage a number of processes, so there are also version issues and other trickiness there too :)

19:24 < ShikharJ> I feel you, totally, I guess this is why people aren't too excited about newer OS releases, especially when it comes to Ubuntu.

19:26 < ShikharJ> Why jump from something that's running fine, to something that might not even work. I realized it only too late.

19:27 < ShikharJ> Actually, the hardware in system is pretty new (early 2017), and hence the only Linux distro that worked ok with that was 17.10. I thought 18.04 would be better altogether.

19:28 < rcurtin> the nvidia drivers are always hard to work with, it seems

19:28 < rcurtin> I don't know too much about bumblebee, but sometimes the apt versions of drivers are too old (or not compatible with the CUDA library versions, etc.)

19:29 < rcurtin> our group's sysadmin insists on installing the drivers directly using whatever .run file nvidia gives on their website

19:29 < rcurtin> but this means every time there is a kernel upgrade, suddenly the drivers are out of date and nothing works, so he has to go fix it by hand, instead of having apt + DKMS auto-rebuild it...

19:30 < ShikharJ> True, you just cannot expect them to work out of the box. Bumblebee (even the old apts) works up fine untill 17.10. But switch to 18.04 and boom! There goes your happiness.

19:31 < ShikharJ> It's going to be a while before people are really going to shift to 18.04 I guess. And I can't even fathom of going through the pain of making a clean install.

19:39 < rcurtin> true

19:39 < rcurtin> I typically just stick with Debian unstable or sid, which is more of a 'rolling release' so there are few 'big upgrades' for me

19:41 < rcurtin> I have a Fedora system I use sometimes, but Fedora moves so fast it seems like every time I get a shell on it, it's out of date...

19:56 < zoq> ShikharJ: Totally feel you, usually, I recommend to use FreeBSD but it's definitely not pain-free, but I agree Docker is a good solution.

20:36 < travis-ci> ShikharJ/mlpack#124 (Deconv - 29bcd76 : Shikhar Jaiswal): The build has errored.

20:36 < travis-ci> Change view : https://github.com/ShikharJ/mlpack/compare/5aa50179465f...29bcd76d8de0

20:36 < travis-ci> Build details : https://travis-ci.org/ShikharJ/mlpack/builds/376524012