Re: Teething problems with new GNU Radio 3.9 installation

[Brendan Horsfield]

Hi David, 

My current configuration is as follows:

• TensorFlow 2.4.1
• CUDA 11.2

I am also running a second version of TensorFlow inside a Docker image that I downloaded from dockerhub:

• https://hub.docker.com/r/tensorflow/tensorflow/tags?page=1&ordering=last_updated

My preferred image is "tensorflow:latest-gpu", but you will see that there are many, many others to choose from, depending on what combination of libraries you want.

The Docker version of TensorFlow is usually much easier to install, as it comes equipped with all of the necessary dependencies (including CUDA drivers).  The only downside is that the image I am using ("latest-gpu") doesn't support graphical tools like Matplotlib, which means you can only print results to the console as simple text.  This makes it great for model training, but not so good for model testing or inference, where you often want to visualise your model's performance with colourful plots. 

Regards,

Brendan.

On Thu, 22 Apr 2021, 20:24 David Scaperoth, <david.scaperoth@gmail.com> wrote:

Brendan,

Out of curiosity, what versions of tensorflow and cuda are you using?

David

On Thu, Apr 22, 2021, 12:50 AM Brendan Horsfield <brendan.horsfield@vectalabs.com> wrote:

Hi Ryan,

I have some good news to report.  

While I was waiting for your reply (the curse of living Down Under), I tried installing the UHD / GNU Radio combo you mentioned in your last email, namely:

• UHD 3.15 (with Python bindings)
• GR 3.8.3 (actually maint-3.8, which installed as v3.8.2.0-112-ge20ffa3c)

Both were installed from source, without using a conda environment.  I fired up gnuradio-companion at the end, and bingo, I'm back in business!  No errors, no warnings, old flowgraphs still working. 

Thanks for your help in getting me back on track -- I was starting to lose hope!

Regards,

Brendan.

 

On Thu, Apr 22, 2021 at 2:05 PM Ryan Volz <ryan.volz@gmail.com> wrote:

Hi Brendan,

On April 21, 2021 7:14:41 PM EDT, Brendan Horsfield <brendan.horsfield@vectalabs.com> wrote:
>Hi Ryan,
>
>Thanks, that sounds like a MUCH easier way of doing things!
>
>A couple of questions:
>
>- I don't have much experience with virtual environments.  If I install
> one of your packages into a conda environment, will I still be able to
>   access applications & libraries that were installed outside of conda
>   (specifically Python, TensorFlow, CUDA & Docker)?  Or will I have to
> reinstall those resources inside the same conda environment as the GNU
>   Radio package?

Everything would have to be installed into the conda environment, assuming you want to use them in the same Python process. (You could probably get tricky with PYTHONPATH or something, but I sure wouldn't recommend it.) Tensorflow and CUDA stuff could be a little tough, because currently that requires mixing the conda-forge channel (community-supported, where uhd and gnuradio are) and the defaults channel (provided by Anaconda Inc., where tensorflow-gpu is available). Depending on how you install conda, you'll want to make sure you also have the defaults channel available at a lower priority than conda-forge:

conda config --append channels defaults

Then

conda install tensorflow-gpu

should work to get you everything (assuming you have the NVIDIA driver installed normally). It would be good to double-check my instructions by consulting some guides, e.g.:

https://docs.anaconda.com/anaconda/user-guide/tasks/tensorflow/

Assuming you mentioned Docker only as a way to get tensorflow set up with GPU support, then you shouldn't need it. Hopefully tensorflow-gpu will make it to conda-forge soon (the infrastructure needed to do the build has only recently been set up, and pytorch-gpu was added just today), and then installation will get even easier.

>  - Slightly off-topic, but still related:  As mentioned in my original
>post, I have already installed UHD & GNU Radio from source on my
>system.
>This was not done inside a virtual environment.  If I now want to
>uninstall
>those items, is it as simple as going into their respective "build"
>folders
>(e.g. ~/workarea-gnuradio/gnuradio/build) and running "sudo make
>uninstall"
>   and "sudo make clean"?  Or is there more to it than that?

That should clean up all of the installed files (the make uninstall part, clean would not be necessary), but configuration files and cache files would still remain. It's almost certainly not necessary to uninstall your other build while trying/using the conda one as I've never experienced interference of that flavor. What could cause trouble are the config/cache files that would remain either way. But now you know about that possibility and know to clean those out if there are any problems. Just make sure to activate your conda environment and you'll be good to go.

>
>Thanks,
>Brendan.
>

Cheers,
Ryan