search

ChrisCummins / paper-end2end-dl

📝 "End-to-end Deep Learning of Optimization Heuristics" (🥇 PACT'17 Best Paper)
https://chriscummins.cc/pub/2017-pact.pdf
GNU General Public License v3.0
72 stars 21 forks source link

Replicating experiments with CUDA version > 8.0 #2

Open zehanort opened 3 years ago

zehanort commented 3 years ago

Hello Chris!

I am trying to replicate your first case study regarding OpenCL heterogeneous mapping and I am trying to use CUDA. When I import tensorflow from the Python 3 CLI on my machine, everything works as expected, loading my version of the CUDA library (libcudart.so.11.2.72), but when I import tensorflow from inside of one of your Jupyter notebooks (which are launched from inside the virtualenv you create using make run) I get an ImportError that the file libcudart.so.8.0 could not be found.

To my understanding, the tensorflow version used in your virtualenv needs CUDA version 8.0 explicitly ( ? ). Is there any workaround for this to work with my CUDA version? Tried a workaround with symlinks, but it seems that there are deeper dependencies and it still does not work.

Thanks a lot in advance!

ChrisCummins commented 3 years ago

Hi Sotiris,

Apologies for my slow response! I haven't touched this codebase in 4 years. If I were you I would try removing the dated tensorflow in the virtualenv and just use the working tensorflow version on your system, though there may be API changes that causes the tensorflow code to break. Hope that helps

Cheers, Chris

xshaun commented 3 years ago

Hi,

One of the useful solution is to use docker or singularity container to run this code. docker://tensorflow/tensorflow:1.14.0-gpu-py3 is available to execute the device mapping code. hope it's helpful.

Cheers,

ChrisCummins commented 3 years ago

Thanks @xshaun that's a great suggestion!