CUDA® Toolkit 9.0 + cuDNN v7.0.5

[obsydion]

Hi there,
As the subject suggests I have been trying to get CSBDeep GPU acceleration to work with CUDA Toolkit 9.0 + cuDNN v7.0.5 with no avail. The reason why I didn't use Toolkit 8.0 as suggested is because I installed 9.0 to run GPU acceleration for our previously installed SVI Huygens deconvolution package.

My question is whether CSBDeep will work with Toolkit 9.0 at all? Similarly, I had to download cuDNN v7.0.5 to match Toolkit 9.0 instead of the suggested cuDNN v6.

Thanks in advance.

[tboo]

Hi - The problem is likely not CSBdeep but Tensorflow. I believe that Tensorflow only runs on cuDNN 6.0 (unless you compile it from source). You can have both cuDNN 6.0 and 7.0 run in parallel though. I have a similar setup.

EDIT: Make sure this works: https://www.tensorflow.org/install/install_linux#run_a_short_tensorflow_program

[HedgehogCode]

Hi,
As @tboo pointed out, TensorFlow requires CUDA 8 and cuDNN 6. They anticipate releasing TensorFlow 1.5 (future version) with CUDA 9 and cuDNN 7 (See the last point of the release notes of 1.4). We have to wait for this release, then update imagej-tensorflow and then we can run CSBDeep with CUDA 9 and cuDNN 7.

Also pointed out by @tboo: You can install multiple versions of CUDA and cuDNN.

[royerloic]

Guys & Gals, I would suggest, maybe, to have one version of CSBDeep per CUDA version… it might be a problem for some to update version because of other dependencies. At least, we should consider the ramifications of a CUDA update… and whether we give the possibility to ’stay behind’ for some time… (one key thing to investigate is whether TF will be backwards compatible… which I doubt) Cheers, Loic

…

On 3. Jan 2018, at 04:48, Benjamin Wilhelm ***@***.***> wrote: Hi, As @tboo <https://github.com/tboo> pointed out, TensorFlow requires CUDA 8 and cuDNN 6. They anticipate releasing TensorFlow 1.5 (future version) with CUDA 9 and cuDNN 7 (See the last point of the release notes of 1.4 <https://github.com/tensorflow/tensorflow/releases/tag/v1.4.0>). We have to wait for this release, then update imagej-tensorflow <https://github.com/imagej/imagej-tensorflow> and then we can run CSBDeep with CUDA 9 and cuDNN 7. Also pointed out by @tboo <https://github.com/tboo>: You can install multiple versions of CUDA and cuDNN. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AByMks4bXRP_TKsUxqHiLzp7hYU6HFMFks5tG3cVgaJpZM4RRNTB>.