Could not create cudnn handle: CUDNN_STATUS_NOT_INITIALIZED

[atyakhtmpg]

Hello, we are having a problem running ResMico: the run fails with the log file including the following:
... tensorflow/stream_executor/cuda/cuda_dnn.cc:373] Could not create cudnn handle: CUDNN_STATUS_NOT_INITIALIZED 2023-09-04 19:28:54.754306: E tensorflow/stream_executor/cuda/cuda_dnn.cc:382] Possibly insufficient driver version: 515.57.0 2023-09-04 19:28:54.755139: W tensorflow/core/framework/op_kernel.cc:1745] OP_REQUIRES failed at conv_ops.cc:1120 : UNIMPLEMENTED: DNN library is not found.

Could there be a conflict between versions of various software packages/modules?

The Conda environment used for the run was based on the following yaml:
`channels:

• conda-forge
• bioconda
  dependencies:
• pigz
• python=3.9
• conda-forge::cudnn
• pip
• pip:
  • numpy==1.23.3
  • scikit-learn==1.1.1
  • pysam==0.19.1
  • git+https://github.com/leylabmpi/ResMiCo.git`

[helgako]

Hi,

It is indeed hard to handle all the dependencies between packages. We specify the required versions in the https://github.com/leylabmpi/resmico/blob/master/environment.yml,
in particular, tensorflow=2.8.1,
and recommend the following steps to install the resmico https://github.com/leylabmpi/resmico/tree/master#from-source-using-pip

The reason for an error could also be that GPU drivers are not up to date and not compatible with the version of tensorflow. I would refer to the official tensorfow installation guide to handle this issue https://www.tensorflow.org/install/pip