keras (tensorflow) + sklearn (GridSearchCV with n_jobs>1) => CUDA_ERROR_NOT_INITIALIZED

[juan...@gmail.com]

Hi everybody,

I follow a simple example from https://machinelearningmastery.com/grid-search-hyperparameters-deep-learning-models-python-keras/ with pima diabetes dataset.

Combine sklearn with GridSearchCV to get the best parameters model is usual and use n_jobs>1 to parallelize jobs, too.

When I use CPU with Tensorflow, there is no problem (only use cores of computer). When I use GPU with CUDA with n_jobs>1 (parallel jobs) it does not work. CUDA error is showed.

Context:

• Ubuntu 16.04 (64 bits)
• Keras 2.0.8 (latest)
• Tensorflow-gpu 1.3.0 (latest)
• sklearn 0.19.0 (latest)
• GPU: NVidia 1080 GTX, CUDA 8.0, cuDNN 6.0 (lastest stable and recomanded to work with Tensorflow)

Any advise?

Regards,

[kadir...@gmail.com]

Can you run two jupyter kernels with tensorflow? I cannot. I always shutdown the unused ones.

I don't have parallel GPUs configured. 

So it seems normal that I cannot run GPU GridSearchCV with n_jobs>1.

Maybe that's your case too.