[Suggestion] Do you think this can be done using tensorflow?

[capaulson]

[Suggestion] Do you think this can be done using tensorflow?

[capaulson]

I like the idea of parallel computing in general. I'm not sure if tensorflow is the right solution here, or if something like MPI is sufficient. The step that takes time is evaluating hyperparameters in a stochastic optimiser. It should be fairly easy enable parallel/clustering on that. In fact, I had it running at one point on a 32 core server in a parallel mode. What specific features do you think tensorflow brings to this toolkit?

[pangard]

I guess using tensorflow or theano is not convenient until any of those projects becomes more widespread, if they ever do. I would like to write code that uses all the cores/gpus/machines available using just one tool, rather that tensorflow or theano for evaluating the function and mpi for krigging, above all because they are nested.
Thanks anyone

[capaulson]

If you want to look into this further, take a look at where the hyper parameter optimisation is done in the code. I use the Inspyred library for the multi-objective optimisation, which does have some built in parallelisation: http://pythonhosted.org/inspyred/reference.html?highlight=parallel#inspyred.ec.evaluators.parallel_evaluation_pp.

[pangard]

Thanks!

[capaulson]

I'm going to close this for now. Thanks for the feedback!