Is it possible to load the saved NeuralPosterior objects from C++?

[hanyk]

Thanks for your amazing work! Is it possible to load the saved NeuralPosterior objects from C++ to allow a more efficient posterior.sample?

[janfb]

thanks!

No, I don't think we have this option. Have you tried increasing sampling speed using the slice_np_vectorized option, or slice_np with num_workers>1?

[hanyk]

Thanks for your suggestion!I will try that later. I am not sure about the structure of NeuralPosterior objects . Is it a kind of PyTorch Model? If so, it may be converted to torch script and then loaded in C++ with libtorch.

[michaeldeistler]

Hi! The NeuralPosterior is not a pytorch module. The returned nflows network is a torch.nn.Module though. I.e.

net = inference.append_simulations(theta, x).train()
--> net is a nn.Module

Hope that helps! Michale

[hanyk]

Thank you very much!

[michaeldeistler]

Closing this for now, feel free to re-open if new questions come up (and feel free to write here if you managed to sample in c++ ;) )