gempy
gempy copied to clipboard
Port to TensorFlow
Is your feature request related to a problem? Please describe. At the moment Gempy is developed on Theano to do kriging to generate the scaler field. As Theano has been stopped developing since 2017, and Tensorflow has become a more popular and easier for debugging tool for auto-differentiation and been actively supported, there is a good reason to switch the core interpolation part of Gempy to Tensorflow.
Describe the solution you'd like Implement the co-Kriging part used to interpolate the scaler filed in Tensorlfow and integrated with Gempy main software.
Describe alternatives you've considered Jax and Pytorch are also possible to be tested
Additional context the second derivative of the posterior
None of this will be merged for 2.1. Highest priority for 2.2 is having this working for a scalar field. More complex models will come in later versions.
What work is remaining here? @GeorgeLiang3 , could you edit the original issue description? Right now it has only the default text from the issue template.
Thanks for editing the description, @GeorgeLiang3 .
Why did you mention the second derivative of the posterior? Is this needed for co-Kriging? Is it unclear how to obtain it with TensorFlow?
I posted this in the software underground slack but there are other frameworks to consider including kepos (https://www.kernel-operations.io/keops/index.html) that seem to scale better with larger datasets. I think for now continue as planned, and by refactoring the code to TensorFlow it may make some of the underlying code clearer and easier to abstract to different implementations
hi,@agzimmerman, Is there a plan to migrate to TensorFlow or Pytorch?
@istrous a implementation using aesara is currently under development and will be released shortly. This release should also solve many other dependency issues.
Cheers Alex
@agzimmerman ,That's great,Thanks
Hey @istrous I just noticed that you have accidentally been tagging me instead of @AlexanderJuestel :)
@istrous see here: https://github.com/GeorgeLiang3/GemPyTF
We ported it to PyTorch one week after finish the TF. TF is deprecated before being born.
Congrats!