search

ldeo-glaciology / LEAP-Cryo-planning

A repo for planning and tracking progress on the LEAP-Cryo project: Learning ice-sheet flow with physics-based and machine learning models.
2 stars 2 forks source link

Racheet/subglacial topography generation #29

Closed racheetmatai closed 11 months ago

racheetmatai commented 11 months ago

Added VAE and GStatSim notebooks for examples on generation of subglacial beds.

Note: 1) These notebooks assume that bed rotation, normalization and sectioning has already been done. 2) GStatSim notebook takes <40min to generate one bed realization. This can be adjusted by changing res 3) It is not advisable to run the VAE notebook without using a GPU. If you do not have access to a GPU, you can use google colab. Pangeo tensorflow gpu notebooks are currently having issues which are being worked upon.

To Do: Need to add a notebook which shows the bed rotation, normalization and sectioning.

jkingslake commented 11 months ago

Thanks for this Racheet! Looks great. The GSatsim stuff was exactly what I was thinking of. Perhaps a little disappointing that it takes 10's on mins to generate one realization. Does this present a significant barrier to generating hundreds or thousands of grids? Or can res be increased to make it practical computationally, but still with a useful resolution?

racheetmatai commented 11 months ago

Thanks for this Racheet! Looks great. The GSatsim stuff was exactly what I was thinking of. Perhaps a little disappointing that it takes 10's on mins to generate one realization. Does this present a significant barrier to generating hundreds or thousands of grids? Or can res be increased to make it practical computationally, but still with a useful resolution?

We can play with res to decrease the time required to generate a realization. It results in a variogram which is a little bit worse but in most realizations not by a 'lot' (ultimately depends on what is acceptable). I am playing with this a bit more to see if the generation time/quality can be improved and will update this notebook if required. I was also looking at parallel generations ,which is what the last function was trying to do, but that only works on google colab for unknown reasons.