Nearest neighbor interpolator #12
Conversation
|
Nearest neighbor |
|
Inverse distance |
earth2grid/_regrid.py
Outdated
| return output | ||
|
|
||
|
|
||
| class S2LinearBarycentricInterpolator(torch.nn.Module): |
There was a problem hiding this comment.
Would be good to add a test that saves a checkpoint of an initialized instance of this and loads that back in. Would be very useful to avoid repeated calls to spatial.KDTree especially in ETL jobs where a regridding script might be invoked multiple times on different inputs.
There was a problem hiding this comment.
I'll make an issue for this. I think it may require some thought and potentially backwards breaking changes to apply this consistently across the suite of re-gridders.
I'm not sure the most elegant way to allow ser/deser of "buffer-only" modules like this. Do you know of an example for this e.g. with normalization.
There was a problem hiding this comment.
I think we don't necessarily need a proper serialization/deserialization pipeline here right? It might be sufficient to create a classmethod that takes in a state dict and returns an initialized instance from that?
There was a problem hiding this comment.
I refactored this so that the S2NearestNeighborInterpolator now returns a serializeable Regridder object. Let me know what you think.
tests/test_regrid.py
Outdated
| n = 10000 | ||
| torch.manual_seed(0) | ||
| lon = torch.rand(n) * 360 | ||
| lat = torch.rand(n) * 180 - 90 |
There was a problem hiding this comment.
| lat = torch.rand(n) * 180 - 90 | |
| lat = torch.asin(2*torch.rand(n)-1) * 180. / np.pi |
There was a problem hiding this comment.
what is the intent here?
There was a problem hiding this comment.
This isn't a super important change, but thought it would be good to remove the polar bias when sampling on the sphere in these tests.
No description provided.