Enh/manage groups #66
Conversation
When open_ncml is called with group="*", every group will be read and they will be flatten in the resulting dataset. If names are conflicting, the dimensions and varaibles names are appended with __n where n is the number of existing simlar names.
|
Converting PR to draft as I discovered flatten is not working properly at the moment. |
xncml/parser.py
Outdated
| ' to preserve the type.' | ||
| ) | ||
| return xr.Variable(data=None, dims=()) | ||
| default_value = nctype(var_type)() |
There was a problem hiding this comment.
I changed (again) the behavior of empty scalar parsing here.
For context, here the ncml describe a scalar with a certain type but without providing any value. We would ideally like to create a placeholder for a scalar variable with a numpy type.
But numpy doesn't allow this for scalar, only with an ndarray we can create a typed empty array of a certain shape.
I first though it would be better to loose the type and create a empty scalar with the value None but not having this type can mess with subsequent processing.
Now I think it's better to preserve the dtype and fill the scalar with the default value of this dtype.
I would appreciate comments/suggestions here.
- improved performances by parsing ncml reprsentation only once - fixed issues of rewritting content when read multiple times
Co-authored-by: David Huard <huard.david@ouranos.ca>
This PR fixes #64
groupargument toopen_ncml'/'is read.The above is similar to xarray's
open_dataset.In addition, using
group='*', it flattens every group into a single datasets (somewhat preserving the current behavior).The names conflicts between groups are solved by appending an incrementing
__nto the variable names, where n is a number.Plus, an attribute
group_pathis added to variables in order to retrieve their original path once they have been flatten.This can be useful to recreate the original structure.