Is there a reason for why DataArray.swap_dims() cannot be done in place like Dataset.swap_dims?

Problem description

This is a problem if I want to swap_dims in DataArray Accessor.

Code Sample

This is what I'm forced to do:

import xarray as xr
import numpy as np
@xr.register_dataarray_accessor('testing')
class TestAccessor(object):
    def __init__(self, xarray_obj):
        self._obj = xarray_obj
    def the_problem(self):
        self._obj = self._obj.swap_dims({'x':'x2'})
        print(self._obj)
        
arr = np.random.rand(4,3,2)
cube = xr.DataArray(arr, dims=['ya', 'x', 'y'], coords={'y':[1,3], 'ya':[1,2,3,6], 'x':[1,2,5]})
cube.coords['x2'] = ('x', [1,2,3])
cube.testing.the_problem()
print(cube)

this prints:

<xarray.DataArray (ya: 4, x2: 3, y: 2)>
array([[[ 0.659583,  0.167555],
        [ 0.357974,  0.46081 ],
        [ 0.85115 ,  0.845257]],

       [[ 0.280308,  0.777399],
        [ 0.512527,  0.542036],
        [ 0.838603,  0.799414]],

       [[ 0.572031,  0.350464],
        [ 0.205219,  0.812232],
        [ 0.687778,  0.984928]],

       [[ 0.803385,  0.63981 ],
        [ 0.089909,  0.499857],
        [ 0.25266 ,  0.967909]]])
Coordinates:
  * y        (y) int32 1 3
  * ya       (ya) int32 1 2 3 6
    x        (x2) int32 1 2 5
  * x2       (x2) int32 1 2 3
<xarray.DataArray (ya: 4, x: 3, y: 2)>
array([[[ 0.659583,  0.167555],
        [ 0.357974,  0.46081 ],
        [ 0.85115 ,  0.845257]],

       [[ 0.280308,  0.777399],
        [ 0.512527,  0.542036],
        [ 0.838603,  0.799414]],

       [[ 0.572031,  0.350464],
        [ 0.205219,  0.812232],
        [ 0.687778,  0.984928]],

       [[ 0.803385,  0.63981 ],
        [ 0.089909,  0.499857],
        [ 0.25266 ,  0.967909]]])
Coordinates:
  * y        (y) int32 1 3
  * ya       (ya) int32 1 2 3 6
  * x        (x) int32 1 2 5
    x2       (x) int32 1 2 3

where the two xarrays are clearly different.

I would want to do:

import xarray as xr
import numpy as np
@xr.register_dataarray_accessor('testing')
class TestAccessor(object):
    def __init__(self, xarray_obj):
        self._obj = xarray_obj
    def the_problem(self):
        self._obj.swap_dims({'x':'x2'}, inplace = True)
        print(self._obj)
        
arr = np.random.rand(4,3,2)
cube = xr.DataArray(arr, dims=['ya', 'x', 'y'], coords={'y':[1,3], 'ya':[1,2,3,6], 'x':[1,2,5]})
cube.coords['x2'] = ('x', [1,2,3])
cube.testing.the_problem()
print(cube)

this would keep the two xarrays the same, as they should be:

<xarray.DataArray (ya: 4, x2: 3, y: 2)>
array([[[ 0.659583,  0.167555],
        [ 0.357974,  0.46081 ],
        [ 0.85115 ,  0.845257]],

       [[ 0.280308,  0.777399],
        [ 0.512527,  0.542036],
        [ 0.838603,  0.799414]],

       [[ 0.572031,  0.350464],
        [ 0.205219,  0.812232],
        [ 0.687778,  0.984928]],

       [[ 0.803385,  0.63981 ],
        [ 0.089909,  0.499857],
        [ 0.25266 ,  0.967909]]])
Coordinates:
  * y        (y) int32 1 3
  * ya       (ya) int32 1 2 3 6
    x        (x2) int32 1 2 5
  * x2       (x2) int32 1 2 3
<xarray.DataArray (ya: 4, x: 3, y: 2)>
array([[[ 0.659583,  0.167555],
        [ 0.357974,  0.46081 ],
        [ 0.85115 ,  0.845257]],

       [[ 0.280308,  0.777399],
        [ 0.512527,  0.542036],
        [ 0.838603,  0.799414]],

       [[ 0.572031,  0.350464],
        [ 0.205219,  0.812232],
        [ 0.687778,  0.984928]],

       [[ 0.803385,  0.63981 ],
        [ 0.089909,  0.499857],
        [ 0.25266 ,  0.967909]]])
Coordinates:
  * y        (y) int32 1 3
  * ya       (ya) int32 1 2 3 6
    x        (x2) int32 1 2 5
  * x2       (x2) int32 1 2 3

I have version 0.10.0, the newest on conda-forge.

swap_dims is not an in-place operation http://xarray.pydata.org/en/stable/generated/xarray.Dataset.swap_dims.html It returns a new dataset with swapped dims. If you want to do it in place, there is a keyword argument for that.

…

Sent from my iPhone

On Dec 1, 2017, at 7:37 AM, Leevi Annala ***@***.***> wrote: Problem description This is a problem if I want to swap_dims in DataArray Accessor. Code Sample This is what I'm forced to do: import xarray as xr import numpy as np @xr.register_dataarray_accessor('testing') class TestAccessor(object): def __init__(self, xarray_obj): self._obj = xarray_obj def the_problem(self): self._obj = self._obj.swap_dims({'x':'x2'}) print(self._obj) arr = np.random.rand(4,3,2) cube = xr.DataArray(arr, dims=['ya', 'x', 'y'], coords={'y':[1,3], 'ya':[1,2,3,6], 'x':[1,2,5]}) cube.coords['x2'] = ('x', [1,2,3]) cube.swap_dims({'x':'x2'}) cube.testing.the_problem() print(cube) this prints: <xarray.DataArray (ya: 4, x2: 3, y: 2)> array([[[ 0.659583, 0.167555], [ 0.357974, 0.46081 ], [ 0.85115 , 0.845257]], [[ 0.280308, 0.777399], [ 0.512527, 0.542036], [ 0.838603, 0.799414]], [[ 0.572031, 0.350464], [ 0.205219, 0.812232], [ 0.687778, 0.984928]], [[ 0.803385, 0.63981 ], [ 0.089909, 0.499857], [ 0.25266 , 0.967909]]]) Coordinates: * y (y) int32 1 3 * ya (ya) int32 1 2 3 6 x (x2) int32 1 2 5 * x2 (x2) int32 1 2 3 <xarray.DataArray (ya: 4, x: 3, y: 2)> array([[[ 0.659583, 0.167555], [ 0.357974, 0.46081 ], [ 0.85115 , 0.845257]], [[ 0.280308, 0.777399], [ 0.512527, 0.542036], [ 0.838603, 0.799414]], [[ 0.572031, 0.350464], [ 0.205219, 0.812232], [ 0.687778, 0.984928]], [[ 0.803385, 0.63981 ], [ 0.089909, 0.499857], [ 0.25266 , 0.967909]]]) Coordinates: * y (y) int32 1 3 * ya (ya) int32 1 2 3 6 * x (x) int32 1 2 5 x2 (x) int32 1 2 3 where the two xarrays are clearly different. I would want to do: import xarray as xr import numpy as np @xr.register_dataarray_accessor('testing') class TestAccessor(object): def __init__(self, xarray_obj): self._obj = xarray_obj def the_problem(self): self._obj.swap_dims({'x':'x2'}, inplace = True) print(self._obj) arr = np.random.rand(4,3,2) cube = xr.DataArray(arr, dims=['ya', 'x', 'y'], coords={'y':[1,3], 'ya':[1,2,3,6], 'x':[1,2,5]}) cube.coords['x2'] = ('x', [1,2,3]) cube.swap_dims({'x':'x2'}) cube.testing.the_problem() print(cube) this would keep the two xarrays the same, as they should be: <xarray.DataArray (ya: 4, x2: 3, y: 2)> array([[[ 0.659583, 0.167555], [ 0.357974, 0.46081 ], [ 0.85115 , 0.845257]], [[ 0.280308, 0.777399], [ 0.512527, 0.542036], [ 0.838603, 0.799414]], [[ 0.572031, 0.350464], [ 0.205219, 0.812232], [ 0.687778, 0.984928]], [[ 0.803385, 0.63981 ], [ 0.089909, 0.499857], [ 0.25266 , 0.967909]]]) Coordinates: * y (y) int32 1 3 * ya (ya) int32 1 2 3 6 x (x2) int32 1 2 5 * x2 (x2) int32 1 2 3 <xarray.DataArray (ya: 4, x: 3, y: 2)> array([[[ 0.659583, 0.167555], [ 0.357974, 0.46081 ], [ 0.85115 , 0.845257]], [[ 0.280308, 0.777399], [ 0.512527, 0.542036], [ 0.838603, 0.799414]], [[ 0.572031, 0.350464], [ 0.205219, 0.812232], [ 0.687778, 0.984928]], [[ 0.803385, 0.63981 ], [ 0.089909, 0.499857], [ 0.25266 , 0.967909]]]) Coordinates: * y (y) int32 1 3 * ya (ya) int32 1 2 3 6 x (x) int32 1 2 5 * x2 (x) int32 1 2 3 I have version 0.10.0, the newest on conda-forge. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

@rabernat

@rabernat, please note that i'm working with dataarrays. There is no inplace keyword.

Ah right. So you would do

cube = cube.swap_dims({'x':'x2'})

Seems like I forgot a confusing line in my example. For example in dataarray_accessor, we have a need for in-place swap.

cube = xr.DataArray(arr, dims=['ya', 'x', 'y'], coords={'y':[1,3], 'ya':[1,2,3,6], 'x':[1,2,5]}) #First dataarray object. cube now referes to this.
cube.coords['x2'] = ('x', [1,2,3])
cube = cube.swap_dims({'x':'x2'}) #Second dataarray object. cube now refers to this and first dataarray object is now forgotten

vs:

cube = xr.DataArray(arr, dims=['ya', 'x', 'y'], coords={'y':[1,3], 'ya':[1,2,3,6], 'x':[1,2,5]}) #First dataarray object. cube now referes to this.
cube.coords['x2'] = ('x', [1,2,3])
cube.swap_dims({'x':'x2'}, inline=True) #This just changes the first dataarray object. cube still refers to the first dataarray object

So the question again was, is there a reason, why the inline keyword is implemented on dataset, but not on dataarray.

If I did it again, I would remove the inplace keyword argument from even Dataset methods. Almost every method in xarray creates new underlying objects, so it's misleading to call them "inplace". (Most of our inplace methods actually create a new object and then assign its properties to an existing object.)

Why not make your accessor method return a new DataArray object? That would be more inline with how most xarray methods work. If you have any special attributes you've set on the accessor you can copy those over to the new result.

Thank you for your explination. I will to look into my problem from another angle.

If I did it again, I would remove the inplace keyword argument from even Dataset methods.

How about deprecating these? Would that be too much of a change you think? (I never use them)

How about deprecating these? Would that be too much of a change you think? (I never use them)

Yes, this is a good idea! Opened #1756 to keep track.

Navigation Menu

Repository navigation

Is there a reason for why DataArray.swap_dims() cannot be done in place like Dataset.swap_dims? #1755

Problem description

Code Sample

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Is there a reason for why DataArray.swap_dims() cannot be done in place like Dataset.swap_dims? #1755

Description

Problem description

Code Sample

Activity

rabernat commented on Dec 1, 2017

leeviannala commented on Dec 1, 2017

rabernat commented on Dec 1, 2017

leeviannala commented on Dec 1, 2017

shoyer commented on Dec 1, 2017

leeviannala commented on Dec 2, 2017

fmaussion commented on Dec 2, 2017

shoyer commented on Dec 2, 2017

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions