[Numpy-discussion] Possible to pickle new state in NDArray subclasses?

Nathan Goldbaum nathan12343 at gmail.com
Wed Dec 14 14:51:26 EST 2016


I'm able to do this in my ndarrary subclass using __reduce__ and
__setstate__:

https://bitbucket.org/yt_analysis/yt/src/yt/yt/units/yt_array.py#yt_array.py-1250

Here it's being used to save the unit information into the pickle for a
unit-aware ndarray subclass.

On Wed, Dec 14, 2016 at 1:45 PM, Stuart Reynolds <stuart at stuartreynolds.net>
wrote:

> I'm trying to subclass an NDArray as shown here:
>    https://docs.scipy.org/doc/numpy/user/basics.subclassing.html
>
> My problem is that when I save the new class' state with pickle, the new
> attributes are lost. I don't seem to be able to override __getstate__ or
> __setstate__ to achieve this?
>
> Is it possible to allow new state to serialized when overriding an NDArray?
>
> In my example below, __setstate__ gets called by pickle but not
> __getstate__.
> In the final line, a RealisticInfoArray has been deserialized, but it has
> no .info attribute.
>
> ----
>
> import cPickle as pickle
> import numpy as np
>
> class RealisticInfoArray(np.ndarray):
>     def __new__(cls, arr, info):
>         obj = np.asarray(arr).view(cls)
>         obj.info = info
>         return obj
>
>     def __array_finalize__(self, obj):
>         if obj is None: return
>         self.info = getattr(obj,"info",None)
>
>     def __setstate__(self, *args):
>         print "SET"
>         return np.ndarray.__setstate__(self,*args)
>
>     def __getstate__(self):
>         print "GET"
>         assert False, "EXPLODE"
>         return np.ndarray.__getstate__(self)
>
> arr = np.zeros((2,3), int)
> arr = RealisticInfoArray(arr, "blarg")
> print arr.info
> arr2 = pickle.loads(pickle.dumps(arr))
> print arr2.info  # no .info attribute!
>
>
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> https://mail.scipy.org/mailman/listinfo/numpy-discussion
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20161214/2edaf711/attachment.html>


More information about the NumPy-Discussion mailing list