Exception pickling kernels

[mbway]

I get this error when pickling a kernel once it has been used by a GP. Since the exception doesn't occur before this, I am assuming that the kernels are supposed to be picklable? The error is the same for other kernels I tried. It is also triggered by using copy.deepcopy.
My intention was to save the unused kernel anyway, and so passing a copy to the GP solved my issue, but this might still be a bug that needs fixing?

#!/usr/bin/env python3
import GPy
import pickle
import numpy as np
import sys

print('python version = ' + sys.version)
print('GPy.__version__ = ' + GPy.__version__)

k = GPy.kern.Matern52(1)

print('before using')
pickle.loads(pickle.dumps(k))
print('k._name = ' + k._name)

gp = GPy.models.GPRegression(np.array([[1]]), np.array([[1]]), kernel=k)

print('after using')
print('k._name = ' + k._name)
print('saving')
saved = pickle.dumps(k)
print('loading')
pickle.loads(saved)

python version = 3.6.3 (default, Oct  3 2017, 21:45:48) 
[GCC 7.2.0]
GPy.__version__ = 1.9.2
before using
k._name = Mat52
after using
k._name = Mat52
saving
loading
Traceback (most recent call last):
  File "pickle_problem.py", line 20, in <module>
    pickle.loads(saved)
  File "/usr/local/lib/python3.6/dist-packages/paramz/parameterized.py", line 340, in __setstate__
    super(Parameterized, self).__setstate__(state)
  File "/usr/local/lib/python3.6/dist-packages/GPy/core/parameterization/priorizable.py", line 18, in __setstate__
    super(Priorizable, self).__setstate__(state)
  File "/usr/local/lib/python3.6/dist-packages/paramz/core/parameter_core.py", line 495, in __setstate__
    super(Parameterizable, self).__setstate__(state)
  File "/usr/local/lib/python3.6/dist-packages/paramz/core/indexable.py", line 56, in __setstate__
    super(Indexable, self).__setstate__(state)
  File "/usr/local/lib/python3.6/dist-packages/paramz/core/pickleable.py", line 125, in __setstate__
    self.observers = ObserverList()
  File "/usr/local/lib/python3.6/dist-packages/paramz/parameterized.py", line 330, in __setattr__
    pnames = self.parameter_names(False, adjust_for_printing=True, recursive=False)
  File "/usr/local/lib/python3.6/dist-packages/paramz/core/parameter_core.py", line 191, in parameter_names
    names.extend([adjust(x.name) for x in self.parameters])
  File "/usr/local/lib/python3.6/dist-packages/paramz/core/parameter_core.py", line 191, in <listcomp>
    names.extend([adjust(x.name) for x in self.parameters])
  File "/usr/local/lib/python3.6/dist-packages/paramz/core/nameable.py", line 63, in name
    return self._name
AttributeError: 'Matern52' object has no attribute '_name'

[cwlgadd]

I have run into the exact same problem. Unfortunately I can't use the same work around though.

[yenicelik]

Exact same problem over here! Did anyone find a work-around or a fix?

[cwlgadd]

My work-around was to not store the GP/kernel and to re-instantiate it every time.. which isn't efficient, but worked for me because I'm using a mixture model. I didn't have the time to look into a fix.

[cwlgadd]

Has anyone been able to look into a fix?

[maxwab]

Same here. This issue makes it impossible to use GPy with multiprocessing for me.

[gregorgebhardt]

Following up on @cwlgadd a possible work-around is to implement the methods _to_dict / to_dict (see here and here) and _from_dict (see here) and use them for serializing the kernel objects. This allows to use GPy with multiprocessing.

[yenicelik]

Another work-around for multiprocessing is to use the https://github.com/uqfoundation/multiprocess which has the exact same interface as the standard python multiprocessing, but uses dill instead of pickle as the backend.

[mbway]

in my testing dill also triggers the exception

[Amir-Arsalan]

@mbway Were you able to use the the to_dict method? It seems pretty easy to use but I have issues using as shown here. Would appreciate if you can help me figure out where I'm making a mistake.

[mbway]

for my use case I didn't actually require pickling kernels that have been used, so I restructured my code to copy the kernel and pass the copy to the GP instead, allowing the object holding the unused kernel to be pickled. This probably doesn't help most cases though, sorry