Pickling issue for the Stumpi

[soja-soja]

Hi,

in case of ongoing streaming of data in different groups (so dealing with more than one stumpi instance), one efficient approach might be to pickle the instances, and then use them whenever new datapoints related to that group is available.

However, due to the implementation I believe the class is not pickleable. Any thought on that?

replicating that would be as easy as :

import pickle
import stumpy

x = stumpy.stumpi(list(map(float, [1,2,3,4,5,6,7,8])), m=3)
pickle.dumps(x)

[seanlaw]

@soja-soja Thank you for your question and welcome to the STUMPY community. Unfortunately, functions like stumpy.stumpi are "decorated" functions and it's not easy/possible to pickle them (see this past attempt). Would it be possible to use the more flexible cloudpickle package instead?

import numpy as np
import stumpy
import cloudpickle


if __name__ == "__main__":
    T = np.random.rand(100)
    m = 50

    stream = stumpy.stumpi(T, m, egress=True)
    stream.update(np.random.rand())
    print(stream.left_P_)

    dump_stream = cloudpickle.dumps(stream)

    stream = cloudpickle.loads(dump_stream)
    print(stream.left_P_)

[NimaSarajpoor]

@seanlaw
Is issue #621 related? Because I think stumpi should be fast if we allow user pass mp (of course, we may want to save some other stuff as well like sliding mean and std). one can save mp (as mp_file.npy) and then just load it and pass it to the stumpi.

[soja-soja]

Thanks for the prompt response, that seems like a good alternative, hopefully I can have cloudpickle in my project.

Also per Nima's suggestion, having pre-computed mp is a great idea and this one is also a use-case for that. I think we cannot rely solely on the current use-cases and I'm sure more long-term use-cases will be using stumpy in the future.