Simple interface for numpy arrays in multiprocessing contexts.
Project description
contextshare
Simple context manager to share numpy arrays in python multiprocessing. Installation on Python >= 3.8:
pip install contextshare
Example
The following code example is complete and compares a serial()
implementation with the parallel()
version thereof implemented with contextshare
.
#!/usr/bin/env python
import contextshare as shared
import numpy as np
def serial(sigmas):
alphas = []
for sigma in sigmas:
K = np.exp(-0.5 * D / (sigma * sigma))
alphas.append(np.linalg.solve(K, b))
return alphas
def parallel(sigmas):
with shared.SharedMemory({"D": D, "b": b}, nworkers=11) as sm:
@sm.register
def one_sigma(sigma):
K = np.exp(-0.5 * D / (sigma * sigma))
return np.linalg.solve(K, b)
for sigma in sigmas:
one_sigma(sigma)
return sm.evaluate(progress=True)
if __name__ == "__main__":
N = 2000
D = np.random.random((N, N))
b = np.random.random(N)
sigmas = 2.0 ** np.arange(1, 12)
s = serial(sigmas)
p = parallel(sigmas)
Documentation
The context manager shared.SharedMemory
takes two arguments: the dictionary of numpy arrays to make available on all parallel workers and the number of workers to create.
# makes variables D and b available under the same name on 11 workers
with shared.SharedMemory({"D": D, "b": b}, nworkers=11) as sm:
The workers are spawned upon entering the context and are stopped upon exiting. Shared memory references are cleaned up automatically. Note that only numpy arrays are supported for sharing. Other arguments should be placed in the arguments of the function below. The function to call (i.e. the body of the serial for loop) needs to be placed in a function and either decorated or called explicitly:
# decorator
@sm.register
def one_sigma(sigma):
pass
# or, equivalently, an explicit call
sm.register(one_sigma)
Calling this decorated function returns immediately and enqueues a function call. Calling
sm.evaluate(progress=True)
starts the calculations and returns the results in order. With progress=True
a progress bar is shown, default is to be silent.
Credits
This package was created with Cookiecutter using the audreyr/cookiecutter-pypackage` template.
======= History
0.1.0 (2021-02-23)
- First release on PyPI.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for contextshare-0.1.1-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | f47bfaa7aff5212afdbcfa472e227e0045d40284d64f9318c9cd5cbababf8df1 |
|
MD5 | c12ad24e99899bdea8a92f18b05fbeb9 |
|
BLAKE2b-256 | 186fed2e1dd3e900b5573546ed2d482c32a8084a3e61582508ae512e37113137 |