psyrun 0.8.0

Easy parameter space evaluation and serial farming.

Psyrun

Psyrun is a Python tool to define parameter spaces and execute an evaluation function for each parameter assignment. In addition Psyrun makes it easy to use serial farming, i.e. evaluating multiple parameter assignments in parallel, on a multicore computers and high-performance clusters.

Documentation

The documentation can be found here.

Overview

Define parameter spaces and evaluate them:

from psyrun import map_pspace, Param

def objective(a, b, c):
    return a * b + c

pspace = (Param(a=np.arange(1, 5))
          * Param(b=np.linspace(0, 1, 10))
          * Param(c=[1., 1.5, 10., 10.5]))
results = map_pspace(objective, pspace)

Or do it in parallel:

from psyrun import map_pspace_parallel
results = map_pspace_parallel(objective, pspace)

Define tasks by placing task_<name>.py files in the psy-tasks` directory:

from psyrun import Param

pspace = (Param(a=np.arange(1, 5))
          * Param(b=np.linspace(0, 1, 10))
          * Param(c=[1., 1.5, 10., 10.5]))

def execute(a, b, c):
    return {'result': a * b + c}

and run them by typing psy run with support for serial farming on high performance clusters.

Installation

pip install psyrun

To be able to use the NPZ store:

pip install numpy
pip install 'psyrun[npz]'

To be able to use the HDF5 store:

pip install numpy
pip install 'psyrun[h5]'

Requirements

• Python >=2.7, >=3.4

• six

Optional requirements

To have faulthandler activated for jobs submitted with psy run in Python 2.7:

• faulthandler

Python 3.4+ already includes the faulthandler module.

To use map_pspace_parallel:

• joblib

To use NPZ files as store:

• NumPy

To use HDF5 files as store:

• NumPy

• pytables

To run the unit tests:

• joblib

• NumPy

• pytables

• pytest >= 2.8

To build the documentation:

• numpydoc

Changes

0.8.0 (June 26, 2018)

New features

• Added the possibility to specify a setup function in task files that will be executed once at the start of each worker process and can be used to acquire resources and pass them to the processing function for individual parameter sets.

0.7.1 (April 19, 2018)

Documentation improvements

• Added documentation for the pool_size task attribute and included it in the task template.

Bug fixes

• Allow to continue tasks if results are missing even if the result file is newer than the task file.

0.7.0 (February 18, 2018)

New features

• Added support for the Slurm Workload Manager.

0.6.0

New features

• Add psy new-task and psy kill commands.

• Added AutodetectStore that determines the appropriate store from the filename extension.

• Added possibility to let psy merge custom stores if provided as psyrun.stores entry point.

• Added capability to set scheduler arguments based on the job name.

0.5.4

Bug fixes

• Fix the psy run continue functionality.

0.5.3

Bug fixes

• Fix psy status and psyrun.backend.distribute.DistributeBackend.get_missing trying to read incompatible data files in the output directory.

• Fix psy status and psyrun.backend.distribute.DistributeBackend.get_missing easily hitting Python’s recursion depth limit.

• Fix merging of npz files with missing integer values by converting them to float where np.nan can be used.

0.5.2

Bug fixes

• Fix incorrect psy status.

• Fix psy run <task1> <task2> ... not running all tasks and run them in order.

0.5.1

Bug fixes

• Fix psy merge always assuming PickleStore.

Documentation improvements

• Add recipe for converting data to Pandas data frame to documentation.

0.5

• Initial release