naklar [na'klaɐ̯]

Naklar is a support framework for managing experiments with a focus on staying out of the way and providing and easy way to select experiment(s) based on SQL style queries.

There is a host of libraries already that allow you to configure and run experiments, naklar does not care, you run your experiments where ever, when ever, how ever. At the end of them there should be a dictionary recording the settings and what ever else you wish to keep track of, that's the point at which naklar comes in.

features

• easy-to-use and flexible database connectors based on sqlalchemy
• autoloading table contents from dictionaries stored on disk
• column value transformations during get/set operations using Python properties

dependencies

Naklar uses a data base to search / store and retrieve experiment results and depends on sqlalchemy and some data base backend.

• sqlalchemy
  • Under the hood naklar uses sqlalchemy to perform data base queries
• SQLite / MySQL / PyMysql
  • for accessing a data base
• six
  • for Py2/3 compatibility

usage

Better and more complete examples will eventually be here http://nbviewer.ipython.org/github/mattilyra/naklar/blob/master/examples/naklar_examples.ipynb

load settings.pkl pickled dictionaries from disk

from naklar import experiment
experiment.initialise('./results/',
                      autoload=True,
                      dict_filename='settings.pkl')

# get all experiments
experiment.select()

# get only those where C=1
# you need to know which parameters are present
for exp in experiment.select(C=1)
    print(exp.C, exp.kernel, exp.path)

load joblib dumps

from functools import partial
from naklar import experiment
import joblib

load_func = partial(joblib.load, mmap_mode='r')
experiment.initialise('./results/',
                      autoload=True, dict_filename='conf.joblib',
                      load_func=load_func)

only load certain keys if some of the values are not compatible with sqlalchemy, for instance NumPy arrays

from functools import partial
from naklar import experiment
import joblib

# use memmap to avoid expensive unnecessary load operations
load_func = partial(joblib.load, mmap_mode='r')
experiment.initialise('./results/',
                    autoload=True, dict_filename='conf.joblib',
                    load_func=load_func,
                    restrict_keys=set(['alpha', 'eta', 'path']))

what's missing

• an easy way to inspect what parameters (the keys) are defined for experiments
• ...