msdnet.train module

Module for training networks.

class msdnet.train.TrainAlgorithm

Bases: abc.ABC

Base class implementing a training algorithm.

abstract step(n, dlist)

Take a single algorithm step.

Parameters

• n – network.Network to train with

• dlist – list of data.DataPoint to train with

abstract to_dict()

Save algorithm state to dictionary

abstract load_dict(dct)

Load algorithm state from dictionary

abstract classmethod from_dict(dct)

Load algorithm from dictionary

classmethod from_file(fn)

Load algorithm from file

to_file(fn)

Save algorithm state to file

class msdnet.train.AdamAlgorithm(network, a=0.001, b1=0.9, b2=0.999, e=1e-08)

Bases: msdnet.train.TrainAlgorithm

Implementation of the ADAM algorithm.

Parameters

• network – network.Network to train with

• a – ADAM parameter

• b1 – ADAM parameter

• b2 – ADAM parameter

• e – ADAM parameter

step(n, dlist)

Take a single algorithm step.

Parameters

• n – network.Network to train with

• dlist – list of data.DataPoint to train with

to_dict()

Save algorithm state to dictionary

load_dict(dct)

Load algorithm state from dictionary

classmethod from_dict(dct)

Load algorithm from dictionary

msdnet.train.restore_training(fn, netclass, trainclass, valclass, valdata, gpu=True)

Restore training from file.

Parameters

• fn – filename to load

• netclass – network.Network class to use

• trainclass – TrainAlgorithm class to use

• valclass – validate.Validation class to use

• valdata – list of data.DataPoint to validate with

• gpu – (optional) whether to use GPU or CPU

Returns

network object, training algorithm object, and validation object

msdnet.train.train(network, trainalg, validation, dataprov, outputfile, val_every=None, loggers=None, stopcrit=inf, progress=False)

Train network.

Parameters

• network – network.Network to train with

• trainalg – TrainAlgorithm object that performs training.

• validation – validate.Validation object that performs validation.

• dataprov – data.BatchProvider object that generates training batches.

• outputfile – file to store trained network parameters in

• val_every – (optional) number of training steps before each validation step

• loggers – (optional) list of loggers.Logger objects to perform logging.

• stopcric – (optional) number of validations steps without improvement before stopping training

• progress – (optional) whether to show progress during training