rdkit.ML.Neural.Trainers module

Training algorithms for feed-forward neural nets

Unless noted otherwise, algorithms and notation are taken from: “Artificial Neural Networks: Theory and Applications”,

Dan W. Patterson, Prentice Hall, 1996
class rdkit.ML.Neural.Trainers.BackProp(speed=0.5, momentum=0.7)

Bases: rdkit.ML.Neural.Trainers.Trainer

implement back propagation (algorithm on pp 153-154 of Patterson)

I don’t think that I’ve made any assumptions about the connectivity of
the net (i.e. full connectivity between layers is not required).
NOTE: this code is currently making the assumption that the activation
functions on the nodes in the network are capable of calculating their derivatives using only their values (i.e. a DerivFromVal method should exist). This shouldn’t be too hard to change.
StepUpdate(example, net, resVect=None)

does a BackProp step based upon the example

Arguments

  • example: a 2-tuple:
    1. a list of variable values values
    2. a list of result values (targets)

  • net: a _Network_ (or something supporting the same API)

  • resVect: if this is nonzero, then the network is not required to classify the _example_

Returns

the backprop error from _network_ before the update

Note

In case it wasn’t blindingly obvious, the weights in _network_ are modified in the course of taking a backprop step.
TrainOnLine(examples, net, maxIts=5000, errTol=0.1, useAvgErr=1, silent=0)

carries out online training of a neural net

The definition of online training is that the network is updated after
each example is presented.

Arguments

  • examples: a list of 2-tuple:
    1. a list of variable values values
    2. a list of result values (targets)

  • net: a _Network_ (or something supporting the same API)

  • maxIts: the maximum number of training epochs (see below for definition) to be run

  • errTol: the tolerance for convergence

  • useAvgErr: if this toggle is nonzero, then the error at each step will be divided by the number of training examples for the purposes of checking convergence.

  • silent: controls the amount of visual noise produced as this runs.

Note

a training epoch is one complete pass through all the training examples
class rdkit.ML.Neural.Trainers.Trainer

Bases: object

“virtual base class” for network trainers