|
AI::NNFlex - A customisable neural network simulator |
AI::NNFlex - A customisable neural network simulator
use AI::NNFlex;
my $network = AI::NNFlex->new([array of hashes],{hash of global config});
use AI::NNFlex::Dataset;
| my $dataset = AI::NNFlex::Dataset->new([ | |
| [INPUTARRAY],[TARGETOUTPUT], | |
| [INPUTARRAY],[TARGETOUTPUT]]); |
my $sqrError = 10;
while ($sqrError >0.01)
{
$sqrError = $dataset->learn($network);
}
$network->lesion({'nodes'=>PROBABILITY,'connections'=>PROBABILITY});
my $outputsRef = $dataset->run($network);
AI::NNFlex is intended to be a highly flexible, modular NN framework. It's written entirely in native perl, so there are essentially no prereq's. The following modular divisions are made:
* NNFlex.pm
the core module. Contains methods to construct and
lesion the network
* feedforward.pm
the network type module. Feedforward is the only type
currently defined, but others may be created and
imported at runtime
* backprop.pm
the learning algorithm. Backprop is the only algorithm
currently defined, but others may be created and
imported at runtime
* <activation>.pm
node activation function. Currently the options are
tanh, linear & sigmoid.
* Dataset.pm
methods for constructing a set of input/output data
and applying to a network.
The code should be simple enough to use for teaching purposes, but a simpler implementation of a simple backprop network is included in the example file bp.pl. This is derived from Phil Brierleys freely available java code at www.philbrierley.com.
AI::NNFlex leans towards teaching NN and cognitive modelling applications. Future modules are likely to include more biologically plausible nets like DeVries & Principes Gamma model.
new ( { HASH OF NETWORK STRUCTURE },{ HASH OF CONFIG OPTIONS });
This should contain sub-hashes for each layer in the network, containing the following options.
nodes=>NUMBER OF NODES
decay=>AMOUNT OF ACTIVATION TO DECAY PER TICK
persistent activation=>TRUE TO RETAIN ACTIVATION BETWEEN
TICKS
random activation=>MAXIMUM VALUE FOR INITIAL ACTIVATION
=head3 config options
This should contain the global config options for the net. The following are defined:
random weights=>MAXIMUM VALUE FOR INITIAL WEIGHT
learning algorithm=>The AI::NNFlex module to import for
training the net
networktype=>The AI::NNFlex module to import for flowing
activation
debug=>[LIST OF CODES FOR MODULES TO DEBUG]
new ( [[INPUT VALUES],[OUTPUT VALUES],[INPUT VALUES],[OUTPUT VALUES],..])
These should be comma separated values. They can be applied to the network with ::run or ::learn
These are the intended or target output values. Comma separated These will be used by ::learn
This is a short list of the main methods. For details on all available methods, please see individual pod pages below, and in individual imported modules.
$network->lesion ({'nodes'=>PROBABILITY,'connections'=>PROBABILITY})
Damages the network.
PROBABILITY
A value between 0 and 1, denoting the probability of a given node or connection being damaged.
Note: this method may be called on a per network, per node or per layer basis using the appropriate object.
$dataset->learn($network)
'Teaches' the network the dataset using the networks defined learning algorithm. Returns sqrError;
$dataset->run($network)
Runs the dataset through the network and returns a reference to an array of output patterns.
See the code in ./examples. For the simplest example, see xor_with_datasets.pl
Phil Brierley, for his excellent free java code, that solved my backprop problem
Dr Martin Le Voi, for help with concepts of NN in the early stages
Dr David Plaut, for help with the project that this code was originally intended for.
Graciliano M.Passos for suggestions & improved code (see SEE ALSO).
AI::NNEasy - Developed by Graciliano M.Passos Shares some common code with NNFlex. Much faster, and more suitable for backprop projects with large datasets.
Lots of things:
clean up the perldocs some more
write gamma modules
write BPTT modules
write a perceptron learning module
speed it up
write a tk gui
v0.11 introduces the lesion method, png support in the draw module and datasets.
v0.12 fixes a bug in reinforce.pm & adds a reflector in feedforward->run to make
$network->run($dataset) work.
v0.13 introduces the momentum learning algorithm and fixes a bug that allowed
training to proceed even if the node activation function module can't be loaded
Copyright (c) 2004-2005 Charles Colbourn. All rights reserved. This program is free software; you can redistribute it and/or modify it under the same terms as Perl itself.
Below are PODs for individual methods
takes as params a reference to a hash. Each hash will be created as a layer.
Valid parameters are currently:
* nodes - number of nodes in the layer
* decay - float, amount that each node in the
layer will decay with each tick.
* persistent activation - whether activation is summed between
ticks.
* adjust error - <NYI>
* random activation - parameter to pass to RAND for random
activation seeding
Additional parameters may be specified as individual param=>value pairs
AFTER the layer hash. These may include:
* random weights - parameter to pass to RAND for random
weights seeding
* random connections - The /probability/ factor of a connection
being made between two nodes in the
network.
(Note that no support exists at present
for restricting the connectivity at a
layer by layer level, although this may
be done by combining two or more network
objects and handling activation from one
to the next programmatically)
* learning algorithm - the learning algorithm (this must be a
valid compatible perl module).
* networktype - E.g. feedforward. Must be a compatible
perl module.
* debug - level of debug information
a different debug level is assigned to
each module type, and the debug property is
an ARRAY of which debugs you require.
0 - error
1 - TBD
2 - NNFlex.pm core debug
3 - networktype debug
4 - learning algorithm debug
5 - activation function debug
6 - GUI/Graphic
Plus other custom settings used in networktype & learning algorithm modules, such as:
* learning rate - A constant for use in e.g. backprop learning
Returns a network object that contains $$network->{'layers'} which is an array of 'layer' objects.
The layer object contains a property 'nodes' which is an array of nodes in that layer. So programmatically if you want to access a particular node (or to interact with the mesh for writing networktypes and learning algorithms) you can access any node directly using the syntax
$network->{'layers'}->[layer number]->{'nodes'}->[node number]->{property}
(HINT: or do foreach's on the arrays)
Copyright (c) 2004-2005 Charles Colbourn. All rights reserved. This program is free software; you can redistribute it and/or modify it under the same terms as Perl itself.
$object->output({``output''=>``1''}); returns the activation of layer 1
else returns activation of last layer as a reference to an array
called from AI::NNFlex::new. no external use required, but not defined as local, in case of debugging use
Init runs through each layer of node objects, creating properties in each node:
* connectedNodesEast - Nodes connected to this node in the layer
to the 'east', i.e. next layer in feedforward
terms.
* connectedNodesWest - Nodes connected to this node in the layer to
the west, i.e. previous layer in feedforward
networks.
These properties are hashes, with the node object acting as a key. Each value is a weight for this connection. This means that you /can/ have connection weights for connections in both directions, since the weight is associated with an incoming connection.
access with the following syntax:
$node->{'connectedNodesWest'}->{'weights'}->{$connectedNode} = 0.12345 $node->{'connectedNodesWest'}->{'nodes'}->[number] = $nodeObject
Note also that the probability of a connection being created is equal to the numeric value of the global property 'random connections' expressed as a decimal between 0 and 1. If 'random connections' is not specified all connections will be made.
The connections are /only/ created from west to east. Connections that already exist from west to east are just copied for the 'connectedNodesWest' property.
No return value: the connections are created in the $network object.
These connectedNodesEast & West are handy, because they are arrays you can foreach them to iterate through the connected nodes in a layer to the east or west.
Internal use, writes to STDOUT parameter 1 if parameter 2 == global variable $DEBUG. or parameter 2 == 1
$network->dump_state({``filename''=>``test.wts''[, ``activations''=>1]});
Dumps the current contents of the node weights to a file.
| useage: | |
| $network->load_state(<filename>); |
Initialises the network with the state information (weights and, optionally activation) from the specified filename.
Note that if you have a file containing activation, but you are not using persistent activation, the activation states of nodes will be reset during network->run
Lesions a node to emulate damage. Syntax is as follows
$network->lesion({'nodes'=>.2,'connections'=>.4});
assigns a .2 probability of a given node being lesioned, and .4 probability of a given connection being lesioned. Either option can be omitted but it must have one or the other to do. If you programmatically need to call it with no lesioning to be done, call with a 0 probability of lesioning for one of the options.
return value is true if successful;
The layer object
Create new layer
Takes the parameters from AI::NNFlex::layer and passes them through to AI::NNFlex::node::new
(Uses nodes=>X to decide how many nodes to create)
Returns a layer object containing an array of node objects
Receives a reference to a hash of parameters. Valid inputs are
* layer - the layer number you want output from
Returns a reference to an array of outputs
Used by AI::NNFlex::output
Lesions a node to emulate damage. Syntax is as follows
$layer->lesion({'nodes'=>.2,'connections'=>.4});
assigns a .2 probability of a given node being lesioned, and .4 probability of a given connection being lesioned. Either option can be omitted but it must have one or the other to do. If you programmatically need to call it with no lesioning to be done, call with a 0 probability of lesioning for one of the options.
return value is true if successful;
the node object
Takes parameters passed from NNFlex via AI::NNFlex::layer
returns a node object containing:
* activation - the nodes current activation
* decay - decay rate
* adjust error - NYI
* persistent activation - if true, activation will be summed on
each run, rather than zeroed
before calculating inputs.
* ID - node identifier (unique across the
NNFlex object)
* threshold - the level above which the node
becomes active
* activation function - the perl script used as an activation
function. Must perform the calculation
on a variable called $value.
* active - whether the node is active or
not. For lesioning. Set to 1
on creation
NOTE: it is recommended that nothing programmatic be done with ID. This is intended to be used for human reference only.
$node->lesion({'nodes'=>.2,'connections'=>.4});
assigns a .2 probability of a given node being lesioned, and .4 probability of a given connection being lesioned. Either option can be omitted but it must have one or the other to do. If you programmatically need to call it with no lesioning to be done, call with a 0 probability of lesioning for one of the options.
return value is true if successful;
Implemented as a method of node to permit node by node or layer by layer lesioning
|
AI::NNFlex - A customisable neural network simulator |