6d82a09a33437be16a6b9ca241b42d42d6a1a23e — dieggsy 4 years ago c7f01ae
Add wiki page, use genann-copy* in example2
2 files changed, 82 insertions(+), 1 deletions(-)

M example2.scm
A genann.wiki
M example2.scm => example2.scm +1 -1
@@ 15,7 15,7 @@
(let loop ((ann (make-genann 2 1 2 1))
           (last-err 1000.0)
           (count 1))
  (let ((save (genann-copy ann)))
  (let ((save (genann-copy* ann)))
    (when (= 0 (modulo count 1000))
      (genann-randomize ann)
      (set! last-err 1000.0))

A genann.wiki => genann.wiki +81 -0
@@ 0,0 1,81 @@
== Genann


=== Introduction

Genann is an chicken egg that provides bindings to the
[[https://github.com/codeplea/genann|genann]] ANSI C neural network library.
The C library is written only in standard C, the egg is written only in
'standard' Chicken 5 (what comes included with C5, no eggs).

=== Genann API

<procedure>(genann-init inputs hidden-layers hidden-neurons outputs)</procedure>
Creates and returns a genann record. Each argument is an integer specifying how
many of that the genann should have. {{hidden-neurons}} is the number of hidden
neurons ''per hidden layer.''

<procedure>(genann-copy genann)</procedure>
Return a new copy of {{genann}}.

<procedure>(genann-free! genann)</procedure>
Free the memory used by {{genann}}.

<procedure>(genann-train genann inputs desired-outputs learning-rate)</procedure>
Does a single backpropagation update on {{genann}}, where {{inputs}} and
{{desired-outputs}} are f64vectors and {{learning-rate}} is a flonum.

<procedure>(genann-run genann inputs)</procedure>
Runs the feedforward algorithm to calculate the ann's output from f64vector
{{inputs}}. Returns outputs as an f64vector.

<procedure>(genann-randomize genann)</procedure>
Sets the weights in {{genann}} randomly. This is called by {{genann-init}}.

<procedure>(genann-read #!optional port)</procedure>
Read a genann record from file port {{port}}.

<procedure>(genann-write #!optional port)</procedure>
Write a genann to file port {{port}}.

=== Added procedures
<procedure>(genann-init* inputs hidden-layers hidden-neurons outputs)</procedure>
<procedure>(make-genann ...)</procedure>
Like {{genann-init}}, but sets {{genann-free!}} as a finalizer to the returned
genann so it can be properly garbage collected. {{make-genann}} is the same
procedure as {{genann-init*}}.

<procedure>(genann-copy* genann)</procedure>
Like {{genann-copy}}, but sets {{genann-free!}} as a finalizer to the returned
genann so it can be properly garbage collected.

<procedure>(genann-inputs genann)</procedure>
<procedure>(genann-hidden-layers genann)</procedure>
<procedure>(genann-hidden-neurons genann)</procedure>
<procedure>(genann-outputs genann)</procedure>
<procedure>(genann-total-weights genann)</procedure>
Returns number of inputs, hidden layers, hidden neurons per layer, and outputs,
respectively, in {{genann}}. These are more or less equivalent to the following
in C:
<enscript highlight="c">
// genann is created with genann_init

<procedure>(genann-weight-set! genann i x)</procedure>
<procedure>(genann-weight-ref genann i)</procedure>
Set or get the ith weight of {{genann}}, with the former being a setter for the
latter, so you can used {{set!}} with {{genann-weight-ref}}. Useful for
training with random search, as in

=== Examples
* [[https://github.com/dieggsy/genann/blob/master/example1.scm|Training using backpropagation]]
* [[https://github.com/dieggsy/genann/blob/master/example2.scm|Training using random search]]
* [[https://github.com/dieggsy/genann/blob/master/example3.scm|Loading a saved ANN]]
* [[https://github.com/dieggsy/genann/blob/master/example4.scm|Training on the IRIS dataset using backpropagation]]