"如何用70行Java代码实现深度神经网络算法" 的delphi版本
版权声明：本文为博主原创文章，未经博主允许不得转载。
=====ann.pas源程序===================================
{
by 阿甘 2016.2.23参考自此篇文档
如何用70行Java代码实现深度神经网络算法
http://geek.csdn.net/news/detail/56086
原文中的代码作者:fourinone
原文中的代码是用java写的，现移植为delphi 供参考
略作修改：权重不含输出层
}
unit ann;
interface
type
Tdbarr=array of double;
TBpDeep=class
private
layer:array of array of double;
layerErr:array of array of double;
layer_weight:array of array of array of double;
layer_weight_delta:array of array of array of double;
mobp,rate:double;
procedure updateWeight(tar:array of double);
public
constructor Create(layernum:array of integer;rate,mobp:double);
function computeout(input:array of double):Tdbarr;
procedure train(input,tar:array of double);
end;
implementation
constructor TBpDeep.Create(layernum:array of integer;rate,mobp:double);
var
a,i,j,k:integer;
begin
self.rate:=rate;
self.mobp:=mobp;
a:=length(layernum);
if a<2 then exit;
setlength(layer,a);
setlength(layerErr,a);
setlength(layer_weight,a-1);
setlength(layer_weight_delta,a-1);
Randomize;
for k:=0 to a-1 do
begin
setlength(layer[k],layernum[k]);
setlength(layerErr[k],layernum[k]);
if k+1<a then
begin
setlength(layer_weight[k],layernum[k]+1,layernum[k+1]);
setlength(layer_weight_delta[k],layernum[k]+1,layernum[k+1]);
for j:=0 to layernum[k]-1 do
for i:=0 to layernum[k+1]-1 do
layer_weight[k][j][i]:=Random;//随机初始化权重
end;
end;
end;
function TBpDeep.computeout(input:array of double):Tdbarr;
var
i,j,k:integer;
z:double;
begin
for k:=1 to high(layer) do
for j:=0 to high(layer[k]) do
begin
z:=layer_weight[k-1][length(layer[k-1])][j];
for i:=0 to high(layer[k-1]) do
begin
if k=1 then layer[k-1][i]:=input[i];
z:=z+layer_weight[k-1][i][j]*layer[k-1][i];
end;
layer[k][j]:=1/(1+exp(-z));
end;
result:=Tdbarr(layer[high(layer)]);
end;
procedure TBpDeep.updateWeight(tar:array of double);
var
i,j,k:integer;
z:double;
begin
k:=high(layer);
for j:=0 to high(layererr[k]) do layerErr[k][j]:=layer[k][j]*(1-layer[k][j])*(tar[j]-layer[k][j]);
while k>0 do
begin
dec(k);
for j:=0 to high(layererr[k]) do
begin
z:=0;
for i:=0 to high(layererr[k+1]) do
begin
if z+k>0 then z:=layerErr[k+1][i]*layer_weight[k][j][i]
else z:=0;
layer_weight_delta[k][j][i]:=mobp*layer_weight_delta[k][j][i]+rate*layerErr[k+1][i]*layer[k][j];//隐含层动量调整
layer_weight[k][j][i]:=layer_weight[k][j][i]+layer_weight_delta[k][j][i];//隐含层权重调整
if j=high(layererr[k]) then
begin
layer_weight_delta[k][j+1][i]:=mobp*layer_weight_delta[k][j+1][i]+rate*layerErr[k+1][i];//截距动量调整
layer_weight[k][j+1][i]:=layer_weight[k][j+1][i]+layer_weight_delta[k][j+1][i];//截距权重调整
end;
end;
layerErr[k][j]:=z*layer[k][j]*(1-layer[k][j]);//记录误差
end;
end;
end;
procedure TBpDeep.train(input,tar:array of double);
begin
computeout(input);
updateWeight(tar);
end;
end.
========================================
测试代码：
procedure TForm1.bttrainClick(Sender: TObject);//训练
var
i,k:integer;
const
data:array[0..3] of array[0..1] of double=((1,2),(2,2),(1,1),(2,1));
tar:array[0..3] of array[0..1] of double=((1,0),(0,1),(0,1),(1,0));
begin
for k:=0 to 499 do
for i:=0 to high(data) do
bp.train(data[i], tar[i]);
end;
procedure TForm1.btcomputeClick(Sender: TObject);//识别
var
rst:Tdbarr;
x:array[0..1] of double;
begin
x[0]:=strtofloat(xx.Text);
x[1]:=strtofloat(yy.Text);
rst:=bp.computeout(x);
memo1.Lines.Append(floattostr(rst[0])+' '+floattostr(rst[1]));
end;
procedure TForm1.FormCreate(Sender: TObject);//初始化
begin
bp:=TBpdeep.Create([2,10,2],0.15,0.8);
end;
=====================================