FANN(Fast Artificial Neural Network Library)を使う～Delphiでお手軽プログラミング

当サイトで他にFANNを使用しているページ: FANNを使って最高気温をAIに予測させる; FANNを使って手書き数字(0～9)をAIに認識させる; FANNを使って手書き数字(0～9)の画像データセットMNISTを利用してみる; FANNを使って2つの1桁の整数の足し算を行わせる

FANNを使用する為のファイルの準備

FANNライブラリ、FANNライブラリ(fannfloat.dll 32Bit版)をDelphiから使用するfann.pas、fannライブラリをクラスとして使用できるライブラリファイル(MamFann.pas)を、本ページ下部からダウンロードします。

Delphiを起動して新規作成を行い、必要なコンポーネントをドラッグ&ドロップする

Delphi起動⇒ファイル⇒新規作成⇒WindowsVCLアプリケーションを選択します。
TButton 4個、 TLabel 1個、 TMemo 1個、 TImage 1個をフォームへドラッグ＆ドロップします。
Image1のプロパティ[Proportional]をTrueに、プロパティ[Stretch]をTrueにします。

プロジェクトフォルダ内にダウンロードしたfann.pas、MamFann.pasファイルを配置します。
fannfloat.dllファイル(32Bit版)をパスの通ったフォルダ(c:\Windows等)に入れます。

ソースコードを記述する

Form1のイベントOnCreateイベント、OnDestroyイベント、、 Button1のOnClickイベント、Button2のOnClickイベント、Button3のOnClickイベント、Button4のOnClickイベント、等に以下ソースコードを記述します。

unit Unit1;

interface

uses
  Winapi.Windows, Winapi.Messages, System.SysUtils, System.Variants, System.Classes, Vcl.Graphics,
  Vcl.Controls, Vcl.Forms, Vcl.Dialogs, Vcl.StdCtrls, Vcl.ExtCtrls,
  fann, MamFann ;

type
  TForm1 = class(TForm)
    Button1: TButton;
    Label1: TLabel;
    Memo1: TMemo;
    Button2: TButton;
    Button3: TButton;
    Button4: TButton;
    Image1: TImage;
    Label2: TLabel;
    procedure Button1Click(Sender: TObject);
    procedure FormDestroy(Sender: TObject);
    procedure Button2Click(Sender: TObject);
    procedure Button3Click(Sender: TObject);
    procedure Button4Click(Sender: TObject);
    procedure FormCreate(Sender: TObject);
  private
    { Private 宣言 }
    MamFann: TMamFann;
  public
    { Public 宣言 }
  end;

var
  Form1: TForm1;

implementation

{$R *.dfm}

procedure TForm1.Button1Click(Sender: TObject);
var NeuronNumInLayer:array of Cardinal;
begin
  //TMamFannクラスのインスタンス化
  if Assigned(MamFann) then
    FreeAndNil(MamFann);

  setlength(NeuronNumInLayer,3);//レイヤー(層)の数
  NeuronNumInLayer[0]:=2;       //入力層のニューロン数
  NeuronNumInLayer[1]:=3;       //中間層のニューロン数
  NeuronNumInLayer[2]:=1;       //出力層のニューロン数

  MamFann:=TMamFann.Create(NeuronNumInLayer);

  Button2.Enabled:=true;
  Button3.Enabled:=false;
  Button4.Enabled:=true;

  ShowMessage('ANNの初期設定完了');
end;
                                                                             //
procedure TForm1.Button2Click(Sender: TObject);
var inputs: array [0..1] of TFann_type;
    outputs: array [0..0] of TFann_type;
    epoch:Integer;   //一連の学習の実施回数
    i,j: Integer;
    Mse: single; //誤差分散
begin
  //一連の学習を3万回実施
  for epoch:=1 to 30000 do
  begin
    inputs[0]:=0;
    inputs[1]:=0;
    outputs[0]:=trunc(inputs[0]) Xor trunc(inputs[1]);
    MamFann.Train(inputs,outputs);

    inputs[0]:=1;
    inputs[1]:=0;
    outputs[0]:=trunc(inputs[0]) Xor trunc(inputs[1]);
    MamFann.Train(inputs,outputs);

    inputs[0]:=0;
    inputs[1]:=1;
    outputs[0]:=trunc(inputs[0]) Xor trunc(inputs[1]);
    MamFann.Train(inputs,outputs);

    inputs[0]:=1;
    inputs[1]:=1;
    outputs[0]:=trunc(inputs[0]) Xor trunc(inputs[1]);
    MamFann.Train(inputs,outputs);

    if epoch mod 100 = 0 then
    begin
      //誤差分散の表示
      Mse:=MamFann.GetMSE;
      Label1.Caption:=Format('%.8f',[mse]);
      Application.ProcessMessages;
    end;
  end;

  //学習結果をファイルに保存
  //ロード[NN.LoadFromFile('abc.net');]すれば学習結果を使用できる
  MamFann.SaveToFile('a.net');

  //ニューロンネットワーク図の描画
  MamFann.CreateNeuronBmp(Image1.Picture.Bitmap);

  Button2.Enabled:=true;
  Button3.Enabled:=true;
  Button4.Enabled:=true;

  ShowMessage('トレーニング終了'+#13#10+'学習ファイル保存');
end;

procedure TForm1.Button3Click(Sender: TObject);
var i,j: integer;
    outputs: array [0..0] of TFann_type;
    inputs : array [0..1] of TFann_type;
begin
  //XORのANN実行
  Memo1.Lines.Clear;

  inputs[0]:=0;
  inputs[1]:=0;
  MamFann.Run(inputs,outputs);
  Memo1.Lines.Add( Format('%1.0f XOR %1.0f = %f',[inputs[0],inputs[1],Outputs[0]]) );

  inputs[0]:=1;
  inputs[1]:=0;
  MamFann.Run(inputs,outputs);
  Memo1.Lines.Add( Format('%1.0f XOR %1.0f = %f',[inputs[0],inputs[1],Outputs[0]]) );

  inputs[0]:=0;
  inputs[1]:=1;
  MamFann.Run(inputs,outputs);
  Memo1.Lines.Add( Format('%1.0f XOR %1.0f = %f',[inputs[0],inputs[1],Outputs[0]]) );

  inputs[0]:=1;
  inputs[1]:=1;
  MamFann.Run(inputs,outputs);
  Memo1.Lines.Add( Format('%1.0f XOR %1.0f = %f',[inputs[0],inputs[1],Outputs[0]]) );
end;

procedure TForm1.Button4Click(Sender: TObject);
begin
  if Assigned(MamFann) then
    FreeAndNil(MamFann);
  if not FileExists('a.net') then exit;
  //学習済みファイルの読み込み
  MamFann:=TMamFann.Create('a.net');

  Button2.Enabled:=true;
  Button3.Enabled:=true;
  Button4.Enabled:=true;

  showmessage('ANNファイルの読み込み完了');
end;

procedure TForm1.FormCreate(Sender: TObject);
begin
  Image1.Proportional:=True;
  Image1.Stretch:=True;

  Button2.Enabled:=false;
  Button3.Enabled:=false;
  Button4.Enabled:=true;
end;

procedure TForm1.FormDestroy(Sender: TObject);
begin
  if Assigned(MamFann) then
    FreeAndNil(MamFann);
end;

end.

実行する

実行ボタンを押して実行します。（デバッグ実行でもOK）

fann.pasソースコード

unit fann;

interface

uses system.types;

Const FANN_DLLFILE = 'fannfloat.dll';

type
  TFann_type = Single;

  PFann_Type  = ^TFann_type;
  PPFann_Type = ^PFann_type;
  Fann_Type_Array = array[word] of TFann_type;
  PFann_Type_Array = ^Fann_Type_Array;
  PPFann_Type_Array = ^PFann_Type_Array;

  //struct _iobuf{
  //  char *_ptr; int _cnt; char *_base; int _flag;
  //  int _file; int _charbuf; int _bufsiz; char *_tmpfname;
  //}; typedef struct _iobuf FILE;
  _iobuf = packed record
    _ptr: PChar;
    _cnt: Integer;
    _base: PChar;
    _flag: Integer;
    _file: Integer;
    _charbuf: Integer;
    _bufsiz: Integer;
    _tmpfname: PChar;
  end;
  TFile = _iobuf;
  PFile = ^TFile;

  TEnumType=Cardinal;

  TFann_activationfunc_enum=(
    FANN_LINEAR,
    //閾値アクティブ化、0又は1の出力
    FANN_THRESHOLD,
    //閾値アクティブ化、-1又は1の出力
    FANN_THRESHOLD_SYMMETRIC,
    //シグモイド活性化、最も使用される、0～1の出力
    FANN_SIGMOID,
    //シグモイド段階的線形近似、シグモイドより高速で精度は劣る、0～1の出力
    FANN_SIGMOID_STEPWISE,
    //対称シグモイド活性化関数、最も使用される、-1～1の出力
    FANN_SIGMOID_SYMMETRIC,
    //対称ｼｸﾞﾓｲﾄﾞ段階的線形近似、対称ｼｸﾞﾓｲﾄﾞより高速で精度は劣る、-1～1の出力
    FANN_SIGMOID_SYMMETRIC_STEPWISE,
    FANN_GAUSSIAN,
    FANN_GAUSSIAN_SYMMETRIC,
    FANN_GAUSSIAN_STEPWISE,
    FANN_ELLIOT,
    FANN_ELLIOT_SYMMETRIC,
    FANN_LINEAR_PIECE,
    FANN_LINEAR_PIECE_SYMMETRIC,
    FANN_SIN_SYMMETRIC,
    FANN_COS_SYMMETRIC,
    FANN_SIN,
    FANN_COS
  );

  PFann_activationfunc_enum = ^Tfann_activationfunc_enum;


  TFann_Neuron = packed record
    first_con: Cardinal;
    last_con : Cardinal;
    sum: TFann_type;
    value: TFann_type;
    activation_steepness: TFann_type;
    activation_function: Cardinal ;
  end;
  PFann_Neuron = ^TFann_Neuron;
  PPFann_Neuron = ^PFann_Neuron;

  TFann_Layer = packed record
    first_neuron: PFann_Neuron;
    last_neuron : PFann_Neuron;
  end;
  PFann_Layer = ^TFann_Layer;

  TFann_nettype_enum = TEnumType;
  TFann_errno_enum = TEnumType;

  TFann_train_enum =(
    //重みが1つのエポックで何度も更新されるバックプロパゲーションアルゴリズム。
    //非常に速くトレーニングするが、高度な問題はうまくトレーニングしません。
    FANN_TRAIN_INCREMENTAL, //0
    //重みはエポック中に1回だけ更新されるので、トレーニングが遅くなります。
    //平均二乗誤差はより正確に計算され、FANN_TRAIN_INCREMENTALより良い。
    FANN_TRAIN_BATCH,       //1
    //適応型の為learning_rateを使用しません。iRPROPアルゴリズム
    FANN_TRAIN_RPROP,       //2
    //learning_rateを使用する、より高度なアルゴリズム。
    FANN_TRAIN_QUICKPROP,   //3
    //SARPROPアルゴリズム
    FANN_TRAIN_SARPROP      //4
  );
  TFann_errorfunc_enum=TEnumType;
  TFann_stopfunc_enum=TEnumType;

  PFann_Train_Data = ^TFann_Train_Data;
  TFann_Train_Data = packed record
    errno_f: TFann_errno_enum;
    erro_log: PFile;
    errstr: PChar;
    num_data: Cardinal;
    num_input: Cardinal;
    num_ouput: Cardinal;
    input: PPFann_Type_Array;
    output: PPFann_Type_Array;
  end;

  PFann = ^TFann;

  TFann_callback = function(
    ann: PFann;
    train: PFann_Train_Data;
    max_epochs: Cardinal;
    epochs_between_reports: Cardinal;
    desired_error: Single;
    epochs: Cardinal): Integer; stdcall;

  TUser_Function = procedure(
    num: Cardinal;
    num_input: Cardinal;
    num_output: Cardinal;
    input: PFann_Type;
    output: PFann_Type); stdcall;

  TFann = packed record
    errno_f: TFann_errno_enum;
    error_log: PFile;
    errstr: PChar;
    learning_rate: Single;
    learning_momentum: Single;
    connection_rate: Single;
    network_type: TFann_nettype_enum;
    first_layer: PFann_Layer;
    last_layer: PFann_Layer;
    total_neurons: Cardinal;
    num_input: Cardinal;
    num_output: Cardinal;
    weights: PFann_Type;
    connections: PPFann_Neuron;
    train_errors: PFann_Type;
    training_algorithm: TFann_train_enum;
    total_connections: Cardinal;
    output: PFann_Type;
    num_MSE: Cardinal;
    MSE_value: Single;
    num_bit_fail: Cardinal;
    bit_fail_limit: TFann_type;
    train_error_function: TFann_errorfunc_enum;
    train_stop_function: Tfann_stopfunc_enum;
    callback: TFann_callback;
    user_data: Pointer;
    cascade_output_change_fraction: single;
    cascade_output_stagnation_epochs: Cardinal;
    cascade_candidate_change_fraction: single;
    cascade_candidate_stagnation_epochs: Cardinal;
    cascade_best_candidate: Cardinal;
    cascade_candidate_limit: TFann_type;
    cascade_weight_multiplier: TFann_type;
    cascade_max_out_epochs: Cardinal;
    cascade_max_cand_epochs: Cardinal;
    cascade_activation_functions: TFann_activationfunc_enum;
    cascade_activation_functions_count: Cardinal;
    cascade_activation_steepnesses: PFann_Type;
    cascade_activation_steepnesses_count: Cardinal;
    cascade_num_candidate_groups: Cardinal;
    cascade_candidate_scores: PFann_Type;
    total_neurons_allocated: Cardinal;
    total_connections_allocated: Cardinal;
    quickprop_decay: single;
    quickprop_mu: single;
    rprop_increase_factor: single;
    rprop_decrease_factor: single;
    rprop_delta_min: single;
    rprop_delta_max: single;
    rprop_delta_zero: single;
    sarprop_temperature:single;
    sarprop_epoch: Cardinal;
    train_slopes: PFann_Type;
    prev_steps: PFann_Type;
    prev_train_slopes: PFann_Type;
    prev_weights_deltas: PFann_Type;
  end;

  TFann_Connection = packed record
    from_neuron: Cardinal;
    to_neuron: Cardinal;
    weight: TFann_type;
  end;
  PFann_Connection = ^TFann_Connection;

  TFann_Error = packed record
    errno_f: TFann_errno_enum;
    error_log: PFile;
    errstr: PChar;
  end;
  PFann_Error = ^TFann_Error;

Const
  //_Fann_Error_Func
  FANN_ERRORFUNC_LINEAR = 0;
  FANN_ERRORFUNC_TANH = 1;

  //_Fann_ErroNo
  FANN_E_NO_ERROR = 0;
  FANN_E_CANT_OPEN_CONFIG_R = 1;
  FANN_E_CANT_OPEN_CONFIG_W = 2;
  FANN_E_WRONG_CONFIG_VERSION = 3;
  FANN_E_CANT_READ_CONFIG = 4;
  FANN_E_CANT_READ_NEURON = 5;
  FANN_E_CANT_READ_CONNECTIONS = 6;
  FANN_E_WRONG_NUM_CONNECTIONS = 7;
  FANN_E_CANT_OPEN_TD_W = 8;
  FANN_E_CANT_OPEN_TD_R = 9;
  FANN_E_CANT_READ_TD = 10;
  FANN_E_CANT_ALLOCATE_MEM = 11;
  FANN_E_CANT_TRAIN_ACTIVATION = 12;
  FANN_E_CANT_USE_ACTIVATION = 13;
  FANN_E_TRAIN_DATA_MISMATCH = 14;
  FANN_E_CANT_USE_TRAIN_ALG = 15;
  FANN_E_TRAIN_DATA_SUBSET = 16;
  FANN_E_INDEX_OUT_OF_BOUND = 17;
  FANN_E_SCALE_NOT_PRESENT = 18;

  //_Fann_Stop_Func
  FANN_STOPFUNC_MSE = 0;
  FANN_STOPFUNC_BIT = 1;

  //_Fann_Net_Type
  FANN_NETTYPE_LAYER = 0;
  FANN_NETTYPE_SHORTCUT = 1;

  function fann_create_standard(num_layers: Cardinal): PFann; stdcall;
  function fann_create_sparse(
    connection_rate: single; num_layers: Cardinal):PFann; stdcall;
  function fann_create_shortcut(connection_rate: single): PFann; stdcall;
  function fann_create_standard_array(
    num_layers: Cardinal; const layers: PCardinal): PFann; stdcall;
  function fann_create_sparse_array(
    connection_rate: single; num_layers: Cardinal;
    const layers: PCardinal): PFann; stdcall;
  function fann_create_shortcut_array(
    num_layers: cardinal;const layers: Pcardinal): PFann; stdcall;
  procedure fann_destroy(Ann: PFann); stdcall;
  function fann_run(ann: PFann; input: PFann_Type): Pfann_type_array; stdcall;
  procedure fann_randomize_weights(
    Ann: PFann; Min_weight: TFann_type; Max_weight: TFann_type); stdcall;
  procedure fann_init_weights(
    Ann: PFann; train_data: PFann_Train_Data); stdcall;
  procedure fann_print_connections(ann: PFann);stdcall;
  procedure fann_print_parameters(ann: PFann);stdcall;
  function fann_get_num_input(Ann: PFann): cardinal;stdcall;
  function fann_get_num_output(Ann: PFann): cardinal;stdcall;
  function fann_get_total_neurons(Ann: PFann): cardinal; stdcall;
  function fann_get_total_connections(Ann: PFann): cardinal; stdcall;
  function fann_get_network_type(Ann: PFann): cardinal; stdcall;
  function fann_get_connection_rate(Ann: PFann): single; stdcall;
  function fann_get_num_layers(Ann: PFann): cardinal; stdcall;
  procedure fann_get_layer_array(Ann: PFann; layers: PCardinal); stdcall;
  procedure fann_get_bias_array(Ann: PFann; bias: PCardinal);stdcall;
  procedure fann_get_connection_array(
    Ann: PFann; connections: PFann_Connection);stdcall;
  procedure fann_set_weight_array(Ann: PFann;
    connections: PFann_Connection; num_connection: Cardinal);stdcall;
  procedure fann_set_weight(Ann: PFann;
    from_neuron: Cardinal; to_neuron: Cardinal; weight: TFann_type);stdcall;
  procedure fann_set_user_data(Ann: PFann; user_data: Pointer);stdcall;
  function fann_get_user_data(Ann: PFann): Pointer; stdcall;
  function fann_create_from_file(
    const configuration_file: PAnsiChar): PFann; stdcall;
  procedure fann_save(Ann: PFann; Const Configuration_File: PAnsiChar);stdcall;
  function fann_save_to_fixed(
    Ann: PFann; Const Configuration_File: PChar): integer;stdcall;
  procedure fann_train(
    Ann: PFann; Input: PFann_Type; Desired_Output: PFann_Type);stdcall;
  function fann_test(
    Ann: PFann; Input: PFann_Type;
    Desired_Output: Pfann_Type): Pfann_type_array;stdcall;
  function fann_get_MSE(Ann: PFann): single;stdcall;
  function fann_get_bit_fail(Ann: PFann): Cardinal;stdcall;
  procedure fann_reset_MSE(Ann: Pfann); stdcall;
  procedure fann_train_on_data(
    Ann: PFann; Data: PFann_Train_Data;
    max_epochs: cardinal;epochs_between_reports: cardinal;
    desired_error: single);stdcall;
  procedure fann_train_on_file(
    Ann: PFann; Filename: PAnsiChar;max_epochs: cardinal;
    epochs_between_reports: cardinal; desired_error: single); stdcall;
  function fann_train_epoch(
    Ann: PFann; data: PFann_Train_Data): single; stdcall;
  function fann_test_data(Ann: PFann; data: PFann_Train_Data): single; stdcall;
  function fann_read_train_from_file(
    const filename: PChar): PFann_Train_Data; stdcall;
  function fann_create_train_from_callback(
    num_data: Cardinal; num_input: Cardinal;
    num_output: Cardinal;
    user_function: TUser_Function): PFann_Train_Data; stdcall;
  procedure fann_destroy_train(train_data: PFann_Train_Data); stdcall;
  procedure fann_shuffle_train_data(Train_Data: PFann_Train_Data);stdcall;
  procedure fann_scale_train(Ann: PFann; data: PFann_Train_Data);stdcall;
  procedure fann_descale_train(Ann: PFann; data: PFann_Train_Data);stdcall;
  function fann_set_input_scaling_params(
    Ann: PFann; const data: PFann_Train_Data;
    new_input_min: single; new_input_max: single): integer;stdcall;
  function fann_set_output_scaling_params(
    Ann: PFann; const data: PFann_Train_Data;
    new_output_min: single; new_output_max: single): integer;stdcall;
  function fann_set_scaling_params(
    Ann: PFann; const data: PFann_Train_Data;
    new_input_min: single; new_input_max: single;
    new_output_min: single; new_output_max: single): integer; stdcall;
  function fann_clear_scaling_params(Ann: PFann): integer; stdcall;
  procedure fann_scale_input(Ann: PFann; input_vector: PFann_type); stdcall;
  procedure fann_scale_output(Ann: PFann; output_vector: PFann_type); stdcall;
  procedure fann_descale_input(Ann: PFann; input_vector: PFann_type); stdcall;
  procedure fann_descale_output(Ann: PFann; output_vector: PFann_type); stdcall;
  procedure fann_scale_input_train_data(
    Train_Data: PFann_Train_Data; new_min: TFann_type;
    new_max: TFann_type); stdcall;
  procedure fann_scale_output_train_data(
    Train_Data: PFann_Train_Data; new_min: TFann_type;
    new_max: TFann_type); stdcall;
  procedure fann_scale_train_data(
    Train_Data: PFann_Train_Data; new_min: TFann_type;
    new_max: TFann_type); stdcall;
  function fann_merge_train_data(
    Data1: PFann_Train_Data; Data2:
    PFann_Train_Data): PFann_Train_Data; stdcall;
  function fann_duplicate_train_data(
    Data: PFann_Train_Data): PFann_Train_Data;stdcall;
  function fann_subset_train_data(
    data: PFann_Train_Data; pos: Cardinal;
    length: Cardinal): PFann_Train_Data; stdcall;
  function fann_length_train_data(
    data: PFann_Train_Data): Cardinal; stdcall;
  function fann_num_input_train_data(
    data: PFann_Train_Data): Cardinal; stdcall;
  function fann_num_output_train_data(
    data: PFann_Train_Data): Cardinal; stdcall;
  function fann_save_train(
    Data: PFann_train_Data; const Filename: PChar):integer;stdcall;
  function fann_save_train_to_fixed(
    Data: PFann_train_Data; const FileName: Pchar;
    decimal_point: cardinal): integer;stdcall;
  function fann_get_training_algorithm(Ann: Pfann):Cardinal;stdcall;
  procedure fann_set_training_algorithm(
    Ann: PFann; Training_Algorithm: Cardinal);stdcall;
  function fann_get_learning_rate(Ann: PFann): single;stdcall;
  procedure fann_set_learning_rate(
    Ann: PFann; Learning_Rate: Single); stdcall;
  function fann_get_learning_momentum(Ann: PFann): single;stdcall;
  procedure fann_set_learning_momentum(
    Ann: PFann; learning_momentum: Single); stdcall;
  function fann_get_activation_function(
    Ann: PFann; layer: integer;
    neuron: integer): TFann_activationfunc_enum; stdcall;
  procedure fann_set_activation_function(
    Ann: PFann; activation_function: Cardinal;
    layer: integer; neuron: integer); stdcall;
  procedure fann_set_activation_function_layer(
    Ann: PFann; activation_function: Cardinal;
    layer: integer); stdcall;
  procedure fann_set_activation_function_hidden(
    Ann: PFann; activation_function: Cardinal);  stdcall;
  procedure fann_set_activation_function_output(
    Ann: PFann; activation_function: Cardinal); stdcall;
  function fann_get_activation_steepness(
    Ann: PFann; layer: integer; neuron: integer): TFann_type; stdcall;
  procedure fann_set_activation_steepness(
    Ann: PFann; steepness: TFann_type;
    layer: integer; neuron: integer); stdcall;
  procedure fann_set_activation_steepness_layer(
    Ann: PFann; steepness: TFann_type; layer: integer); stdcall;
  procedure fann_set_activation_steepness_hidden(
    Ann: PFann; steepness: TFann_type); stdcall;
  procedure fann_set_activation_steepness_output(
    Ann: PFann; steepness: TFann_type); stdcall;
  function fann_get_train_error_function(Ann: PFann): cardinal;stdcall;
  procedure fann_set_train_error_function(
    Ann: PFann; Train_Error_Function: cardinal); stdcall;
  function fann_get_train_stop_function(Ann: PFann): Cardinal; stdcall;
  procedure fann_set_train_stop_function(
    Ann: PFann; train_stop_function: cardinal); stdcall;
  function fann_get_bit_fail_limit(Ann: PFann): TFann_type; stdcall;
  procedure fann_set_bit_fail_limit(
    Ann: PFann; bit_fail_limit: TFann_type); stdcall;
  procedure fann_set_callback(Ann: PFann; callback: TFann_Callback); stdcall;
  function fann_get_quickprop_decay(Ann: PFann): single;stdcall;
  procedure fann_set_quickprop_decay(
    Ann: Pfann; quickprop_decay: Single);stdcall;
  function fann_get_quickprop_mu(Ann: PFann): single;stdcall;
  procedure fann_set_quickprop_mu(Ann: PFann; Mu: Single);stdcall;
  function fann_get_rprop_increase_factor(Ann: PFann): single;stdcall;
  procedure fann_set_rprop_increase_factor(
    Ann: PFann;rprop_increase_factor: single);stdcall;
  function fann_get_rprop_decrease_factor(
    Ann: PFann): single;stdcall;
  procedure fann_set_rprop_decrease_factor(
    Ann: PFann;rprop_decrease_factor: single); stdcall;
  function fann_get_rprop_delta_min(Ann: PFann): single; stdcall;
  procedure fann_set_rprop_delta_min(
    Ann: PFann; rprop_delta_min: Single); stdcall;
  function fann_get_rprop_delta_max(Ann: PFann): single;stdcall;
  procedure fann_set_rprop_delta_max(
    Ann: PFann; rprop_delta_max: Single); stdcall;
  function fann_get_rprop_delta_zero(Ann: PFann): single;stdcall;
  procedure fann_set_rprop_delta_zero(
    Ann: PFann; rprop_delta_zero: Single); stdcall;
  procedure fann_set_error_log(errdat: PFann_Error; Log_File: PFile);stdcall;
  function fann_get_errno(errdat: PFann_Error): cardinal;stdcall;
  procedure fann_reset_errno(errdat: PFann_Error);stdcall;
  procedure fann_reset_errstr(errdat: PFann_Error);stdcall;
  function fann_get_errstr(errdat: PFann_Error): PChar;stdcall;
  procedure fann_print_error(Errdat: PFann_Error);stdcall;
  procedure fann_cascadetrain_on_data(
    Ann: PFann; data: PFann_Train_Data;
    max_neurons: Cardinal; neurons_between_reports: Cardinal;
    desired_error: single); stdcall;
  procedure fann_cascadetrain_on_file(Ann: PFann; const filename: PChar;
    max_neurons: Cardinal; neurons_between_reports: Cardinal;
    desired_error: single); stdcall;
  function fann_get_cascade_output_change_fraction(
    Ann: PFann): single; stdcall;
  procedure fann_set_cascade_output_change_fraction(
    Ann: PFann; cascade_output_change_fraction: single); stdcall;
  function fann_get_cascade_output_stagnation_epochs(
    Ann: PFann): cardinal; stdcall;
  procedure fann_set_cascade_output_stagnation_epochs(
    Ann: PFann; cascade_output_stagnation_epochs: cardinal); stdcall;
  function fann_get_cascade_candidate_change_fraction(
    Ann: PFann): single; stdcall;
  procedure fann_set_cascade_candidate_change_fraction(
    Ann: PFann; cascade_candidate_change_fraction: single); stdcall;
  function fann_get_cascade_candidate_stagnation_epochs(
    Ann: PFann): cardinal; stdcall;
  procedure fann_set_cascade_candidate_stagnation_epochs(
    Ann: PFann; cascade_candidate_stagnation_epochs: cardinal); stdcall;
  function fann_get_cascade_weight_multiplier(
    Ann: PFann): TFann_type; stdcall;
  procedure fann_set_cascade_weight_multiplier(
    Ann: PFann; cascade_weight_multiplier: TFann_type); stdcall;
  function fann_get_cascade_candidate_limit(
    Ann: PFann): TFann_type; stdcall;
  procedure fann_set_cascade_candidate_limit(
    Ann: PFann; cascade_candidate_limit: TFann_type); stdcall;

  function fann_get_cascade_max_out_epochs(Ann: PFann): cardinal; stdcall;
  procedure fann_set_cascade_max_out_epochs(
    Ann: PFann; cascade_max_out_epochs: cardinal); stdcall;
  function fann_get_cascade_max_cand_epochs(
    Ann: PFann): cardinal; stdcall;
  procedure fann_set_cascade_max_cand_epochs(
    Ann: PFann; cascade_max_cand_epochs: cardinal); stdcall;
  function fann_get_cascade_num_candidates(
    Ann: PFann): cardinal; stdcall;
  function fann_get_cascade_activation_functions_count(
    Ann: PFann): cardinal; stdcall;
  function fann_get_cascade_activation_functions(
    Ann: PFann): PCardinal; stdcall;
  procedure fann_set_cascade_activation_functions(Ann: PFann;
    cascade_activation_functions: PCardinal;
    cascade_activation_functions_count: Cardinal); stdcall;
  function fann_get_cascade_activation_steepnesses_count(
    Ann: PFann): cardinal; stdcall;
  function fann_get_cascade_activation_steepnesses(
    Ann: PFann): pfann_type; stdcall;
  procedure fann_set_cascade_activation_steepnesses(
    Ann: PFann; cascade_activation_steepnesses: PFann_Type;

    cascade_activation_steepnesses_count: Cardinal); stdcall;

  function fann_get_cascade_num_candidate_groups(Ann: PFann): cardinal; stdcall;
  procedure fann_set_cascade_num_candidate_groups(
    Ann: PFann; cascade_num_candidate_groups: cardinal); stdcall;

implementation

  function fann_create_standard; external FANN_DLLFILE;
  function fann_create_sparse; external FANN_DLLFILE;
  function fann_create_shortcut; external FANN_DLLFILE;
  function fann_create_standard_array;
    external FANN_DLLFILE name '_fann_create_standard_array@8';
  function fann_create_sparse_array;
    external FANN_DLLFILE name '_fann_create_sparse_array@12';
  function fann_create_shortcut_array; external FANN_DLLFILE;
  procedure fann_destroy(Ann: PFann);
    external FANN_DLLFILE name '_fann_destroy@4';
  function fann_run; external FANN_DLLFILE name '_fann_run@8';
  procedure fann_randomize_weights;
    external FANN_DLLFILE name '_fann_randomize_weights@12';
  procedure fann_init_weights; external FANN_DLLFILE;
  procedure fann_print_connections; external FANN_DLLFILE;
  procedure fann_print_parameters; external FANN_DLLFILE;
  function fann_get_num_input; external FANN_DLLFILE;
  function fann_get_num_output; external FANN_DLLFILE;
  function fann_get_total_neurons; external FANN_DLLFILE;
  function fann_get_total_connections;
    external FANN_DLLFILE name '_fann_get_total_connections@4';
  function fann_get_network_type; external FANN_DLLFILE;
  function fann_get_connection_rate;
    external FANN_DLLFILE name '_fann_get_connection_rate@4';
  function fann_get_num_layers;
    external FANN_DLLFILE name '_fann_get_num_layers@4';
  procedure fann_get_layer_array;
    external FANN_DLLFILE name '_fann_get_layer_array@8';
  procedure fann_get_bias_array;
    external FANN_DLLFILE name '_fann_get_bias_array@8';
  procedure fann_get_connection_array;
    external FANN_DLLFILE name '_fann_get_connection_array@8';
  procedure fann_set_weight_array; external FANN_DLLFILE;
  procedure fann_set_weight; external FANN_DLLFILE;
  procedure fann_set_user_data; external FANN_DLLFILE;
  function fann_get_user_data; external FANN_DLLFILE;
  function fann_create_from_file;
    external FANN_DLLFILE name '_fann_create_from_file@4';
  procedure fann_save; external FANN_DLLFILE name '_fann_save@8';
  function fann_save_to_fixed; external FANN_DLLFILE;
  procedure fann_train; external FANN_DLLFILE name '_fann_train@12';
  function fann_test; external FANN_DLLFILE;
  function fann_get_MSE; external FANN_DLLFILE name '_fann_get_MSE@4';
  function fann_get_bit_fail; external FANN_DLLFILE;
  procedure fann_reset_MSE; external FANN_DLLFILE name '_fann_reset_MSE@4';
  procedure fann_train_on_data;
    external FANN_DLLFILE name '_fann_train_on_data@20';
  procedure fann_train_on_file;
    external FANN_DLLFILE name '_fann_train_on_file@20';
  function fann_train_epoch; external FANN_DLLFILE;
  function fann_test_data; external FANN_DLLFILE;
  function fann_read_train_from_file; external FANN_DLLFILE;
  function fann_create_train_from_callback; external FANN_DLLFILE;
  procedure fann_destroy_train; external FANN_DLLFILE;
  procedure fann_shuffle_train_data; external FANN_DLLFILE;
  procedure fann_scale_train; external FANN_DLLFILE;
  procedure fann_descale_train; external FANN_DLLFILE;
  function fann_set_input_scaling_params; external FANN_DLLFILE;
  function fann_set_output_scaling_params; external FANN_DLLFILE;
  function fann_set_scaling_params; external FANN_DLLFILE;
  function fann_clear_scaling_params; external FANN_DLLFILE;
  procedure fann_scale_input; external FANN_DLLFILE;
  procedure fann_scale_output; external FANN_DLLFILE;
  procedure fann_descale_input; external FANN_DLLFILE;
  procedure fann_descale_output; external FANN_DLLFILE;
  procedure fann_scale_input_train_data; external FANN_DLLFILE;
  procedure fann_scale_output_train_data; external FANN_DLLFILE;
  procedure fann_scale_train_data; external FANN_DLLFILE;
  function fann_merge_train_data; external FANN_DLLFILE;
  function fann_duplicate_train_data; external FANN_DLLFILE;
  function fann_subset_train_data; external FANN_DLLFILE;
  function fann_length_train_data; external FANN_DLLFILE;
  function fann_num_input_train_data; external FANN_DLLFILE;
  function fann_num_output_train_data; external FANN_DLLFILE;
  function fann_save_train; external FANN_DLLFILE;
  function fann_save_train_to_fixed; external FANN_DLLFILE;
  function fann_get_training_algorithm;
    external FANN_DLLFILE name '_fann_get_training_algorithm@4';
  procedure fann_set_training_algorithm;
    external FANN_DLLFILE name '_fann_set_training_algorithm@8';
  function fann_get_learning_rate;
    external FANN_DLLFILE name '_fann_get_learning_rate@4';
  procedure fann_set_learning_rate;
    external FANN_DLLFILE name '_fann_set_learning_rate@8';
  function fann_get_learning_momentum;
    external FANN_DLLFILE name '_fann_get_learning_momentum@4';
  procedure fann_set_learning_momentum;
    external FANN_DLLFILE name '_fann_set_learning_momentum@8';
  function fann_get_activation_function; external FANN_DLLFILE;
  procedure fann_set_activation_function;
    external FANN_DLLFILE name '_fann_set_activation_function@16';
  procedure fann_set_activation_function_layer; external FANN_DLLFILE;
  procedure fann_set_activation_function_hidden;
    external FANN_DLLFILE name '_fann_set_activation_function_hidden@8';
  procedure fann_set_activation_function_output;
    external FANN_DLLFILE name '_fann_set_activation_function_output@8';
  function fann_get_activation_steepness; external FANN_DLLFILE;
  procedure fann_set_activation_steepness;
    external FANN_DLLFILE name '_fann_set_activation_steepness@16';
  procedure fann_set_activation_steepness_layer;
    external FANN_DLLFILE name '_fann_set_activation_steepness_layer@12';
  procedure fann_set_activation_steepness_hidden;
    external FANN_DLLFILE name '_fann_set_activation_steepness_hidden@8';
  procedure fann_set_activation_steepness_output;
    external FANN_DLLFILE name '_fann_set_activation_steepness_output@8';
  function fann_get_train_error_function; external FANN_DLLFILE;
  procedure fann_set_train_error_function; external FANN_DLLFILE;
  function fann_get_train_stop_function; external FANN_DLLFILE;
  procedure fann_set_train_stop_function; external FANN_DLLFILE;
  function fann_get_bit_fail_limit; external FANN_DLLFILE;
  procedure fann_set_bit_fail_limit; external FANN_DLLFILE;
  procedure fann_set_callback; external FANN_DLLFILE;
  function fann_get_quickprop_decay; external FANN_DLLFILE;
  procedure fann_set_quickprop_decay; external FANN_DLLFILE;
  function fann_get_quickprop_mu; external FANN_DLLFILE;
  procedure fann_set_quickprop_mu; external FANN_DLLFILE;
  function fann_get_rprop_increase_factor; external FANN_DLLFILE;
  procedure fann_set_rprop_increase_factor; external FANN_DLLFILE;
  function fann_get_rprop_decrease_factor; external FANN_DLLFILE;
  procedure fann_set_rprop_decrease_factor; external FANN_DLLFILE;
  function fann_get_rprop_delta_min; external FANN_DLLFILE;
  procedure fann_set_rprop_delta_min; external FANN_DLLFILE;
  function fann_get_rprop_delta_max; external FANN_DLLFILE;
  procedure fann_set_rprop_delta_max; external FANN_DLLFILE;
  function fann_get_rprop_delta_zero; external FANN_DLLFILE;
  procedure fann_set_rprop_delta_zero; external FANN_DLLFILE;
  procedure fann_set_error_log; external FANN_DLLFILE;
  function fann_get_errno; external FANN_DLLFILE;
  procedure fann_reset_errno; external FANN_DLLFILE;
  procedure fann_reset_errstr; external FANN_DLLFILE;
  function fann_get_errstr; external FANN_DLLFILE;
  procedure fann_print_error; external FANN_DLLFILE;
  procedure fann_cascadetrain_on_data; external FANN_DLLFILE;
  procedure fann_cascadetrain_on_file; external FANN_DLLFILE;
  function fann_get_cascade_output_change_fraction; external FANN_DLLFILE;
  procedure fann_set_cascade_output_change_fraction; external FANN_DLLFILE;
  function fann_get_cascade_output_stagnation_epochs; external FANN_DLLFILE;
  procedure fann_set_cascade_output_stagnation_epochs; external FANN_DLLFILE;
  function fann_get_cascade_candidate_change_fraction; external FANN_DLLFILE;
  procedure fann_set_cascade_candidate_change_fraction; external FANN_DLLFILE;
  function fann_get_cascade_candidate_stagnation_epochs; external FANN_DLLFILE;
  procedure fann_set_cascade_candidate_stagnation_epochs; external FANN_DLLFILE;
  function fann_get_cascade_weight_multiplier; external FANN_DLLFILE;
  procedure fann_set_cascade_weight_multiplier; external FANN_DLLFILE;
  function fann_get_cascade_candidate_limit; external FANN_DLLFILE;
  procedure fann_set_cascade_candidate_limit; external FANN_DLLFILE;

  function fann_get_cascade_max_out_epochs; external FANN_DLLFILE;
  procedure fann_set_cascade_max_out_epochs; external FANN_DLLFILE;
  function fann_get_cascade_max_cand_epochs; external FANN_DLLFILE;
  procedure fann_set_cascade_max_cand_epochs; external FANN_DLLFILE;
  function fann_get_cascade_num_candidates; external FANN_DLLFILE;
  function fann_get_cascade_activation_functions_count; external FANN_DLLFILE;
  function fann_get_cascade_activation_functions; external FANN_DLLFILE;
  procedure fann_set_cascade_activation_functions; external FANN_DLLFILE;
  function fann_get_cascade_activation_steepnesses_count; external FANN_DLLFILE;
  function fann_get_cascade_activation_steepnesses; external FANN_DLLFILE;
  procedure fann_set_cascade_activation_steepnesses; external FANN_DLLFILE;

  function fann_get_cascade_num_candidate_groups; external FANN_DLLFILE;
  procedure fann_set_cascade_num_candidate_groups; external FANN_DLLFILE;

end.

MamFann.pasソースコード

unit MamFann;

interface

uses fann, System.Classes, System.SysUtils, VCL.Graphics,System.Types;

Type

  TMamFann=class
  private
    FFann:PFann;
    FConnectionRate: Single;//接続割合
    FLearningRate: Single;  //学習割合
    FLearningMomentum: Single;
    FActivationFunctionHidden:TFann_activationfunc_enum;
    FActivationFunctionOutput:TFann_activationfunc_enum;
    FTrainingAlgorithm:TFann_train_enum;

    //学習慣性
    procedure SetLearningMomentum(Const Value: Single);
    //学習割合
    procedure SetLearningRate(Const Value:Single);
    //学習アルゴリズム
    procedure SetTrainingAlgorithm(Const Value:TFann_train_enum);
    //非表示レイヤーのアクティベーション関数を設定
    procedure SetActivationFunctionHidden(Const Value: TFann_activationfunc_enum);
    //出力層のアクティベーション関数を設定
    procedure SetActivationFunctionOutput(Const Value: TFann_activationfunc_enum);

  protected
  public
    //コンストラスター(接続割合,各層のニューロン数)
    //  接続割合:
    //    1だと完全に接続されたネットワーク(fann_create_standard_array)
    //    1未満だと完全に接続されないネットワーク(fann_create_sparse_array)
    //  各層のニューロン数:各層のニューロン数を与える
    constructor Create(NeuronNumInLayer: array of Cardinal;
      ConnectionRate:Single=1.0);overload;
    //コンストラクター(ファイル名)
    //  fann_create_from_fileを使用する場合
    constructor Create(LoadFromFileName:String);overload;
    destructor Destroy();override;
    //誤差分散の取得
    function GetMSE():Single;
    //誤差分散のリセット
    procedure ResetMse();
    //トレーニング
    procedure Train(Inputs,Outputs: array of TFann_type);
    procedure TrainOnFile(FileName: String; MaxEpochs: Cardinal; DesiredError: Single);
    //ANN実行
    procedure Run(Inputs: array of TFann_type; var Outputs: array of TFann_type);

    procedure SaveToFile(FileName: String);
    //ファイルからANNを読み込む
    procedure LoadFromFile(Filename: string);
    //ネットワーク接続数を取得する
    function GetTotalConnections():Cardinal;
    //ネットワーク接続を取得する
    procedure GetConnectionArray(pcns:PFann_Connection);
    //レイヤーの数を取得する
    function GetNumLayers():Cardinal;
    procedure GetLayerArray(pc:PCardinal);
    //バイアスを取得する
    procedure GetBiasArray(pc:PCardinal);
    //ニューロンネットワーク図の作成
    procedure CreateNeuronBmp(bmp:TBitmap);
    function GetTrainingAlgorithm():Cardinal;
    //学習慣性(デフォルト0.0、与える場合は0.9程度らしい)
    property LearningMomentum:Single
      read FLearningMomentum write SetLearningMomentum;
    property LearningRate:single
      read FLearningRate write SetLearningRate;
    property TrainingAlgorithm:TFann_train_enum
      read FTrainingAlgorithm write SetTrainingAlgorithm;
    property ActivationFunctionHidden:TFann_activationfunc_enum
      read FActivationFunctionHidden write setActivationFunctionHidden;
    property ActivationFunctionOutput:TFann_activationfunc_enum
      read FActivationFunctionOutput write setActivationFunctionOutput;
  end;

implementation

{ TMamFann }

constructor TMamFann.Create(
  NeuronNumInLayer: array of Cardinal;ConnectionRate:Single=1.0);
var LayerNum: Cardinal; //レイヤーの数
begin
  if Length(NeuronNumInLayer)<2 then
    raise Exception.Create('２レイヤー以上が必要です');
  FConnectionRate  :=ConnectionRate;//接続割合
  FLearningRate    :=0.7;//学習割合
  FLearningMomentum:=0.0;//学習慣性
  //全ての非表示レイヤーのアクティベーション関数の設定
  FActivationFunctionHidden:=TFann_activationfunc_enum.FANN_SIGMOID;
  //全ての出力レイヤーのアクティベーション関数の設定
  FActivationFunctionOutput:=TFann_activationfunc_enum.FANN_SIGMOID;
  //練習アルゴリズム
  FTrainingAlgorithm:=TFann_train_enum.FANN_TRAIN_RPROP;
  LayerNum:=Length(NeuronNumInLayer);

  if(FConnectionRate<1) then
    FFann:=fann_create_sparse_array(
      FConnectionRate,LayerNum,@NeuronNumInLayer[0])
  else
    FFann:=fann_create_standard_array(LayerNum,@NeuronNumInLayer[0]);

  fann_randomize_weights(FFann,-0.5,0.5);
  fann_set_learning_rate(FFann,FLearningRate);
  fann_set_learning_momentum(FFann,FLearningMomentum);
  fann_set_activation_steepness_hidden(FFann,0.5);
  fann_set_activation_steepness_output(FFann,0.5);
  fann_set_activation_function_hidden(
    FFann,Cardinal(ord(FActivationFunctionHidden)));
  fann_set_activation_function_output(
    FFann,Cardinal(ord(FActivationFunctionOutput)));
  fann_set_training_algorithm(
    FFann,Cardinal(ord(FTrainingAlgorithm)));

  self.ResetMse;
end;

constructor TMamFann.Create(LoadFromFileName: String);
begin
  self.LoadFromFile(LoadFromFileName);
end;

procedure TMamFann.CreateNeuronBmp(bmp: TBitmap);
//ニューロンネットワーク図の作成
Type
  TNeuron=record
    x,y:integer;
    isBias:Boolean;
  end;
var ConnectionNum: Cardinal;//ネットワークの接続数
    Connections: array of TFann_Connection;//ネットワークの接続
    LayerNum: Cardinal;//レイヤー(層)数
    NeuronNums:array of Cardinal; //各レイヤー(層)ごとのニューロン数
    BiasNums:array of Cardinal;   //各レイヤー(層)ごとのバイアスの数
    Neurons:array of TNeuron;
    NeuronCount:integer;
    i,j:integer;
    max_x,max_y:integer;
    FromPoint,ToPoint:TPoint;
    ConWeight:integer;
    ConColor:TColor;
const
  span:integer=100;
begin
  //ネットワークの接続数を取得
  ConnectionNum:=Self.GetTotalConnections();
  SetLength(Connections,ConnectionNum);
  //ネットワーク内の接続を取得します
  Self.GetConnectionArray(@Connections[0]);

  //レイヤー数の取得(バイアス含まず)
  LayerNum:=Self.GetNumLayers();
  SetLength(NeuronNums, LayerNum);
  SetLength(BiasNums  , LayerNum);
  //各レイヤーのニューロン数を取得(バイアス含まず)
  Self.GetLayerArray(@NeuronNums[0]);
  //各レイヤーのバイアス数を取得
  Self.GetBiasArray(@BiasNums[0]);

  NeuronCount:=0;
  for i := 0 to LayerNum-1 do
  begin
    for j := 0 to NeuronNums[i]-1 do
    begin
      inc(NeuronCount);
      setLength(Neurons,NeuronCount);
      Neurons[NeuronCount-1].x:=j;
      Neurons[NeuronCount-1].y:=i;
      Neurons[NeuronCount-1].isBias:=false;
    end;
    for j := 0 to BiasNums[i]-1 do
    begin
      inc(NeuronCount);
      setLength(Neurons,NeuronCount);
      Neurons[NeuronCount-1].x:=NeuronNums[i]+j;
      Neurons[NeuronCount-1].y:=i;
      Neurons[NeuronCount-1].isBias:=True;
    end;
  end;

  max_y:=LayerNum;
  max_x:=0;
  for i := 0 to Length(Neurons)-1 do
  begin
    if max_x<Neurons[i].x then
      max_x:=Neurons[i].x;
  end;
  inc(max_x);
  if not Assigned(bmp) then
    bmp:=TBitmap.Create;

  bmp.Width :=max_x*span;
  bmp.Height:=max_y*span;

  bmp.Canvas.Brush.Color:=$FFFFFF;
  bmp.Canvas.Brush.Style:=TBrushStyle.bsSolid;
  bmp.Canvas.FillRect(Rect(0,0,bmp.Width,bmp.Height));

  for i := 0 to ConnectionNum-1 do
  begin
    FromPoint.X:=Neurons[Connections[i].from_neuron].x;
    FromPoint.Y:=Neurons[Connections[i].from_neuron].y;
    ToPoint.X  :=Neurons[Connections[i].to_neuron].x;
    ToPoint.Y  :=Neurons[Connections[i].to_neuron].y;
    ConWeight  :=trunc(Connections[i].weight);
    if ConWeight>0 then
      ConColor:=$000000
    else
      ConColor:=$0000FF;
    ConWeight:=ABS(ConWeight);
    if ConWeight=0 then ConWeight:=1;
    bmp.Canvas.Pen.Color:=ConColor;
    bmp.Canvas.Pen.Width:=ConWeight;
    bmp.Canvas.MoveTo(
      FromPoint.X*span + span div 2, FromPoint.Y*span + span div 2);
    bmp.Canvas.LineTo(
      ToPoint.X  *span + span div 2, ToPoint.Y  *span + span div 2);
  end;
  for i := 0 to NeuronCount-1 do
  begin
    if Neurons[i].isBias then
      ConColor:=$0099FF
    else
      ConColor:=$FF0000;

    bmp.Canvas.Pen.Color:=ConColor;
    bmp.Canvas.Brush.Color:=ConColor;
    bmp.Canvas.Ellipse(
      Neurons[i].x*span + span div 2 - span div 8,
      Neurons[i].y*span + span div 2 - span div 8,
      Neurons[i].x*span + span div 2 + span div 8,
      Neurons[i].y*span + span div 2 + span div 8
    );
  end;
end;

destructor TMamFann.Destroy;
begin
  fann_destroy(FFann);
  inherited;
end;


procedure TMamFann.GetBiasArray(pc: PCardinal);
begin
  fann_get_bias_array(FFann,pc);
end;

procedure TMamFann.GetConnectionArray(pcns: PFann_Connection);
begin
  fann_get_connection_array(FFann,pcns);
end;

procedure TMamFann.GetLayerArray(pc: PCardinal);
begin
  fann_get_layer_array(FFann,pc);
end;

function TMamFann.GetMSE: Single;
begin
  result:=fann_get_MSE(FFann);
end;

function TMamFann.GetNumLayers: Cardinal;
begin
  result:=fann_get_num_layers(FFann);
end;

function TMamFann.GetTotalConnections: Cardinal;
begin
  result:=fann_get_total_connections(FFann);
end;

function TMamFann.GetTrainingAlgorithm: Cardinal;
begin
  result:=fann_get_training_algorithm(FFann);
end;

procedure TMamFann.LoadFromFile(Filename: string);
begin
  if Assigned(FFann) then
    fann_destroy(FFann);
  FFann:=fann_create_from_file(PAnsiChar(AnsiString(Filename)));
end;

procedure TMamFann.ResetMse;
begin
  fann_reset_mse(FFann);
end;

procedure TMamFann.Run(
  Inputs: array of TFann_type;var Outputs: array of TFann_type);
var out_array: Pfann_type_array;
    i: Integer;
begin
  out_array:=fann_run(FFann,@Inputs[0]);
  for i:=0 to High(outputs) do
  begin
    Outputs[i]:=out_array[i];
  end;
end;

procedure TMamFann.SaveToFile(FileName: String);
begin
  fann_save(FFann,PAnsiChar(AnsiString(Filename)));
end;

procedure TMamFann.SetActivationFunctionHidden(
  const Value: TFann_activationfunc_enum);
begin
  FActivationFunctionHidden:=Value;
  fann_set_activation_function_hidden(
    FFann,Cardinal(ord(FActivationFunctionHidden)));
end;

procedure TMamFann.SetActivationFunctionOutput(
  const Value: TFann_activationfunc_enum);
begin
  FActivationFunctionOutput := Value;
  fann_set_activation_function_output(
    FFann,Cardinal(ord(FActivationFunctionOutput)));
end;

procedure TMamFann.SetLearningMomentum(const Value: Single);
begin
  FLearningMomentum:=Value;
  fann_set_learning_momentum(FFann,FLearningMomentum);
end;

procedure TMamFann.SetLearningRate(const Value: Single);
begin
  FLearningRate:=Value;
  fann_set_learning_rate(FFann,FLearningRate);
end;

procedure TMamFann.SetTrainingAlgorithm(const Value: TFann_train_enum);
begin
  FTrainingAlgorithm:=Value;
  fann_set_training_algorithm(
    FFann,Cardinal(ord(FTrainingAlgorithm)) );
end;

procedure TMamFann.Train(Inputs,Outputs: array of TFann_type);
begin
  fann_train(FFann, @Inputs[0], @Outputs[0]);
end;

procedure TMamFann.TrainOnFile(FileName: String; MaxEpochs: Cardinal;
  DesiredError: Single);
begin
  fann_train_on_file(
    FFann,PAnsiChar(AnsiString(FileName)),MaxEpochs,1000,DesiredError);
end;

end.

一式ダウンロード

fann_mam.zip(118KB)