Taiyo Mineo
Simple Recurrent Neural Network Predictor
Diff: SRNN.cpp
- Revision:
- 4:9d94330f380a
- Parent:
- 2:d623e7ef4dca
- Child:
- 5:026d42b4455f
--- a/SRNN.cpp Mon Feb 16 07:53:23 2015 +0000
+++ b/SRNN.cpp Wed Feb 18 15:01:17 2015 +0000
@@ -51,9 +51,12 @@
SRNN::~SRNN(void)
{
- delete [] sample; delete [] sample_maxmin;
- delete [] predict_signal; delete [] Win_mid;
- delete [] Wmid_out; delete [] expand_in_signal;
+ delete [] sample;
+ delete [] sample_maxmin;
+ delete [] predict_signal;
+ delete [] Win_mid;
+ delete [] Wmid_out;
+ delete [] expand_in_signal;
delete [] expand_mid_signal;
}
@@ -70,7 +73,7 @@
void SRNN::predict(float* input)
{
float *norm_input = new float[this->dim_signal];
-
+
// output signal
float* out_signal = new float[dim_signal];
// value of network in input->hidden layer
@@ -116,12 +119,14 @@
for (int n=0; n < dim_signal; n++) {
predict_signal[i_predict * dim_signal + n] = expand_signal(out_signal[n],sample_maxmin[n * 2],sample_maxmin[n * 2 + 1]);
}
-
+
}
-
+
// 領域解放
- delete [] norm_input; delete [] out_signal;
- delete [] in_mid_net; delete [] mid_out_net;
+ delete [] norm_input;
+ delete [] out_signal;
+ delete [] in_mid_net;
+ delete [] mid_out_net;
}
@@ -141,11 +146,24 @@
// 係数行列の更新量
float* dWin_mid = new float[row_in_mid * col_in_mid];
float* dWmid_out = new float[row_mid_out * col_mid_out];
+
// 前回の更新量:慣性項に用いる.
float* prevdWin_mid = new float[row_in_mid * col_in_mid];
float* prevdWmid_out = new float[row_mid_out * col_mid_out];
+
float* norm_sample = new float[len_seqence * dim_signal]; // 正規化したサンプル信号; 実際の学習は正規化した信号を用います.
+ // 出力層の信号
+ float* out_signal = new float[dim_signal];
+
+ // 入力層->中間層の信号和
+ float* in_mid_net = new float[num_mid_neuron];
+ // 中間層->出力層の信号和.
+ float* mid_out_net = new float[dim_signal];
+
+ // 誤差信号
+ float* sigma = new float[dim_signal];
+
// 係数行列の初期化
for (int i=0; i < row_in_mid; i++)
for (int j=0; j < col_in_mid; j++)
@@ -156,28 +174,17 @@
MATRIX_AT(Wmid_out,col_mid_out,i,j) = uniform_rand(width_initW);
// 信号の正規化:経験上,非常に大切な処理
- for (int seq=0; seq < len_seqence; seq++) {
- for (int n=0; n < dim_signal; n++) {
- MATRIX_AT(norm_sample,dim_signal,seq,n) =
- normalize_signal(MATRIX_AT(this->sample,dim_signal,seq,n),
- MATRIX_AT(this->sample_maxmin,2,n,0),
- MATRIX_AT(this->sample_maxmin,2,n,1));
- // printf("%f ", MATRIX_AT(norm_sample,dim_signal,seq,n));
+ for (int i_seq=0; i_seq < len_seqence; i_seq++) {
+ for (int dim_n=0; dim_n < dim_signal; dim_n++) {
+ MATRIX_AT(norm_sample,dim_signal,i_seq,dim_n) =
+ normalize_signal(MATRIX_AT(this->sample,dim_signal,i_seq,dim_n),
+ MATRIX_AT(this->sample_maxmin,2,dim_n,0),
+ MATRIX_AT(this->sample_maxmin,2,dim_n,1));
+ // printf("%f ", MATRIX_AT(norm_sample,dim_signal,i_seq,dim_n));
}
- // printf("\r\n");
+ //printf("\r\n");
}
- // 出力層の信号
- float* out_signal = new float[dim_signal];
-
- // 入力層->中間層の信号和
- float* in_mid_net = new float[num_mid_neuron];
- // 中間層->出力層の信号和.
- float* mid_out_net = new float[dim_signal];
-
- // 誤差信号
- float* sigma = new float[dim_signal];
-
// 前回の二乗誤差値:収束判定に用いる.
float prevError;
squareError = FLT_MAX;
@@ -221,42 +228,49 @@
} else {
// コンテキスト層 = 前回のコンテキスト層の出力
// 前回の中間層信号との線形和をシグモイド関数に通す.
- for (int d = 0; d < num_mid_neuron; d++) {
- expand_in_signal[dim_signal + d] = sigmoid_func(alpha_context * expand_in_signal[dim_signal + d] + expand_mid_signal[d]);
+ for (int d_in = 0; d_in < num_mid_neuron; d_in++) {
+ expand_in_signal[dim_signal + d_in] = sigmoid_func(alpha_context * expand_in_signal[dim_signal + d_in] + expand_mid_signal[d_in]);
}
}
+
+ // printf("%d matrix calc start. \r\n", iteration);
+
// バイアス項は常に1に固定.
expand_in_signal[dim_signal + num_mid_neuron] = 1;
-
+ // printf(" in bias OK \r\n");
// 入力->中間層の出力信号和計算
multiply_mat_vec(Win_mid,
expand_in_signal,
in_mid_net,
num_mid_neuron,
dim_signal + num_mid_neuron + 1);
+ // printf(" in->mid OK \r\n");
// 中間層の出力信号計算
sigmoid_vec(in_mid_net,
expand_mid_signal,
num_mid_neuron);
+ // printf(" mid sigmoid OK \r\n");
expand_mid_signal[num_mid_neuron] = 1;
+ // printf(" mid bias OK \r\n");
// 中間->出力層の出力信号和計算
multiply_mat_vec(Wmid_out,
expand_mid_signal,
mid_out_net,
dim_signal,
num_mid_neuron + 1);
+ // printf(" mid->out OK \r\n");
// 出力層の出力信号計算
sigmoid_vec(mid_out_net,
out_signal,
dim_signal);
-
+ // printf(" out sigmoid OK \r\n");
- for (int i = 0; i < dim_signal; i++) {
- predict_signal[i] = expand_signal(out_signal[i],
- MATRIX_AT(sample_maxmin,2,i,0),
- MATRIX_AT(sample_maxmin,2,i,1));
+ for (int i_dim = 0; i_dim < dim_signal; i_dim++) {
+ predict_signal[i_dim] = expand_signal(out_signal[i_dim],
+ MATRIX_AT(sample_maxmin,2,i_dim,0),
+ MATRIX_AT(sample_maxmin,2,i_dim,1));
}
- // printf("predict : %f %f %f \r\n", predict_signal[0], predict_signal[1], predict_signal[2]);
+ // printf("%d predict : %f %f %f \r\n", iteration, predict_signal[0], predict_signal[1], predict_signal[2]);
// print_mat(Wmid_out, row_mid_out, col_mid_out);
@@ -281,7 +295,7 @@
predict_signal[i] = expand_signal(out_signal[i],
MATRIX_AT(sample_maxmin,2,i,0),
MATRIX_AT(sample_maxmin,2,i,1));
- //printf("%f ", predict_signal[i]);
+ // printf("%f ", predict_signal[i]);
}
break;
}
@@ -311,7 +325,7 @@
}
// 中間->入力層の係数の変更量計算
- register float sum_sigma;
+ float sum_sigma;
for (int i = 0; i < num_mid_neuron; i++) {
// 誤差信号を逆向きに伝播させる.
sum_sigma = 0;
@@ -355,10 +369,15 @@
}
- delete [] dWin_mid; delete [] dWmid_out;
- delete [] prevdWin_mid; delete [] prevdWmid_out;
- delete [] norm_sample; delete [] out_signal;
- delete [] in_mid_net; delete [] mid_out_net;
+ delete [] dWin_mid;
+ delete [] dWmid_out;
+ delete [] prevdWin_mid;
+ delete [] prevdWmid_out;
+ delete [] norm_sample;
+ delete [] out_signal;
+ delete [] in_mid_net;
+ delete [] mid_out_net;
+ delete [] sigma;
return squareError;
}