Taiyo Mineo
Simple Recurrent Neural Network Predictor
Diff: SRNN.cpp
- Revision:
- 5:026d42b4455f
- Parent:
- 4:9d94330f380a
- Child:
- 6:e97ccc643bf1
--- a/SRNN.cpp Wed Feb 18 15:01:17 2015 +0000
+++ b/SRNN.cpp Thu Feb 19 08:20:23 2015 +0000
@@ -18,6 +18,7 @@
// sample/sample_maxmin allocation
this->sample = new float[len_seqence * dim_signal];
+ this->norm_sample = new float[len_seqence * dim_signal];
this->sample_maxmin = new float[dim_signal * 2];
memcpy(this->sample, input_sample, sizeof(float) * len_seqence * dim_signal);
@@ -34,6 +35,30 @@
expand_in_signal = new float[dim_signal + num_mid_neuron + 1];
expand_mid_signal = new float[num_mid_neuron + 1];
+ // For BUG FIX...
+ // 係数行列のサイズ
+ int row_in_mid = num_mid_neuron;
+ int col_in_mid = dim_signal + num_mid_neuron + 1;
+ int row_mid_out = dim_signal;
+ int col_mid_out = num_mid_neuron + 1;
+
+ // 行列のアロケート
+ // 係数行列の更新量
+ this->dWin_mid = new float[row_in_mid * col_in_mid];
+ this->dWmid_out = new float[row_mid_out * col_mid_out];
+
+ // 前回の更新量:慣性項に用いる.
+ this->prevdWin_mid = new float[row_in_mid * col_in_mid];
+ this->prevdWmid_out = new float[row_mid_out * col_mid_out];
+ // 出力層の信号
+ this->out_signal = new float[dim_signal];
+ // 入力層->中間層の信号和
+ this->in_mid_net = new float[num_mid_neuron];
+ // 中間層->出力層の信号和.
+ this->mid_out_net = new float[dim_signal];
+ // 誤差信号
+ this->sigma = new float[dim_signal];
+
// Parameter settings (Tuning by taiyo)
this->squareError = FLT_MAX; // (large value)
this->maxIteration = 5000;
@@ -52,6 +77,7 @@
SRNN::~SRNN(void)
{
delete [] sample;
+ delete [] norm_sample;
delete [] sample_maxmin;
delete [] predict_signal;
delete [] Win_mid;
@@ -72,8 +98,8 @@
/* Predict : predicting next sequence of input */
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
@@ -142,6 +168,7 @@
int row_mid_out = dim_signal;
int col_mid_out = num_mid_neuron + 1;
+ /*
// 行列のアロケート
// 係数行列の更新量
float* dWin_mid = new float[row_in_mid * col_in_mid];
@@ -151,7 +178,7 @@
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.
+ // 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];
@@ -163,6 +190,7 @@
// 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++)
@@ -173,16 +201,17 @@
for (int j=0; j < col_mid_out; j++)
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
+ float buf_float;
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));
+ 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("\r\n");
+ // printf("%f %f %f \r\n", MATRIX_AT(norm_sample,dim_signal,i_seq,0), MATRIX_AT(norm_sample,dim_signal,i_seq,1),MATRIX_AT(norm_sample,dim_signal,i_seq,2));
}
// 前回の二乗誤差値:収束判定に用いる.
@@ -191,6 +220,8 @@
/* 学習ループ */
while (1) {
+ // printf("dWin_out : %p, dWmid_out : %p, prevdWin_mid : %p, prevdWmid_out : %p, out_signal : %p, in_mid_net : %p, mid_out_net : %p, sigma : %p \r\n", dWin_mid, dWmid_out,prevdWin_mid, prevdWmid_out, out_signal, in_mid_net, mid_out_net, sigma);
+
// 終了条件を満たすか確認
if (!end_flag) {
end_flag = !(iteration < maxIteration
@@ -217,10 +248,21 @@
memset(prevdWmid_out, float(0), sizeof(float) * row_mid_out * col_mid_out);
}
+ memset(dWin_mid, float(0), sizeof(float) * row_in_mid * col_in_mid);
+ memset(dWmid_out, float(0), sizeof(float) * row_mid_out * col_mid_out);
+
/* 学習ステップその1:ニューラルネットの出力信号を求める */
// 入力値を取得
- memcpy(expand_in_signal, &(norm_sample[seq * dim_signal]), sizeof(float) * dim_signal);
+ memcpy(expand_in_signal, &(norm_sample[(seq * dim_signal)]), sizeof(float) * dim_signal);
+
+ /*
+ if ( iteration >= 0 ) {
+ printf("%d first 3 norm_sample : %f(%p) %f(%p) %f(%p) \r\n", iteration, norm_sample[seq * dim_signal], &(norm_sample[seq * dim_signal]), norm_sample[seq * dim_signal + 1], &(norm_sample[seq * dim_signal + 1]), norm_sample[seq * dim_signal + 2], &(norm_sample[seq * dim_signal + 2]));
+ printf("%d first 3 expand_in_signal : %f(%p) %f(%p) %f(%p) \r\n", iteration, expand_in_signal[0], &(expand_in_signal[0]), expand_in_signal[1], &(expand_in_signal[1]), expand_in_signal[2], &(expand_in_signal[2]));
+ }
+ */
+
// SRNN特有:入力層に中間層のコピーが追加され,中間層に入力される.
if (iteration == 0) {
// 初回は0埋めする
@@ -270,10 +312,11 @@
MATRIX_AT(sample_maxmin,2,i_dim,0),
MATRIX_AT(sample_maxmin,2,i_dim,1));
}
+
+ // printf("%d sample : %f %f %f \r\n", seq, sample[dim_signal * seq], sample[dim_signal * seq + 1], sample[dim_signal * seq + 2]);
+ // printf("%d output : %f(%p) %f(%p) %f(%p) \r\n", iteration, out_signal[0], &(out_signal[0]), out_signal[1], &(out_signal[1]), out_signal[2], &(out_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);
-
// この時点での二乗誤差計算
squareError = 0;
// 次の系列との誤差を見ている!! ここが注目ポイント
@@ -286,6 +329,11 @@
}
}
squareError /= dim_signal;
+ // printf("%f \r\n", squareError);
+ if ( isnan(squareError) || isinf(squareError) ) {
+ fprintf( stderr, "SRNN LEARNING ERROR!! Learning failed (detected %f) \r\n", squareError);
+ exit(1);
+ }
/* 学習の終了 */
// 終了フラグが立ち,かつ系列の最後に達していたら学習終了
@@ -305,17 +353,24 @@
end_flag = 1;
}
+ /*
+ if ( iteration > 40 ) {
+ printf("%d sample : %f %f %f \r\n", seq, sample[dim_signal * seq], sample[dim_signal * seq + 1], sample[dim_signal * seq + 2]);
+ printf("%d norm_sample : %f %f %f \r\n", iteration, MATRIX_AT(norm_sample,dim_signal,seq,0), MATRIX_AT(norm_sample,dim_signal,seq,1), MATRIX_AT(norm_sample,dim_signal,seq,2));
+ }
+ */
+
/* 学習ステップその2:逆誤差伝搬 */
// 誤差信号の計算
- for (int n = 0; n < dim_signal; n++) {
+ for (int n_dim = 0; n_dim < dim_signal; n_dim++) {
if (seq < len_seqence - 1) {
- sigma[n] = (out_signal[n] - MATRIX_AT(norm_sample,dim_signal,seq+1,n)) * out_signal[n] * (1 - out_signal[n]);
+ sigma[n_dim] = (out_signal[n_dim] - MATRIX_AT(norm_sample,dim_signal,(seq+1),n_dim)) * out_signal[n_dim] * (1 - out_signal[n_dim]);
} else {
/* 末尾と先頭の誤差を取る (大抵,大きくなる) */
- sigma[n] = (out_signal[n] - MATRIX_AT(norm_sample, dim_signal,0,n)) * out_signal[n] * (1 - out_signal[n]);
+ sigma[n_dim] = (out_signal[n_dim] - MATRIX_AT(norm_sample,dim_signal,0,n_dim)) * out_signal[n_dim] * (1 - out_signal[n_dim]);
}
}
- // printf("Sigma : %f %f %f \r\n", sigma[0], sigma[1], sigma[2]);
+ // printf("%d Sigma : %f %f %f \r\n", iteration, sigma[0], sigma[1], sigma[2]);
// 出力->中間層の係数の変更量計算
for (int n = 0; n < dim_signal; n++) {
@@ -369,15 +424,15 @@
}
- 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 [] sigma;
+ // 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;
}
@@ -385,5 +440,5 @@
// サンプルの（リ）セット
void SRNN::set_sample(float* sample_data)
{
- memcpy(sample, sample_data, sizeof(float) * len_seqence * dim_signal);
+ memcpy(this->sample, sample_data, sizeof(float) * len_seqence * dim_signal);
}