第一次成功整合了卷积核的bias，达到89750新高：# train_num 3000， test_num 400， epoch 300，…

… lr 0.03-100-0.003， filter_bias G(0, 0.2)， minus_maxi True， filter_bias_update True

include/model.h

@@ -217,7 +217,7 @@ _type convMatrix(CVLayer * input, CVLayer * conv, int k, int x, int y, int index
  */
 void UpdateFilter(CVLayer * filter[], CVLayer * input, CVLayer * conv, double alpha) {
     for (int index = 0; index < filter_num; ++ index) {
-//        filter[index]->bias -= alpha * sumMatrix(conv, index);
+        filter[index]->bias -= alpha * sumMatrix(conv, index);
         for (int k = 0; k < filter[index]->H; ++ k)
             for (int i = 0; i < filter[index]->L; ++ i)
                 for (int j = 0; j < filter[index]->W; ++ j)
@@ -232,11 +232,7 @@ void UpdateFilter(CVLayer * filter[], CVLayer * input, CVLayer * conv, double al
  * @param alpha 学习率参数
  * @param label Ground Truth标签
  */
-void backPropagation(Network * CNN, int step, Alpha * alpha, int label) {
-    for (int i = 0; i < class_num; ++ i) {
-        CNN->fc_output->deltas[i] = CNN->fc_output->values[i];
-        if (i == label) CNN->fc_output->deltas[i] -= 1.0;
-    }
+void backPropagation(Network * CNN, int step, Alpha * alpha) {
     for (int i = 0; i < CNN->fc_input->L; ++ i) {
         CNN->fc_input->deltas[i] = 0;
         for (int j = 0; j < class_num; ++ j) {

include/network.h

@@ -129,7 +129,7 @@ CVLayer * Convol2D_(int h, int l, int w) {
 
 CVLayer * Filter2D_(int h, int l, int w) {
     CVLayer * layer = (CVLayer *)malloc(sizeof(CVLayer));
-    layer->destroy = killCVLayer; layer->bias = GaussRand(2.75, 0.2);
+    layer->destroy = killCVLayer; layer->bias = GaussRand(0.1, 0.2);
     layer->L = l; layer->W = w; layer->H = h; layer->deltas = NULL;
     layer->values = (_type ***)malloc(h * sizeof(_type **));
     for (int k = 0; k < h; ++ k) {

src/main.c

@@ -11,9 +11,10 @@
     return 0; \
 }while(0)
 
-int train_num = 3000; // 训练样本数
-int test_num  = 400; // 测试样本数
+int train_num = 6000; // 训练样本数
+int test_num  = 600; // 测试样本数
 int epoch_num = 300; // 训练轮数
+int batch_size = 1;
 
 Vector1D labels_train;
 Vector2D images_train;
@@ -45,10 +46,20 @@ void train(Network * CNN, Alpha * alpha, const char * fileName) {
     printf("Begin Training ...\n");
     for (int step = 0; step < epoch_num; ++ step) {
         _type err = 0;
-        for (int i = 0; i < train_num; i ++) {
-            forePropagation(CNN, i, images_train);
-            err -= log(CNN->fc_output->values[(int)labels_train.data[i]]);
-            backPropagation(CNN, step, alpha, (int)labels_train.data[i]);
+        for (int i = 0; i < train_num / batch_size; i ++) {
+            for (int k = 0; k < class_num; ++ k)
+                CNN->fc_output->deltas[k] = 0;
+            for (int j = 0; j < batch_size; ++ j) {
+                forePropagation(CNN, i * batch_size + j, images_train);
+                int label = (int)labels_train.data[i * batch_size + j];
+                for (int k = 0; k < class_num; ++ k)
+                    CNN->fc_output->deltas[k] += CNN->fc_output->values[k];
+                CNN->fc_output->deltas[label] -= 1.0;
+                err -= log(CNN->fc_output->values[label]);
+            }
+            for (int k = 0; k < class_num; ++ k)
+                CNN->fc_output->deltas[k] /= batch_size;
+            backPropagation(CNN, step, alpha);
         }
         new_score = test(CNN);
         printf("step: %3d loss: %.5f prec: %.5f\n",
@@ -102,7 +113,7 @@ int main(int argc, char * argv[]) {
             images_test.data[i][j] /= 255.0;
 
     // 训练
-    train(CNN, alpha, "model.sav");
+    train(CNN, alpha, "../model.sav");
 
     // 释放内存
     delete_p(CNN);