dropout convolutions

Author: ab

Sources/convolutional_layer.hpp

@@ -43,11 +43,20 @@ class convolutional_layer : public layer {
         assert(in.size() == in_feature_maps_ * in_width_ * in_width_);
         assert(weights_.size() == in_feature_maps_*out_feature_maps_*filter_width_*filter_width_);
 
+        if (dropout_prob_ > 0.0)
+            sample_dropout();
+
         for (uint_t out_fm=0; out_fm<out_feature_maps_; out_fm++) {
 
             for (uint_t ox=0; ox<out_width_; ox++) {
                 for (uint_t oy=0; oy<out_width_; oy++) {
 
+                    /* dropout while training: [0,1), dropout while testing: 1 */
+                    if (dropout_prob_!= 0 && dropout_prob_!=1 && dropout_sample_[(out_fm*out_width_+ ox)*out_width_ + oy]<dropout_prob_) {
+                        output_[(out_fm*out_width_+ ox)*out_width_ + oy] = 0.0;
+                        continue;
+                    }
+
                     float_t sum = 0.0;
                     for (uint_t in_fm=0; in_fm<in_feature_maps_; in_fm++) {
 
@@ -61,6 +70,10 @@ class convolutional_layer : public layer {
                             }
                         }
                     }
+                    /* dropout while testing */
+                    if (dropout_prob_==1)
+                        sum = sum*0.5;
+
                     output_[(out_fm*out_width_+ ox)*out_width_ + oy] = A_.f(sum + bias_[out_fm]);
                 }
             }

Sources/fullyconnected_layer.hpp

@@ -47,7 +47,7 @@ class fullyconnected_layer : public output_layer {
 
             /* dropout while testing */
             if (dropout_prob_==1)
-                sum = sum*0.5;
+                sum = sum*0.1;
 
             output_[o] = ActFunc.f(sum + bias_[o]);
         }

Sources/main.cpp

@@ -155,9 +155,8 @@ cnn_training_test_mnist()
     vec_t soll {0,0,0,0,0,0,0,0,0,0};
     int last_label = 0;
     for(int s=1; s<=steps; s++) {
-        
-        
-        S4.set_dropout_prob(0.5);
+
+        S4.set_dropout_prob(0.9);
 #if TRAINING_MOMENTUM
         if (s%1000==0) {
             nn.set_learning_rate( s==0? 0.00085 : nn.learning_rate()*0.85);

Sources/subsampling_layer.hpp

@@ -29,8 +29,10 @@ class subsampling_layer : public layer {
     }
 
     void forward(const vec_t& in /*[in_feature_map * in_width_ * in_width_]*/) {
-        
         assert(in.size() == feature_maps_ * in_width_ * in_width_);
+        
+        if (dropout_prob_ > 0.0)
+            sample_dropout();
 
         for (uint_t fm=0; fm<feature_maps_; fm++) {