merged

dlib/image_transforms/spatial_filtering.h

@@ -1233,15 +1233,25 @@ namespace dlib
             << "\n\t is_same_object(in_img,out_img): " << is_same_object(in_img,out_img) 
         );
 
+        if (sigma < 18)
+        {
             typedef typename pixel_traits<typename image_traits<out_image_type>::pixel_type>::basic_pixel_type type;
             typedef typename promote<type>::type ptype;
-
             const matrix<ptype,0,1>& filt = create_gaussian_filter<ptype>(sigma, max_size);
-
             ptype scale = sum(filt);
             scale = scale*scale;
-
             return spatially_filter_image_separable(in_img, out_img, filt, filt, scale);
+        }
+        else
+        {
+            // For large sigma we need to use a type with a lot of precision to avoid
+            // numerical problems.  So we use double here.
+            typedef double ptype;
+            const matrix<ptype,0,1>& filt = create_gaussian_filter<ptype>(sigma, max_size);
+            ptype scale = sum(filt);
+            scale = scale*scale;
+            return spatially_filter_image_separable(in_img, out_img, filt, filt, scale);
+        }
 
     }
 

examples/dnn_mnist_advanced_ex.cpp

@@ -56,7 +56,9 @@ template <typename SUBNET> using ares_down = base_ares<2,SUBNET>;
 // Now that we have these convenient aliases, we can define a residual network
 // without a lot of typing.  Note the use of a repeat layer.  This special layer
 // type allows us to type repeat<9,res<SUBNET>> instead of
-// res<res<res<res<res<res<res<res<res<SUBNET>>>>>>>>>.
+// res<res<res<res<res<res<res<res<res<SUBNET>>>>>>>>>.  It will also prevent
+// the compiler from complaining about super deep template nesting when creating
+// large networks.
 const unsigned long number_of_classes = 10;
 using net_type = loss_multiclass_log<fc<number_of_classes,
                             avg_pool<11,11,11,11,