Made the batch normalization code assign the parameter gradient outputs instead

of add to them so that it's consistent with how the layer interface expects
this to be done.

dlib/dnn/cpu_dlib.cpp

@@ -224,6 +224,9 @@ namespace dlib
             DLIB_CASSERT(num == beta_grad.size(),"");
             DLIB_CASSERT(have_same_dimensions(gradient_input, src),"");
             DLIB_CASSERT(have_same_dimensions(gradient_input, src_grad),"");
+
+            beta_grad = 0;
+            gamma_grad = 0;
             auto p_grad = gradient_input.host();
             auto p_src = src.host();
             const auto p_gamma = gamma.host();   

dlib/dnn/cuda_dlib.cu

@@ -268,6 +268,8 @@ namespace dlib
             {
                 dvars[i] = 0;
                 dmeans[i] = 0;
+                gamma_grad[i] = 0;
+                beta_grad[i] = 0;
 
                 for (long n = 0; n < num_samples; ++n)
                 {
@@ -646,6 +648,8 @@ namespace dlib
             dmeans.copy_size(means);
             dvars = 0;
             dmeans = 0;
+            gamma_grad = 0;
+            beta_grad = 0;
 
             _cuda_batch_normalize_conv_gradient1<<<512,512>>>(
                 gradient_input.device(),

dlib/dnn/tensor_tools.h

@@ -245,8 +245,8 @@ namespace dlib { namespace tt
                 - Let f(src,gamma,beta) == dot(gradient_input, dest output of
                   batch_normalize(dest,means,invstds,src,gamma,beta))
                 - Adds the gradient of f() with respect to src to #src_grad.
-                - Adds the gradient of f() with respect to gamma to #gamma_grad.
-                - Adds the gradient of f() with respect to beta to #beta_grad.
+                - Assigns the gradient of f() with respect to gamma to #gamma_grad.
+                - Assigns the gradient of f() with respect to beta to #beta_grad.
         !*/
     private:
 #ifdef DLIB_USE_CUDA
@@ -312,8 +312,8 @@ namespace dlib { namespace tt
                 - Let f(src,gamma,beta) == dot(gradient_input, dest output of
                   batch_normalize_conv(dest,means,invstds,src,gamma,beta))
                 - Adds the gradient of f() with respect to src to #src_grad.
-                - Adds the gradient of f() with respect to gamma to #gamma_grad.
-                - Adds the gradient of f() with respect to beta to #beta_grad.
+                - Assigns the gradient of f() with respect to gamma to #gamma_grad.
+                - Assigns the gradient of f() with respect to beta to #beta_grad.
         !*/
     private:
 #ifdef DLIB_USE_CUDA