merged

dlib/dnn/layers.h

@@ -80,7 +80,7 @@ namespace dlib
             // allocate params for the filters and also for the filter bias values.
             params.set_size(num_inputs*_num_filters + _num_filters);
 
-            dlib::rand rnd("con_"+cast_to_string(num_outputs+num_inputs));
+            dlib::rand rnd(std::rand());
             randomize_parameters(params, num_inputs+num_outputs, rnd);
 
             filters = alias_tensor(_num_filters, sub.get_output().k(), _nr, _nc);
@@ -663,7 +663,7 @@ namespace dlib
             else
                 params.set_size(num_inputs, num_outputs);
 
-            dlib::rand rnd("fc_"+cast_to_string(num_outputs));
+            dlib::rand rnd(std::rand());
             randomize_parameters(params, num_inputs+num_outputs, rnd);
 
             weights = alias_tensor(num_inputs, num_outputs);

dlib/dnn/trainer.h

@@ -513,8 +513,10 @@ namespace dlib
                 for (size_t i = 0; i < devices.size(); ++i)
                     losses[i] = std::async(std::launch::async,[&,i](){ return compute_parameter_gradients(i, next_job, pick_which_run_update); });
                 // aggregate loss values from all the network computations.
+                double theloss = 0;
                 for (auto&& loss : losses)
-                    record_loss(loss.get());
+                    theloss += loss.get();
+                record_loss(theloss/losses.size());
 
                 // Now, if there is more than one active device we need to synchronize the
                 // gradient updates between devices.  So we do that now.