The structural SVM optimizers had checks that the risk never became negative.

This only happens if the user implements the separation oracle wrong.  In this
case I had the code setup to force the risk to zero but this just ends up
obscuring the fact that the user's separation oracle is incorrect.  So I'm
removing this check.  This should make debugging separation oracles a little
easier.

dlib/svm/structural_svm_distributed.h

@@ -568,8 +568,7 @@ namespace dlib
 
                 subgradient /= num;
                 total_loss /= num;
-                // Include a sanity check that the risk is always non-negative.
-                risk = std::max<scalar_type>(total_loss + dot(subgradient,w), 0);
+                risk = total_loss + dot(subgradient,w);
             }
 
             std::vector<std::pair<std::string,unsigned short> > nodes;

dlib/svm/structural_svm_problem.h

@@ -348,8 +348,7 @@ namespace dlib
 
             subgradient /= num;
             total_loss /= num;
-            // Include a sanity check that the risk is always non-negative.
-            risk = std::max<scalar_type>(total_loss + dot(subgradient,w), 0);
+            risk = total_loss + dot(subgradient,w);
         }
 
         virtual void call_separation_oracle_on_all_samples (