cross_validate_graph_labeling_trainer() that can incorporate per node
loss values.

problems.

interface.

"file to include" not a link anymore since you probably shouldn't
be clicking on it when consulting the documentation.

times faster.  The behavior should be exactly the same.  The change is basically just
a switch to using the streambuf directly.

integer instead of the usual dlib variable width encoding.

as well as the normal version.

Google protocol buffer objects.

letting them fly out and terminate the application.

Now it can be anything.

for the factor_value_disagreement() values since the code actually supports this.

to make it robust to very large images.

But nevertheless, it's fixed.