Old Change Logs

Minor changes to avoid compiler warnings from clang++

Minor changes to avoid compiler warnings in gcc

Added the histogram intersection kernel for sparse and dense vectors.

The last dlib release added a max_iterations parameter to the
svm_c_linear_trainer and svm_c_ekm_trainer objects.  However,
there was a bug which made them only ever do at most 16 iterations,
which is too few to solve many problems.  This has been fixed.

added more tests

Clarified spec

Added unit tests for the new sequence labeling stuff

Simplified the sequence_labeler interface a bit.

Made unit test a little more robust.

Minor change to avoid a compiler error in clang

Changed a little to work around a bug in visual studio 2005

changed to avoid a compiler error

Changed code a little to make the reject_labeling() routine optional.
Now if you don't define it you get a reasonable default behavior
rather than a compiler error.

Filled in the rest of the interface for the structural_sequence_labeling_trainer
and cleaned up a few other related things.

fixed grammar

Refactoring and spec improvement.  Still some work left to do though.

fleshed out the spec

minor cleanup

Refactored code a little and added more comments to spec.

Made this object inherit from noncopyable

setup serialization for the sequence_labeler.

Added a set of tools to allow a user to easily learn to do sequence
labeling using dlib's structural SVM implementation.

Added the is_sequence_labeling_problem() routine.

Added some missing validation of the user supplied number of folds
to the cross_validate_multiclass_trainer() routine.  Not it will
throw an exception if the number of folds is too big rather than
just producing a confusing result.

Clarified the spec regarding the type of sparse vector supported
by these two objects.

Clarified the sparse vector spec and formally introduced the concept
of unsorted sparse vectors.

Fixed a bug in find_max_factor_graph_viterbi() which occurred when a
zero order model had negative factor values.

Clarified spec

Changed find_max_factor_graph_viterbi() so you can use run-time
defined order and num_states parameters.

Fixed a bug in the statement::get_column_as_text() routine.  It didn't
work right if the column contained a NULL.

Minor change to avoid a warning in gcc

Removed unused variables.

Changed makefile slightly so it compiles in Ubuntu 11.10

Added the is_array2d template.

Fixed a bug in is_const_type.  It didn't work right for reference types.

Added unit tests and fixed a few issues.

Clarified spec

Improved spec

Added functions for computing the convolution and cross-correlation between
two matrices.

Added flip()

fixed typo in spec

Fixed invalid assert

Added the segment_image() routine.

Added randomly_color_image()

improved spec

Added an object to help iterate over the border elements of a rectangle.

Added an implementation of the union-find algorithm.

Added missing requirement to the equalize_histogram() and get_histogram() routines.

fixed typo

clarified spec

Optimized the trust region solver a little bit.

Changed LAPACK bindings so that really small matrices don't use LAPACK
for eigenvalue_decomposition or the triangular solver.

Changed chol() so it only calls LAPACK if the matrix is big enough
to benefit from LAPACK.

Added a constructor to array2d that lets you set its size.

clarified spec

Changed the behavior of gaussian_blur() to make it a little more user friendly.

Added a parameter to control the maximum number of iterations for the
two OCA based svm optimizers.

Fixed a bug pointed out by Justin Solomon which could cause the svr_trainer and
svm_c_trainer to produce incorrect results in certain unusual cases.  Also added
unit tests to make sure the bug stays fixed.

Fixed typo in spec

Added a cast to avoid compiler errors in visual studio.

Added missing assert and made minor refinements to spec.

Added some tests for the label_connected_blobs() function.

Added the label_connected_blobs() routine.

updated spec

Changed suppress_non_maximum_edges() to use the L2 norm instead of L1 norm
for measuring the strength of an edge since this produces a slightly better
result.

Removed an unneeded cast.

Added a convenience routine for blurring an image with a Gaussian filter.

Added is_float_type.

Fixed a bug in the matrix class.  Expressions of the form mat *= mat(0) would
evaluate incorrectly because the *= operator took the right hand side by reference
and thus experienced an aliasing problem.  The other op= operators had similar
problems and have also been fixed.

Increased usability of the dlib CMakeLists.txt file slightly.

setup cmake file to link against sqlite3 if it can find it.

Just removed the sqlite namespace and unindended everything.

Added missing asserts and clarified spec.

Added a C++ wrapper for SQLite.

Changed shared_ptr slightly to avoid warnings from gcc.

Changed find_max_factor_graph_nmplp() to use a simple hash table instead
of std::map.  This is significantly faster.  I also added some missing asserts
to validate that the map problems supplied by the user are valid.

Changed the suppress_non_maximum_edges() routine so that it is more robust
to a wider range of pixel types.

Another minor change to avoid a bug in visual studio 2005.

Changed code to avoid a bug in visual studio 2005

Previously, when using spatially_filter_image(), the intermediate filter computations
were carried out using variables of the same basic type as the input image's pixels.
I changed this to use the scalar type used by the actual filter given to this function
since this should result in fewer surprises (since the filter usually determines what
type is appropriate, not the input image).

Minor change to avoid a compiler warning from gcc.

Added some casts and a few other minor changes to get the code compiling
in gcc 4.1.1.

Fixed some corner cases in find_max_factor_graph_viterbi() and also
added unit tests.

Added missing requirement

Minor clarification.

Added the find_max_factor_graph_viterbi() routine for performing MAP
inference in chain-structured factor graphs.

Renamed a function so auto complete in Vim works more conveniently.

Changed code slightly to avoid a bug in visual studio 2005.

Added unit tests for the new object detection tools

Minor changes to avoid compiler warnings

Fixed grammar.

Fixed spelling error.

Just renamed get_scales() to get_hash_bin_sizes() to make things more clear.

Just renamed find_map_nmplp() to find_max_factor_graph_nmplp()

Minor change to avoid a linker error.

Changed scan_image() so that once the max number of detections is reached
it begins randomly subsampling possible detection locations rather than
just stopping the detection procedure.

clarified the spec

Added some comments.

Fleshed out the spec.

Fixed bugs in asserts.

Added the cross_validate_object_detection_trainer() and test_object_detection_function()
routines.

Improved the error messages generated by the structural_svm_object_detection_problem.

Increased the default max detections per template from 2000 to 10000 since
setting this value too low can sometimes result in the failure of the
structural_svm_object_detection_problem.

clarified spec

Added an initial version of the structural_object_detection_trainer.  This is
a tool for learning the parameters for an object like scan_image_pyramid.

Defaulted the overlap tester type to test_box_overlap.

added missing requirement.

Fixed the width_step() member functions.  They weren't defined in quite the
way they should have been.

Minor change to avoid a compiler warning.

Improved the error messages output when a set of truth rectangles can't be
represented by an image scanner.

Changed the test_regression_function() and cross_validate_regression_trainer()
routines so they return both the MSE and R-squared values rather than just the
MSE.

- Renamed spatially_filter_image() for the separable case to spatially_filter_image_separable()
 - Changed the behavior of spatially_filter_image().  Now it won't truncate signed pixel values
   to 0 if they go negative.

Changed the interface for the spatially_filter_image() routines to take the filter
as a matrix rather than C-array.  I also fixed a bug which showed up when using
non-square filters.  The bug would cause the edges of the output image to be incorrect.

Increased the max size matrix which will allocate its memory on the stack
from 128 bytes to 256 bytes.

changed to avoid a compiler warning

Added an overload of spatially_filter_image() that uses a separable filter.

Added functions to the interface that allow a user to control the relative
loss from a false alarm vs. a missed detection.

setup separate measures of loss for false alarms and missed truth boxes.

Added comments clarifying the thread safety requirements of the new
object detection tools.

Removed some testing/debugging code like internal asserts, cout
statements, and other cruft.

Added yet more missing asserts

Added a missing assert

Added a missing assert and fixed some typos.

Added missing assert

simplified this object a little bit.

Fixed a bug in the cpp_tokenizer that only shows up on newer versions of
gcc.  It wasn't tokenizing double quoted strings right.

filled out spec

improved the spec

Added some comments.  Still need to complete the spec.

filled in the spec

Added a functor for testing if two rectangles overlap.

Added the structural_svm_object_detection_problem object.

Added the object_detector object.

Added some functions for creating detection templates.

Added the scan_image_pyramid object.

Added the hashed_feature_image object.

Fixed typos in spec.

Added missing requirement.

Added some optimized serialization overloads for std::vectors of
characters to take advantage of the fact that you can just write
the contents of memory out in this case.

Fixed typo

Fixed grammar in comment.

Added a copy_configuration() routine to the hog_image.

Added a get_num_dimensions() method to the hog_image.

Changed the set_current_dir() implementation so that it compiles in visual studio
even when compiling with unicode strings enabled.

Added cast to avoid compiler warning in visual studio 2010

Added the scan_image() routine.  It is a tool for sliding a set of rectangles
over an image space and finding the locations where the sum of pixels in
the rectangles exceeds a threshold.

clarified spec

Added a missing requires clause

clarified the spec

updated the spec

Added some unit tests for the find_map_nmplp() routine.

clarified spec

Added the find_map_nmplp() function for performing approximate
MAP inference.

Fixed a bug in the png_loader.  If you loaded an image with an
alpha channel into something without an alpha channel there were
uninitialized values being alpha blended into the image.

Added some overloads of serialize/deserialize for array2d objects
which contain simple 8bit POD pixel types.  This makes the serialize
routines much faster in these cases.

- Formally defined how array2d objects must lay their components out in memory
 - Added width_step() to array2d to help define the memory layout.  Also added
   it to cv_image to keep the interfaces compatible.
 - Fixed a typo in the deserialize for array2d objects.  The template wasn't
   declared properly.

Changed the serialization routine for bgr_pixels to store the pixels
in BGR order rather than RGB.

Added a mem_manager_type typedef so the interface better matches array2d.

Fixing incorrect print statement.

Moved the docs for some functions to the top so that htmlify links to them right.

Changed the interface to the ridge regression trainer objects so that
they report the entire set of LOO prediction values rather than a
summary statistic like mean squared error.

Added the following functions for computing statistics on vectors:
mean_sign_agreement(), correlation(), covariance(), r_squared(),
and mean_squared_error()

Fixed typo in spec

Cleaned up the interface to draw_line()

Added missing inline to avoid compiler errors.

Added a function for computing the center of a rectangle.

Added a test that the pyramid down objects do the coordinate mapping correctly.

Fixed a minor bug in the list_box.  It's behavior was a little awkward
when using the multiple selection feature.

Changed the pyramid down objects so that they make better use of the
available image data.  Previously, more pixels on the right and bottom
of the image were cropped than was necessary.

Updated the block spatial filtering functions so they can operate on
rectangular areas instead of just square areas.

Fixed a bug pointed out by Martin Müllenhaupt which caused the windows socket
code to not compile when used with the mingw-cross-env project.

Made the behavior of the image_window a little more convenient.  Now it will
automatically size itself properly when a zoom is in effect and you place a new
image into it.

The image_display didn't display overlay rectangles quite right.  If you zoomed
in you could see that some of the pixels which are inside the rectangle were
outside the overlay.  Specifically, the right column and bottom row was outside
the overlay rectangle.  This has been fixed.  Now all pixels which are supposed
to be part of a rectangle are drawn as being inside the overlay rectangle.

Made the integral_image more general by making it templated on the
type of scalar used to store the sums.

Added a global operator+ for rectangle+point to make this kind of operation
compile in more contexts.

fixed minor typo in comment

Made hog_image clearable and serializable.

Made running_stats serializable.

Clarified spec.

added more unit tests for hog coordinate transforms

Removed unneeded code.

Added a more complete set of functions for converting between image space and the downsampled hog
grid.  So now you can convert from image to hog instead of just hog to image.  Also added some
related unit tests.

Added fill_rect() for images.

Clarified spec

Added some tests for the new versions of pyramid_down.

Updated image pyramid spec to include new pyramid functions.

Added a version of pyramid_down which scales at a ratio of 5:4

Added a 4:3 pyramid down tool.

Moved the /256 from the spatial filtering routine to it's proper place in the pyramid code.

Made code a little easier to read.

Adding initial version of a 3:2 pyramid

Added some functions for running a 3x3 separable filter on part of an image.

Fixed grammar in comment.

Added a file_exists() function.

Fixed spelling error in comment.

Changed code around so that the save_png.h header file doesn't need to #include png.h.
I did this because png.h will cause the compile to fail if the user #includes setjmp.h
along with png.h.

Added an event handler for when the user selects an overlay rectangle.

Added the ability for the user to delete overlay rectangles from the image_display.
Also added interface functions to programmatically observe what is happening to the
overlays.

Fixed a bug in the svm_struct test.  The test would always fail if run
more than once.

Simplified the CMakeLists.txt file for the unit tests.  Also made it
use Release mode by default.

Minor change to suppress a compiler warning.

removed tabs

Added support for user drawn rectangle overlays and selectable overlays to the
image_display widget.

Added the distance_to_rect_edge() routine.

Made a minor change which apparently avoids a bug in gcc 5.4.2 on Ubuntu 11.04.

Simplified the cmake files by putting the stuff to make Release
the default build into a single file.

Fixed a bug in the image_display widget.  If you switched it between
images of a different size while any kind of zoom was in effect
it could cause a segmentation fault.

Made the global mutex construction for the get_current_dir() and set_current_dir()
routines more robust.

Fixed the code so it compiles on windows.

Added a set_current_dir() function.

Made the horizontal scroll speed for the list_box more reasonable.

Added match_endings to the dir_nav utils.

Added the check_sub_option() method to the command line parser check
object.

Minor change to avoid a compile time error in visual studio 2008.

I improved the way calls to the separation oracle get dispatched to the
thread_pool.  Previously, each separation oracle call was dispatched
to a thread individually.  This is inefficient when there are a whole
lot of samples (and thus separation oracle calls which need to be made).
So now entire batches of separation oracle calls are dispatched to
each thread.  This minimizes the thread switching and synchronization
overhead.

Added extra #include for image_io.h so that future linker errors from
this component will be caught by the unit tests.

Added a bunch of overloads for add_task_by_value()

Removed spurious statement.

Moved the serialize and deserialize function definitions outside the body of the
type_safe_union class to avoid a compiler bug in an older version of gcc.

Fixed a bug which caused multiply defined symbol errors during linking
if the PNG saving routine was #included.

Clarified spec and docs

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.

Added some tests for the new add_task_by_value() methods of the thread_pool.

added a test

Added an add_task_by_value() function to the thread_pool.

Clarified spec again

clarified docs

Switched some member_function_pointers from checking mode to non-checking mode.

Fixed comment

Removed cruft

Made the conversion to and from HSI a little more robust.

Removed time based seed to make test more repeatable.

Changed code slightly to avoid compiler error in visual studio 2005

Added another test

updated spec

Made find_clusters_using_kmeans() work with dlib::vector objects.

Added a kmeans test

Changed code so that an error message is always printed if we run out of memory while
cross_validate_trainer_threaded() is executing.

Increased the hysteresis_threshold() recursion depth a bit.

Made threshold_image() handle signed pixels properly.

Changed hysteresis_threshold() so it properly handles signed pixel types.

Added comments explaining the conversion that might happen when images
are saved to disk by the various save_*() functions.

Fixed a bug in assign_pixel_intensity() that happened when
the target pixel was an RGB pixel with an alpha channel.

Added more tests

Made get_pixel_intensity() more efficient.

Minor change to avoid a compiler error in visual studio 2010

Switched the default hash function for std::strings (when used with hash tables)
to dlib::hash().

minor cleanup

Added another test

Improved hashing tests and clarified spec.

Clarified spec

Added comments regarding endianness and hashing.

Added a hash() for matrices.

Fixed a linker error that happens on gcc.

Added a unit test for the new hashing functions.

Added Austin Appleby's excellent MurmurHash3 hashing code and setup some
additional convenience functions.

Changed comment so htmlify will put an anchor on it.

Changed code to use the new DLIB_ASSERT_HAS_STANDARD_LAYOUT to check
if a type can be swapped rather than the explicit union trick.

Moved DLIB_ASSERT_HAS_STANDARD_LAYOUT into assert.h.

Added a macro to assert on a type being a POD or a standard layout type.
This way I can change just this one macro to use the better C++0x
method of doing this check once C++0x becomes available.

Added some unit tests for the save_png() and load_png() functions.

Switched all the uses of array2d to not use the array2d<type>::kernel_1a
syntax but to use the array2d<type> syntax.

Changed the array2d object so you don't have to say array2d<type>::kernel_1a
anymore to declare it.  Now you just say array2d<type>.

Removed debugging cout statement.

Fixed incorrect exception being thrown.

Fixed compile time test

Added the load_image() function which just looks at a file's extension and
then calls the appropriate image loading function.

Added a bunch of overloads for all the various save_*() and load_*() image
I/O options dlib supports.  Also added ones which can work directly with
a string file name.

Fixed code so it throws the right type of exception.

Added a function for saving to the PNG image format.

updated png_loader so it can load 16bit data.

Made the png_loader::get_image() function a little faster.

Updated png_loaded so it can handle paletted images.

setup proper error handling for libpng in the png_loader.  Now if the PNG
file is corrupted in the middle it won't just print a message and abort
the program.

Added missing requires clause to hysteresis_threshold()

fixed typo in spec

Fixed a spelling error.

Fixed spelling error in a comment.

Added a few more tests

removed tabs

Added a matrix_type typedef to structural_svm_problem_threaded
so it matches the interface of the structural_svm_problem.

Minor change to avoid compiler warning in gcc.

Changed the type_safe_union a little bit to avoid the C4503 warnings
you sometimes get from visual studio about
"decorated name length exceeded, name was truncated".

What I did was introduce a new type called _void as the deafult
template parameter for the type_safe_union rather than use the T1
type as the default for the 19 other parameters.  Since _void is
always a short word it helps keep the warning from showing up.

Removed random seed to make the test more repeatable.

Changed code to avoid compiler error in visual studio.

Added a bunch of unit tests for the various forms of structured svm.

Clarified spec

Worked on the spec and cleaned up the code a little.

clarified spec

Added initial version of a distributed structural svm solver.

Added some requirements and asserts.

Added a function to determine how many threads a structural_svm_problem_threaded
object is going to use.

Simplified code a little bit.

Made caching object a little more robust.

Added some more comments.

Cleaned up the code a little by pulling the caching logic out into its
own class.

Minor change to avoid a compiler warning.

Fixed a bug in the DLIB_TEST_MSG macro for the unit testing suite.  This
bug was introduced a few days ago.

Clarified recent change to the spec

Made the bridge code a little more robust.

Made test more robust

Added a clear() to the bridge and improved some comments and tests.

Added tests for the bridge.

Polished the bridge object a bit.

Yet more code cleanup

minor cleanup

Added initial version of the bridge object.  It is a tool for bridging
pipe objects from one computer to another.

Updated the unit test application so that it prints the number of
individual testing statements which were executed.

Upgraded the type_safe_union with a constructor and assignment operator
which can copy an object into the type safe union.

Gave the pipe the ability to disable dequeue operations.

Fixed a bug which caused the structural_svm_problem_threaded object to
crash when caching was used.

Added a version of the structural_svm_problem which can use multiple threads to
call the separation_oracle() function.

Added a missing #include

Increased the default values for the initial basis size and basis size
increment a little.

Added more comments.

Fixed typo in spec

Renamed all the rand::kernel_1a and rand::float_1a declarations
to just say rand.

Changed the rand object so you don't have to say rand::kernel_1a or
rand::float_1a anymore to declare it.  Now you just say rand.

Changed the stopping condition for the structured svm to the one suggested
by the Joachims paper.  I also added another useful reference to the docs.

Improved the feature vector caching in the structural_svm_problem object.

Changed the oca_problem interface to the oca optimizer.  Now the optimization_status()
function reports back the current risk and risk gap in addition to just the overall
objective value and objective gap.  Note that this change breaks backwards compatibility
with the previous interface (but only slightly).

Made assign_image() and assign_image_scaled() a little more general.  Now
they can use source images which are matrices in addition to array2d objects.
This change also enables the image_window to display matrices.

Added missing #includes

Added a unit test for the pyramid_down object.

Upgraded the pyramid_down function object so it can create color
pyramids.  Not just grayscale ones.

Added some functions which define the coordinate transforms
between different layers in an image pyramid.

A minor change to avoid getting a compiler error about functions
being ambiguous when the rectangle is used with floating point
vectors.

Added a missing const keyword.

Added some comments to clarify what exactly is a valid loss function.

The deserialize() for std::string was written under the assumption that std::string
doesn't store it's data in a contiguous block.  However, that isn't quite the case.
See this discussion:
http://stackoverflow.com/questions/1986966/does-s0-point-to-contiguous-characters-in-a-stdstring

Anyway, I changed deserialize() for std::string to be more efficient and read
directly into the string rather than into a contiguous buffer and then do a copy.

A minor usability improvement, I defaulted the feature_vector_type_
to matrix_type.

Cleaned up tests and added some additional ones.

A minor optimization.

Fixed grammar.

Added the max_sum_submatrix() function.

Changed the config_reader object so you don't have to say config_reader::kernel_1a
anymore to declare it.  Now you just say config_reader.

Changed the base64 object so you don't have to say base64::kernel_1a anymore to
declare it.  Now you just say base64.

You can now declare byte_orderer by saying just byte_orderer rather than
byte_orderer::kernel_1a.

Added a unit test for the byte orderer component

Changed the pipe so you don't have to say pipe<type>::kernel_1a anymore.  Now you
just say pipe<type>.

Changed the crc32 object so you don't have to say crc32::kernel_1a anymore to
declare it.  You just say crc32.  I also added some unit tests for it as well
as a few minor convenience functions.

Made it so you can reset a console_progress_indicator and use
it more than once.

The last commit left something in checking (i.e. _c mode).
Just switching it back to non-checking mode.

Added a missing requires clause to the get_sum_of_area() function
in the integral_image.  I also made the integral_image work
with signed pixel types.

Modified the probabilistic() trainer adapter (and the train_probabilistic_decision_function() routine)
so that they work with objects which have an interface compatible with std::vector rather than strictly
just std::vector objects.  For example, random_subset_selector objects can be used (as was the case in
some previous versions of dlib).

Added a value_type to the random_subset_selector so that it is more compatible
with the std::vector interface.

Fixed a compile time error that happens on debian's gcc 4.1.2

fixed grammar

minor change to avoid compiler warning in visual studio.

Added a unit test for the new svm_multiclass_linear_trainer.

Filled out spec and cleaned up code.

Added some missing DLIB_ASSERT statements.

Moved the new multiclass svm trainer into dlib.  Still need to clean up the code
and setup the abstract file.

Fixed compile time error

improved spec

updated spec

Fixed grammar

Added a structural SVM solver.

Minor cleanup, I removed unnecessary code.

Removed inappropriate uses of const on the be_verbose() functions.

Added some more tests for the max_cost_assignment() function.

Minor change to avoid a compiler warning in gcc.

Added the multiclass_linear_decision_function object.

It occurred to me that in some cases it is very easy to end up with dense vectors
which are slightly smaller than their corresponding sparse vectors.  So I changed
the requires clause of the dot(sparse,dense) and dot(dense,sparse) functions to not
require that the dense vector's size > sparse vector's size.  Now any combination
of sizes is legal.

Renamed assign_dense_to_sparse() to assign() and made it capable of also
assigning from sparse to sparse.

Clarified spec.

Added a dot() function which can dot a sparse vector with a dense vector.

This change is mainly about adding a result_type typedef to the various function objects.
Prior to this change, different function objects declared their return type in different ways,
now this has all been reconciled.  Now they all declare it as a public typedef named result_type.

I also simplified the cross_validate_multiclass_trainer(), cross_validate_trainer(),
test_binary_decision_function(), and test_multiclass_decision_function().  They now always
return double matrices regardless of any other consideration.

Added some tests for the LIBSVM formatted data IO functions
and related routines.

Added the fix_nonzero_indexing() function.

made test more robust

Simplified the oca implementation slightly.  It no longer keeps track of the
best point seen so far.  This change will allow me to use the optimizer in conjunction
with methods for quickly approximating subgradients and risk values without needing
to worry about the optimizer permanently fixating on what it erroneously thinks
is the best point.  Now the optimizer will just keep solving the cutting plane
subproblem and will use that solution as the current iterate and best known
solution.

Renamed max_index_value_plus_one() (a function for working with graphs) to
max_index_plus_one() so that it uses the same name as the essentially identical
function for working with sparse vectors.

Moved some functions for dealing with sparse vectors out of the svm_c_linear_trainer
implementation and into the sparse_vector.h header.  These are the add_to(), subtract_from(),
and max_index_plus_one() functions.

Added a value_type typedef to matrix_exp so it's easier to write templates
which operate on STL containers and matrix objects.

Removed some old comments

Fixed a bug in the kernel_matrix() function.  It didn't compile when used with
sparse samples which were of type std::vector<std::pair<> >.  Moreover, some of
the trainers have a dependency on kernel_matrix() so this fix makes those trainers
also work with this kind of sparse sample.

Added an is_pair template for telling if a type is a std::pair object.

Fixed a bug in load_libsvm_formatted_data().  Forgot to clear the contents
of the labels output vector before adding the loaded label data.

Fixed typo in spec

Changed the oca optimizer so that it warm starts the QP subproblem
rather than resolving it from scratch during each iteration. This
improves the speed and stability of the algorithm.

Changed code slightly to avoid a warning from gcc.

Made max_cost_assignment() take a matrix expression rather than an actual matrix.

Fixed bug in assert.

Added an implementation of the Hungarian algorithm for solving the optimal
assignment problem (in the new max_cost_assignment() routine).

Fixed typo in comment.

Fixed a problem which prevented the any_function unit test from compiling in visual studio 2008.

Added a test for split()

Added a function for splitting up strings which contain a sequence
of delimited tokens.

Added a unit test for the any_function.

Fixed a bug in any_function which prevented the .get() function
from compiling.

Made the is_function template work with functions of up to 10 arguments.
Previously it only worked up to 5 arguments.

Fixed a typo in the spec.

Added overloads of all the GUI event handlers so that now you can
use general functions as callbacks (via any_function).  This way,
if you have a C++0x compiler, you can use lambda functions with the
event handlers.

Switched the event handlers in the gui_widgets from member_function_pointers
to any_functions.

Fixed a typo in the spec for the is_empty() functions.  I also added an
is_set() to the any_function so that it's interface is similar to the
member_function_pointer.

Added a factory function to easily create temporary member function pointers.

Added an any_function which supports the same functionality as std::function from the
upcoming C++0x standard.  I added this so dlib can be modified to easily support lambda
functions while still being compilable on compilers which don't support the new
std::function.

Added tests for the new features of the distance_function object.

Fixed a runtime bug in the distance_function's + and - operators which triggered when
distance_functions with no basis vectors in them were added or subtracted.

Made the interface to the approximate_distance_function() a little cleaner and
improved its specification a bit.

Refactored the code in the reduced_decision_function_trainer2.  Part of it has been turned into
a global function called approximate_distance_function() which performs the main optimization.  The
reduced_decision_function_trainer2 now depends on this global function.  This changes makes this
function optimizer available for other purposes besides use in the reduced_decision_function_trainer2
object.

Improved the distance_function object by turning it into a properly encapsulated class
rather than just a simple struct.  I also added overloaded +, -, *, and / operators
for this object so you can do the kind of arithmetic you would expect on an object
which represents a point in a vector space.  This breaks backwards compatibility
with the previous interface though as the member variables are now private.

Fixed a few more compile time errors when dlib is compiled with clang.

I made some minor changes in a few places so that dlib compiles with the
clang compiler.

The any unit test was broken in a way which prevented it from actually running.
This has been fixed.

Applied a patch from Nils Labugt which fixes a bug in the gui_core component (Bug ID: 3171720).
The bug details are below:
    dlib GUI applications don't work anymore after I upgraded to Ubuntu 10.10 (AMD64). They
    consistently crash in base_window::set_title. XwcTextListToTextProperty returns
    XLocaleNotSupported instead of initializing 'property', but dlib ignores the return value,
    resulting in a crash in XLib when XSetWMName is called with garbage input or when
    property.value is XFreed without having been allocated.

Fixed a bug in an assert

Fixed errors in the spec

Fixed some comments.

Fixed the typo in the last commit which prevented compilation when ENABLE_ASSERTS
was defined.

Refactored the krr_trainer into two objects.  A rr_trainer which just does
linear ridge regression and the krr_trainer which uses the empirical_kernel_map
to do non-linear ridge regression.  No changes were made to the behavior of the
krr_trainer.  This update is just to allow the use of linear ridge regression
without a superfluous empirical_kernel_map running at the beginning of the training
process.

Adding a copy of the krr_trainer that will be just for use with linear kernels.

Added an overloaded operator/ to allow you to say things like 3.0/my_matrix

Removed the requirement that the lisf be non-empty when used to load an
EKM.  Instead, the load() function now will accept non-empty lisf objects
and throw a non-fatal exception.  This behavior should be slightly less
surprising to users, especially since certain degenerate datasets can give
rise to empty lisf objects when you might not expect it.

Minor fix to avoid a compile time error when using the kernel_matrix() function.

Updated the image_display widget so that you can zoom in and out using the mouse wheel.

Made test more robust

clarified a comment

Added wrappers for the new overloads which take std::string.

Added overloads to the config_reader's methods to allow it to load directly from
a file name given as a string in addition to taking istream objects.

Accidentally removed a needed #include yesterday.  Just putting it back.

Removed the ASSERT that the B matrix be symmetric since compiler optimizations
can sometimes cause B to be very slightly non-symmetric.

Added a comment

All I did in this change was rename memory_manager<char>::kernel_1a to default_memory_manager.
default_memory_manager is way more readable and clear.

Added a default_memory_manager typedef at the top of algs.h.

fixed typo in spec

Added a test to make sure the probabilistic() trainer adapter works right
and also that it works with the one_vs_all_trainer without issue.

Added a probabilistic trainer adapter.

Renamed generic_probabilistic_decision_function to probabilistic_function.

- Made the train_probabilistic_decision_function() more general by making it work
   with any kind of trainer object rather than only ones which produce
   dlib::decision_function objects.  I also made it work with trainers that only
   take std::vectors.
 - train_probabilistic_decision_function() no longer accepts column vectors of
   samples and labels.  Now it only accepts std::vectors of samples and labels.
 - Added a new generic_probabilistic_decision_function object which is returned by
   the train_probabilistic_decision_function().  The old probabilistic_decision_function
   still exists and can be constructed from a generic_probabilistic_decision_function.
   This way backwards compatibility is maintained with older code so long as that
   code used std::vectors to hold samples and labels.

Added a one vs. all multiclass trainer.

clarified spec

Moved null_df into its own file.

minor cleanup

Added a test for the cross_validate_regression_trainer() function.

Added cross_validate_regression_trainer() and test_regression_function()

Removed svn:executable from some files.

Minor change to avoid compile time error in gcc.

Added some unit tests for the one_vs_one multiclass training code.

clarified spec

minor change to avoid warning in gcc

minor cleanup

Added a bunch of new multiclass classification tools.

clarified spec

Removed an unused variable.

Fixed a minor bug in unordered_pair

I just retabbed this file

Added two new any object.  One for containing just decision function style objects and
another for trainers.

Fixed grammar

Moved basic_type out of any and made it a first class object with
proper documentation.

Removed the kernel_type typedef from the normalized function since this
meta-object should be capable of working with non-kernel decision functions.

Added the unordered_pair object.

Added typedefs for each of the template arguments to the type_safe_union.

Fixed a bug in the deserialization code for the sparse kernels.  The bug
prevented code which used the deserialization routine from compiling.

Added serialization support for std::pair objects.

Modified the type_safe_union so that you can call apply_to_contents() on const
type_safe_unions.

Added unit tests for the svr_trainer and svm_one_class_trainer.

Added an epsilon-insensitive support vector regression class.

Clarified spec

Simplified requires clause of train() function.

Added the is_learning_problem() predicate and used it to make a few
requires clauses more straight forward.

made spec clearer

Added a missing #include

fixed grammar

Added another citation

moved calculate_rho_and_b back into the svm_nu_trainer

Fixed a minor bug and did some cleanup

Added the svm_one_class_trainer object.

Made alpha initialization a little more robust.

Removed a bit of code that didn't actually do anything.

Fixed typo in comment

Added missing #include

Added some tests for the svm_c_trainer

Added the svm_c_trainer.  It uses the solve_qp3_using_smo optimizer I just added.

Moved calculate_rho_and_b() into it's own file since it will be used
by other tools.  I also clarified the spec slightly.

Added another version of the quadratic solver.  It's basically a copy
of solve_qp2_using_smo but this one solves the problem associated with
C-SVMs and a few other things.

Made the solver return the number of iterations performed.

fixed minor bug in unit tests

Clarified spec

A minor change to avoid a compiler warning from gcc.

Clarified some comments

Refactored the svm_nu_trainer.  Specifically, I pulled the quadratic
solver out and made it a separate class.  The kernel_matrix_cache has also
been removed in favor of the new symmetric_matrix_cache.  Finally, the
remaining bits of the svm_nu_trainer have been moved into svm_nu_trainer.h

Also note that invalid_svm_nu_error has been renamed to invalid_nu_error.

updated spec

removed some cruft

More code cleanup and spec refinement.  I also added a unit test for the symmetric_matrix_cache.
Finally, I put the new structs for determining the return types from colm/rowm/diag
into the proper places.

code cleanup

Added rough version of the new symmetric_matrix_cache() function.

Changed the kernel_matrix() function a little so it is easier for compilers to optimize.

Removed a pair of unused variables.

Changed the check for the radius being zero a little more robust against
floating point weirdness.

Minor change to avoid a compiler warning in gcc

made tests a little more robust

Made the requirement that the two images be different objects
a little more clear.

Improved the behavior of assign_image_scaled()

Fixed typo in comment

Fixed a bug in the signed/unsigned comparison code in the assign_pixel function.

Added a function called assign_image_scaled() which copies an image but also
does an intelligent scaling if the source and destination images have differing
dynamic ranges.  I also modified the image display widgets to use assign_image_scaled()
instead of assign_image().

I went through the image transforms and updated their specs and a few other
things so that they still make sense now that I'm letting pixels be signed
as well as unsigned.

Added missing requirement

Changed promote so that 8bit integers promote to 32bit integers.

changed to avoid compiler warnings in visual studio

Added the promote template which is a tool for converting from a smaller
scalar type to a bigger one.  The change in vector.h is just renaming to
avoid a naming conflict.

Changed the pixel_traits so that signed grayscale pixel types are allowed.  This
involved adding a few new fields to pixel_traits.  I also changed the get_pixel_intensity()
function so that its return value is of the same type as the basic pixel type rather
than always being unsigned long.

Added missing declaration for int64 in the is_built_in_scalar_type template.

Fixed typo

made timer test more robust

Fixed spelling in a comment

minor fix to a test

previous checkin broke some ASSERT statements.  Fixing this problem

made tests more robust

made test a little more reasonable

minor change to avoid a compiler warning in gcc

Changed code for the LM/quasi-newton model around a little to avoid
repeated calculation of things and also added some checks for division
by zero.

Changed the LM function model so that it computes the gradient
and hessian without creating an explicit Jacobian matrix.

Added a test for the any object.

Added an any container object.

Cleaned up a comment slightly.

Clarified spec

minor changes to avoid a compiler bug in visual studio 2005

Told the eigenvalue decomposition to always use the symmetric
eigen solver.

Fixed a compiler warning and also a runtime bug in sort_basis_vectors().
The bug triggered when all the basis vectors were included in the final
answer.

Replaced a funny character with the normal ASCII one.

Added a spec for the least squares stuff

Minor cosmetic changes to spec

- Added a missing operator*() for diagonal by diagonal matrix multiplication.
   Without it you would get an error about multiplication being ambiguous
   in this case.
 - Added an overload to catch expressions of the form diag_matrix*regular_matrix*diag_matrix
   and turn them into a form which is slightly more numerically stable in some cases.

Fixed grammar in a comment.

Made the rosen test a little more reasonable.

Minor change to avoid compiler warning.

Added tests for the least squares code.

Added some more optimization tests.

Added an #include for the least squares stuff

Added an initial version of some least squares solvers.

Minor changes to make code more robust.

added missing public: qualifier.

Made the trust region solver more robust.

Upgraded the derivative() function so it can work on functions with two arguments.

Added a trust region optimizer.

Added some comments to the spec.

Added some comments about how chol() behaves.

Added a bunch of overloads to catch operations on diagonal matrices
and use more efficient code paths for them.  For example, inv(diagm(d))
turns into diagm(reciprocal(d)).

Fixed tabbing

Added missing comment about the QP needing to be convex.

Added the ability to add/subtract scalar values to/from all the elements of a matrix
using the - and + operators.

Moved serialize definition outside class since some compilers will fail to compile
friend functions when they aren't declared in this way.

Updated a comment.

Added missing *= operator

Updated makefile

Increased the number of MCMC rounds so that the bayes nets test
is less likely to indicate failure.

Made the cmake file for finding blas more robust.

Added some code to avoid a bug in an older LAPACK implementation on Debian Etch.

Added flipud() and fliplr() matrix functions.  I also renamed
the scale_rows_columns.cpp test file to matrix4.cpp.

Added a missing requirement to sort_basis_vectors()

- Renamed the linearly_independent_subset_finder's dictionary_size() member function to
   size().  This way, linearly_independent_subset_finder objects can be used in many
   templated functions which expect objects which look like arrays.
 - Generalized the kernel_matrix() function slightly so that it can work with anything
   that looks like an array.  This now includes linearly_independent_subset_finder objects.

Removed the old is_same_object function from set_utils
since it conflicts with the new one in some cases.

Made this function capable of accepting anything that can be converted to
a matrix via vector_to_matrix()

made spec more clear

Added a function for performing supervised basis selection.

Added the is_same_object() function.

Updated spec to reference the correct LAPACK routine.

Clarified spec

Added some tests for the string_assign object.

Improved the string_assign utility.

Added the string_assign tool

Added a tool for measuring how long a task will take.

Switched the eigenvalue_decomposition from using syev to syevr since
syevr is supposed to be better.

Improved the way we link to the Intel MKL.  Now it will link
right against newer versions.

Minor change to avoid compiler warning.

Added missing std::abs() calls.

Made the code in chol() more robust to indefinite matrices.

Removed two cout statements I used for testing but forgot to remove before
committing code to subversion.  Oops.

Added a new scale_rows() function.  I also overloaded the * operator so that the
expressions mat*diagm(v) and diagm(v)*mat get bound to calls to scale_columns() and
scale_rows() respectively.

Added a comment about aliasing and the kernel_matrix expression.

Improved the alias detection capability of kernel_matrix() expressions.  Now statements
of the form: sample = kernel_matrix(kern, *, sample) can be used since the aliasing of
sample will be handled.

Changed the assert macros so that they don't use __PRETTY_FUNCTION__
with gcc 4.4.5 since, on Ubuntu at least, this version of gcc segfaults
when __PRETTY_FUNCTION__ is used within certain template constructs.

Added get_parent_directory()

Made X11 code a little more robust (I think).

Fixed file headers

Added a class for reading JPEG image files.

Improved the way we look for available BLAS and LAPACK libraries.

Added a compile time assert which makes sure you don't accidentally
give a non-linear kernel to the svm_c_linear_trainer.

Adjusted the lambda search code so that when two lambdas give the same
leave-one-out-error we always pick the bigger lambda.  This should be
slightly better since, all other things equal, bigger lambda usually
gives better generalization.

Added a test for make_symmetric()

Added some tests to exercise the column major code paths.

Minor change to avoid compiler warning in visual studio.

Setup the qr_decomposition to use LAPACK when available.  Also removed the
qr_decomposition::get_householder() function since I don't currently have any
way to test it or precisely define what it does.

Added some static_casts to avoid compiler warnings.

Added an overloaded matrix_assign() that handles symmetric kernel_matrix()
expressions more efficiently by only evaluating the upper triangular part
of the matrix.

Changed code to make sure it always calls the symmetric eigenvalue solver.

Made cholesky_decomposition use the triangular solver

Added file headers

Setup the LU decomposition code to use LAPACK when available.  I also removed the older
version from numerical recipes and made everything depend on the lu_decomposition object
instead.  Finally, I added in a triangular solver that uses BLAS when available and made
the lu_decomposition object us it.

Fixed typo in comment

Changed pinv() so it uses svd3() so that it can use LAPACK when available.

Changed code to avoid compile time bug.

Various minor fixes to the LAPACK bindings.

Added a make_symmetric() statement so that the most efficient eigenvalue_decomposition
method is used.

Made tests slightly more reasonable.

Setup the SVD routines to use LAPACK when available.  I also changed the svd functions
so that you can't supply output matrices which use both column and row major layouts.
Now all the output matrices need to use the same memory layout.

Made bindings more robust

Made the cholesky decomposition code use the xPOTRF routines in LAPACK

Pushed all the work variables into the LAPACK binding functions.

Setup the eigenvalue_decomposition to use LAPACK

Added some logic to find and link LAPACK libraries.

polished LAPACK bindings

Fixed misnamed function.

Added a bunch of LAPACK bindings.