New Features and Improvements:
   - Added Beta distribution to dlib::rand.
   - Added directory_exists.
   - Added bgr_alpha_pixel type.
   - Added support for loading and saving WebP images.
   - Deep learning tooling:
      - Added ReOrg layer.
      - Added SiLU layer.
      - Added visitor to draw network architectures using the DOT language.
      - Made Barlow Twins loss much faster for high dimensionality inputs.
      - Added Focal loss gamma to loss_multibinary_log_.
      - Added SmeLU activation layer.

Non-Backwards Compatible Changes:
   - Do not round coordinates in rectangle_transform (PR #2498).

Bug fixes:
   - Fix missing ssize_t definition on Windows (PR #2492)
   - Fix Barlow Twins loss gradient (PR #2518)
   - Fix bug with some versions of cuDNN (PR #2527)
   - Fix python import errors on some systems.
   - Various cmake adjustments to avoid build failures on some systems.

New Features and Improvements:
   - Added serialization support for std::variant and std::optional.
   - Added upgrades to dlib::type_safe_union: type_safe_union_size, type_safe_union_alternative_t,
     made into variadic template, more generic visit() method.
   - Added dlib/invoke.hh which contains C++11 backports of std::invoke and related tooling.
   - Added letterbox_image()
   - Exposed momentum_filter to Python.
   - Deep learning tooling:
      - Added Clipped ReLU and ELU activations
      - Made input_layer() work with networks that contain repeat layers.
      - Added loss_yolo_ for training YOLO style object detectors.
      - Added support for fused convolutions, see dlib::fuse_layers().
      - Added dnn self supervised learning example: dnn_self_supervised_learning_ex.cpp

Non-Backwards Compatible Changes:

Bug fixes:
   - Fix input/output mappings with repeat layers (PR #2337)
   - Fix disable_duplicative_biases when the input is a skip layer (PR #2367)
   - Fix cannot compile iso only code (PR #2384)
   - Fix running gradient crashing sometimes (PR #2401)
   - Fix trainer with unsupervised loss (PR #2436)
   - Fix dnn_trainer trying to decrease the learning rate (PR #2442)
   - Fix Layer Normalize (PR #2489)

New Features and Improvements:
   - Deep learning tooling:
      - Added loss_multibinary_log_
      - Added scale_prev layer
      - Various ease of use improvements to the deep learning tooling, such as improved layer
        visitors and increased DNN training stability.
      - Added CUDA implementation for loss_multiclass_log_per_pixel_weighted.
      - Add GELU activation layer
      - Add Layer Normalization
      - Add CIFAR-10 dataset loader: load_cifar_10_dataset()
   - Add probability_values_are_increasing() and probability_values_are_increasing_robust().
   - Expanded list of serializable types and added DLIB_DEFINE_DEFAULT_SERIALIZATION, a macro that
     lets you make a class serializable with a single simple declaration.
   - Added exponential and Weibull distributions to dlib::rand.
   - For dlib::matrix:
         - Added soft_max() and pointwise_pow()
         - The FFT methods now support arbitrary sized FFTs and are more performant.
   - Added user definable stopping condition support to find_min_global() and find_max_global().

Non-Backwards Compatible Changes:
   - Rename POSIX macro to DLIB_POSIX to avoid name clashes with some libraries.
   - Dropped support for gcc 4.8.

Bug fixes:
   - Fixed bug in loss_mmod that degraded the quality of bounding box regression.  Now
     bounding box regression works a lot better.
   - Fixes for code not compiling in various environments and support newer CUDA tooling.

New Features and Improvements:
   - Added support for cuDNN 8.0.
   - Added support for CUDA in Python 3.8 on Windows.
   - Allow forwarding initial function evaluations into find_max_global().
   - Can now load PNG images from a memory buffer.

Non-Backwards Compatible Changes:

Bug fixes:
   - Improved error messages for python users trying to install dlib
     with CUDA enabled or who are using windows.
   - Fix random forest regression not doing quite the right thing.

New Features and Improvements:
   - Added CUDA implementation for loss_mean_squared_per_channel_and_pixel.
   - Added DCGAN example in examples/dnn_dcgan_train_ex.cpp
   - Added transfer learning example in examples/dnn_introduction3_ex.cpp
   - Added leaky_relu activation layer
   - Added mish activation layer 
   - Added loss_multiclass_log_weighted
   - Added loss_binary_log_per_pixel
   - Minor API simplifications in deep learning tooling.
   - imglab can automatically select the number of image clusters when --cluster 0 is given.
   - Added a relative epsilon termination option to svm_c_linear_trainer
   - Support new version of OpenCV that doesn't have IplImage anymore.

Non-Backwards Compatible Changes:

Bug fixes:
   - Corrected interpolate_bilinear for lab_pixel.
   - Fix build errors in CUDA 10.2
   - Make equal_error_rate() handle degenerate case where all scores are equal.
   - Fix DLIB_ISO_CPP_ONLY not working
   - Fix build errors in C++20 build mode.
   - Fixed function_evaluation_request::set() invalidating function_evaluation_request::x()

New Features and Improvements:
   - Made find_min_global() and find_max_global() much faster when using them
     to optimize inexpensive functions.  These tools now measure the runtime of
     the objective function and automatically tune themselves to balance solver
     overhead vs objective function execution time.
   - Improved handling of label and ignore boxes in loss_mmod layer.
     See here for more details: https://github.com/davisking/dlib/issues/1894
   - Better auto-detection of build configurations on some platforms.
   - Shiny new deep instance segmentation examples:
     dnn_instance_segmentation_ex.cpp and dnn_instance_segmentation_train_ex.cpp.  

Non-Backwards Compatible Changes:

Bug fixes:
   - Fix build error in some versions of visual studio in some setups.
   - Fix find_max() going into an infinite loop in some cases when a
     non-differentiable function is given.

New Features and Improvements:
   - Make dlib.full_object_detection take list of dlib.point or dlib.points.
   - CMake scripts will automatically include the Intel MKL's iomp dll in the
     output folder to reduce confusion for windows users who haven't added the
     Intel MKL to their PATH.  This avoids a dll not found error that would
     otherwise result in that situation.
   - load_jpeg() now supports loading from a memory buffer.
   - Added input_grayscale_image_pyramid DNN input layer.
   - Added loss_mean_squared_per_channel DNN input layer.
   - Added methods for getting keyboard and mouse clicks to image_window's python API.
   - Made pkg-config report all needed include and link settings to use dlib.

Non-Backwards Compatible Changes:

Bug fixes:
   - Setting a point's y coordinate changed x instead, when done from Python.
     This has been fixed.
   - Fixed global_function_search's initialization being wrong if explicitly
     given an empty list of initial function evaluations.
   - Fixed compile time errors on some uncommon system configurations.
   - Work around bugs in CUDA runtime on some systems.
   - Various improvements to CMake built scripts to handle unusual system configurations.
   - Fixed bug in multi-GPU training of semantic segmentation models.

New Features and Improvements:
   - Added weighted labels to loss_binary_log layer.
   - Added padding option to Python compute_face_descriptor() methods.
   - Added pointwise_divide function.
   - Added U-net style skip connections to the semantic-segmentation example.
   - Exposed dlib::chinese_whispers() directly in Python API.
   - Added python api that generates desciptor(s) from the aligned image(s).
   - Added support for CUDA 10.1.
   - Dlib's python install process always automatically enables all supported features.  
     So --yes <option> was removed as it does nothing.
   - Added --box-images to imglab.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed build errors on some systems or buggy compilers.
   - Workaround for bug in cudaStreamSynchronize() causing hanging in some rare cases.
   - Fixed broken CUDA memcpy overload.
   - Fixed CMake not finding cusolver in CUDA 10.1. 

New Features and Improvements:
   - Allow 64bit seeds for murmur hash.
   - Some code paths are a little faster.
   - Added an option to do bounding box regression to the loss_mmod layer.
   - Added resize_image(img,scale) to Python API.

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed build errors in some environments.
   - Make older versions of CMake work with CUDA 10.
   - Fixed is_image_type template always reporting false. This also caused
     have_same_dimensions() to be uncallable for images, so that's fixed now too.

New Features and Improvements:

Non-Backwards Compatible Changes:

Bug fixes:
   - Fixed sub_image() in the Python API not working correctly for non-8bit images.
   - Last release broke have_same_dimensions(). The version specific to images
     was being used for tensors, which is incorrect and caused some of the DNN
     tooling to malfunction.