GitLab

The loss value should reflect the impact of the gamma parameter, as well.

Warnings about unused variables are quenched by simply incrementing the
variables in question. This makes the try_compile() call in CMake
succeed without warnings when checking for C++11 compatibility and using
-Wall.

When propagating the option "-Xclang -fcxx-exceptions" to targets
linking to dlib, the options with quotes are unknown to clang-cl.
Removing the quotes and splitting the arguments into two works fine.

Clang 7 increased the default template-depth to 1024.

When compiling with other compilers than MSVC, error messages about
C++11 support for versions below MSVC version 19.0.24210.0 were printed.
This happens, for example, when using clang-cl installed with the Visual
Studio Installer.

* Add focal gamma to loss_multibinary_log

* update release notes

* Update docs

* escape & in git_logs_as_xml

Notably, set_all_bn_running_stats_window_sizes and fuse_layers.

But also I took the chance to remove the superflous separators and
change the attribute of upsample layers from stride to scale.

* Refactor visitors into their own header

* Update visitor links in term index

* Move documentation from layers_abstract

* Add visitor to draw nets in the DOT language

* Add documentation

Problem: `Error while calling cudnnGetConvolutionForwardWorkspaceSize( context(), descriptor(data), (const cudnnFilterDescriptor_t)filter_handle, (const cudnnConvolutionDescriptor_t)conv_handle, descriptor(dest_desc), (cudnnConvolutionFwdAlgo_t)forward_algo, &forward_workspace_size_in_bytes) in file C:\a\2\s\3rdparty\dlib\dlib\cuda\cudnn_dlibapi.cpp:1029. code: 9, reason: CUDNN_STATUS_NOT_SUPPORTED` (#2532)

Solution: when this happens, select the best algorithms again - but this time bypassing the cache

Problem: Error while calling cudnnGetConvolutionBackwardFilterWorkspaceSize( context(), descriptor(data), descriptor(dest_desc), (const cudnnConvolutionDescriptor_t)conv_handle, (const cudnnFilterDescriptor_t)filter_handle, (cudnnConvolutionBwdFilterAlgo_t)backward_filters_algo, &backward_filters_workspace_size_in_bytes) in file C:\a\1\s\3rdparty\dlib\dlib\cuda\cudnn_dlibapi.cpp:1048. code: 9, reason: CUDNN_STATUS_NOT_SUPPORTED (#2527)

Solution: Do not apply the cuDNN 5.1 workaround if we have cuDNN version greater than or equal to 7

* Fix Barlow Twins loss gradient

* Update reference test accuracy after fix

* Round the empirical cross-correlation matrix

Just a tiny modification that allows the values to actually reach 255 (perfect white).

* Make github actions run the python unit tests

* why don't the pyhton tests run?

* run python tests

* don't run tests twice

* don't run cmake tests twice

* rename the CMake test to C++ since it's testing the C++ part of the library

And the python build uses cmake too, which makes the naming somewhat
confusing

* work around macos

* oops, fix spelling error

* Refactor not to use a matrix

* Add clang build

* Add job on clang-13 as well

* Disable the `gcc-11` job for now

* Add get_random_beta

* simplify beta distribution generation

* add tests

* fix condition and use full names for shape parameters

* remove unneeded include

* match test types to prevent Windows from failing tests

* Revert "remove unneeded include"

This reverts commit 35f55a96e91f4c4340a0bda344c494f1ce68d521.

* Revert "match test types to prevent Windows from failing tests"

This reverts commit a990307066e99cb56e15ae5de63e67e29cc72911.

* Revert "Revert "remove unneeded include""

This reverts commit 59be002c9e8b8a6ac395d87e6bf9d4b57af19d28.

* fix outdated message in DLIB_CASSERT

* relax mpc condition again

* Revert "relax mpc condition again"

This reverts commit 1d208c5dcf1ddc15e09e68e1e2e11145506729ae.

* Loop while both alpha and beta are zero

* Update dlib/rand/rand_kernel_abstract.h
Co-authored-by: Davis E. King <davis685@gmail.com>