* move some headers around
* add nuget
* Refactor bindingutilities and outputhelper into .cpp and header files
* remove dllload.cpp
* Remove windows.ai.machinelearning header
* Fix testing
* Don't change sln
* Fix x86 build
* Added different configurations
* Add preprocessor definitions to separate between headers and namespaces
* Change naming of EXE for nuget builds
* Update OutputHelper.cpp
remove pragma in outputhelper.cpp
* remove pragma from cpp
* Switch between native header and update documentation
* use const cast
* rename nuget to NuGet and address PR comments
Co-authored-by: Ryan Lai <ryalai96@gamil.com>
* check session options version
* Create Session Considering Support for Session Options
* Only set batchsize to 1
* Made batch size override always 1
* Merged
* Update minimum SDK
* Add -tensor preprocessing function options [Identity|ScaleMeanStdDev].
* Unify the csv and image reading mechanisms to unify the preprocessing code.
* Eliminate unnecessary copies of the csv data, expect for preprocessing pass.
* Update unit tests, including new DenseNet121 test for functional verification of preprocessing.
* -Tensor Normalize - update command to accept actual means and std devs on command line. Fixup tests and target csvs with new values.
* Tensorize edits to update spacing, and add a throw if fetching native tensor buffer fails.
* Evaluate folder of images
* Evaluate folder of images
* Updated usage
* Test change
* test changes back to test method cleanup
* Fix build break
* remove test
* Added test
* Fix comment typo
* Don't print out image path if using -terse flag
* Remove commitments
* Add feature to bind GPU tensors
* Add test to check for invalid arg
* Spacing
* Make heap properties and resource desc more simple
* Moved GPU upload tensor code under if statement
* Added support for models with image denotation
* Added bind image support for models with image denotation
* Added mnist and mnist input files
* Model tensor inputs should have channel 3 or 1 if binding image
* Add mnist tests
* Removed modelbinding.h
* PRcomments
* Fixed is image, is csv input check
* Add extra checking for fp16 output tensor tests
* Fixed var name
* Adjust values
* Made populating tensors more generic and tensor vectors are passed by reference
* Fixed formatting
* PR comments for renaming method
* Added comment
* Memset garbage binding input to 2 instead of 0
* Make FP16 tests more lenient
* Refactored Tensor creation to be cleaner. Garbage data is now arbitrary
* Added check when creating commandline args so that user cannot save tensors if no input data is provided
* Added FP16 tensor output support
* Using winmlrunner tests from master
* Added index 0 to for loop when printing out tensor
* Renamed tensor output files
* Added fp16 tests and made CompareTensors helper method more strict for tensor comparisons
* unused helper method
* Changed template for fp16 case
* Fixed spacing
* Adding tensor output expected files
* Addressed PR comments. Removed unneccessary code from template function
* Update GPUAdapterIndex option
* Support MCDM but use DIRECT command queue to allow GPU tensorization of image
* Test out DXCore
* Compatibility: Use DXGI adapter, 11_0 feature level and DIRECT command queue for regular GPU
* Use delay-load for dxgi and d3d11, use 18298 SDK and use static lib for debug build
* Remove dependency on kernel32.dll
* Support WCOSHeadless which doesn't have DXGI and D3D11
* Handle exception from delayed load
* Generate full PDB for debug build
* Handle delayed load SEH in a CreateDXGIFactorySEH
* Removing workaround for X86 DXCore, fix is grfx_dev 18318 and EEAP rs_prerelease 18323.1000.190119-1745 or newer
* Fix merge erros
* Delayload dxcore
* Throws error message if dxcore isn't supported
* Added extra checking for gpuadapter index flag argument
* Removed unnecessary files
* Added IsApiSetIMplemented
* Static link onecore on other architectures as well
* Added error checking to see if dxcore is available
* Prints out all adapter enumerations
* Added extra warning information about dxcore support.NEeds SDKat least 18323
* Stop outputting perf file for ever perf run
* Fixed indexing for commandline args
* PR feedback, removed WinMLRunnerDLL because this needs more work
* Changed flag from -output to -PerfOutput in WinMLRunnerTests
* Removed const for primitive types in commandlineargs
* Changed outputhelper reset so that it is more meaningful to only reset bind and eval times.
* Changed output directory for x86 to x86 instead of win32, to make it more uniform with x64 output directory
* Added ARM64 support
* Made release build static link vcruntime140.dll and MSVCP140.dll
* Changed SDK version on README.md
* Removed winml.lib dependency because not needed
* Removed winml.lib from x86 / debug
* Add a lot more perf metrics and granularity
* Add the Tensor Input Data Type
* Make the existing tests pass
* Add dynamic DLL loading at runtime
* Fix warnings
* Fix x86 warning
* Update README.md and UseGPU() function
* Add more tests for garbage input with the new flags
* Add tests for garbage data
* Revert temporary code
* Fixed typo
* Change old code
* Change tabs for spaces
* Fix bad IF statement
* Fix bad IF statement
* Revert non-backward-compatible change
* Revert non-backward-compatible changes
* Fix typo in README.md
* Fix "totalTime" showing "nan" in the CSV
* Retarget the test Windows SDK
If both projects target the same Windows SDK, we don't need to install
an older SDK just to run the tests.
* Fix the random garbage generator
* Fix the last failing test
* Update README.md
* add more entries to squeezenet and mnist test
* Add style transfer test
* add teardown & fix timeout handling for style transfer test
This change includes Squeezenet Samples test (CS project) using kitten image.
Both squeezenet test and mnist test will be run with vstest with MSTest adapter.
Brian Martin
Ryan Lai
Xiang Zhang
pmbrown1055
Ryan Lai
Ori Levari
Ryan Lai
Young Kim
laks0209
Patrice Vignola
Kumar Rajeev