* New approach to offline transformations
* Include paths fix
* Imports fix
* offline_transformations target dependency simplification
* MakeStatefulTransformation exposed to python
* Python bindings build configuration fix
* Test variable name refactor
* Stop crying (snow)flake
* Snake cased offline transformations API
* Removal of old offline transformations from python
* Imports adaptation to new code style
* offline_transformations as a part of the common wheel
* Cmake simplification and refactor
* Correct transform invocation
* CI fix
* Proper dependency check in MO
* _pyngraph as a dependency of MO in cmake
* IR serialization fix in MO
* POT adaptation to the new API
* Revert "Removal of old offline transformations from python"
This reverts commit f9a0551ead.
* Merge of old& new bindings for offline_transformations
* Revert "POT adaptation to the new API"
This reverts commit 499554e68c.
* Obsolete cmake line removal
* Missing comma and merge conflict fix
* Offline transformations tests fix
* IE imports removal from check_ie_bindings
* Installation of opevino/__init__.py fix
* Obsolete line removal
* MO serialization switched to the new API
* Revert of preliminary MO adaptation to the new API
* Another magic spell that will hopefully make CI pass
* Python api cmake dependencies reorg
* Temporary solution for the CI/cpack errors
* Installation fix and code formatting
* ie_api & pyopenvino dependency removal
* Explicit cpack configuration for the new API
* cpack configuration adaptation
* Revert of obsolete cpack changes
Co-authored-by: Alexander Zhogov <alexander.zhogov@intel.com>
* enable itt trace for AUTO
Signed-off-by: fishbell <bell.song@intel.com>
* set itt to real loadnetwork path
Signed-off-by: fishbell <bell.song@intel.com>
* refine itt tracing
Signed-off-by: fishbell <bell.song@intel.com>
* formatting
Signed-off-by: fishbell <bell.song@intel.com>
* [LPT] optimize Subtract with zero_point == 0 when it's not a direct Constant
zero_point may be a Constant, but also it may come in form of
Constant->Convert(from low to high precision) subgraph.
If we handle both cases, we can reduce redundant Subtract node
e.g. between weights and Convolution which makes Convolution
to work fully on low precision inputs which significantly improves
performance.
* don't round shift if it's already in low precision
* Fix 'cannot estimate element if precision is UNSPECIFIED' error caused by LPT
In some models after ConvMulFusion and FakeQuantizeMulFusion, output_low contains
denorms. That is problematic since LayerTransformation::getPrecisionDetails function checks
if output_low is close to zero and if it is - the function sets 'signedPrecision' flag to false.
In that case, both 'signedPrecision' and 'unsignedPrecision' are set to false and that makes
getPrecisionDetails return element::undefined.
This patch changes zeroThreshold to handle denorms.
Ticket: 65375
* fix FakeQuantizeTransformation tests
* Add test cases for PReLU in cpu plugin
* For case when slope is vector
* Add Constant template plugin reference tests
* Update CMakeLists.txt and delete constant.in.cpp
* Add tests of tensor_2constant and constant_multi_use
* Add test of constant_equality_bool
* Remove wrong comments
* Remove some of strange if
* Merge to one CreateFunction
* Remove test names and update test for types
* Add bf16 and f64 tests
* Add missing type tests
* Clear actualOutData to allow multiple use of Validate()
* Update SetUp and CreateFunction to support CentOS CI
* Remove inputData = {}
* Remove fp16 of Convert layer test from skip_tests.config.cpp as it works now
* update repo
* add op reference test of experimental_detectron_detection_output and experimental_detecteion_generate_proposals and remove ngraph backend test
* modify visitor api test of experimental_detectron_detection for clang-format
* modify visitor api test of experimental_detectron_detection for clang-format again
* modify visitor api test of experimental_detectron_detection for clang-format again2
