* Specification for the NMS-4 operation (updated shape infer function)
* Enabled NMS-4 in the Model Optimizer
* Changed opset version for NMS with dynamic outputs and namespace to be "dynamic"
* Added NMS-4
* Added opset4 to the nGraph
* Added unit tests for NMS-4 type infer
* Renamed UpgradeNMS3ToNMS4 to UpgradeNMS3ToNMSDynamic. Added stub for ConvertNMS4ToLegacy
* Make IE aware of opset4 ops
* Updated NMSIE to have different shape infer function based on the NMS it was converted from. Implemented NMS4->NMSIE conversion
* Apply code style
* Updated StaticShapeNonMaximumSuppression op in the VPU
* Introduced new version of NMSIE operation with shape infer function from v4::NMS
* Fixed dynamicToStaticNonMaxSuppression transformation
* Added new version of NMSIE op with updated shape infer function
* Fixed NMS4 to NMSIE2 transformation
* Fixed constructors for nGraph ops v4::NM and dynamic::NMS
* Updated text in the opset4 specification document
* Code style fixes
* Fixed constructors for StaticShapeNMS + fixed test
* Minor change to the NMS op in the MO
* Fixed typo in the dynamic_to_static_shape_non_max_suppression transformation
* Removed redundant checks
* Refactored NMS infer and validate functions
* Added more checks to the validate_and_infer_types functions for NMS-3 and NMS-4
* Fixed compilation issue on Windows for op NMS
* Code style fixes
* Fixed typos in the NMSIE and NMSIE2 to CNNLayer op conversion
* Fixed typo in the ie_cnn_layer_builder_ngraph.cpp
* Fixed the NMSToLegacyNMS transformation. Added unit tests
* Apply code review comments
* Refactored NMSIE to use visitors
* Removed calling ConvertNMS4ToLegacy in the common optimizations
* Moved NMS4ToNMSLegacy to convert1_to_legacy group of transformations
* Removed useless include statement
* Removed copy-paste issue
Co-authored-by: Evgeny Lazarev <elazarev.nnov@gmail.com>
* Fixed deleting Transpose layers after and before Interpolate layers.
* Added run_after() for the transformation InterpolateTranspose.
* Some checks were moved from the replacement function to the pattern.
* Added a check of the attribute 'axes' into the pattern.
* LayerNorm(PyTorch/HuggingFace pattern)->MVN+Mul+Add. Improves perf on BERT by 5%
* deducing the across_channels from axes passed to the MVN op.
axes are normalized. if no axes is specified, falling back to the (previously) default across_channel value
Co-authored-by: myshevts <maim.y.shevtsov@intel.com>
* Removed back phase transformations related to IRv7
* Fixed setting value for the input port using the 'set_value' method
* Removed front and middle phase transformations related to IRv7
* Cleanup the rest of the Model Optimizer transformations from IRv7 specific transformations
* Final cleanup of the deprecated IR v7 related code
* Removed 'blobs_as_input' usage in the Model Optimizer.
* Removed function '_fuse_add' from the Model Optimizer since it is not used anymore.
* Removed 'keep_in_IR' node attribute for FakeQuantize ops in the MO
* Disabled failing gpu_engine.user_context test
* [MO] Implement EmbeddingBag_3
* Transform dynamic sub-graph of Wide and Deep into EmbeddingSegmentsSum
- Expressed SparseWeightedSum sub-graph through EmbeddingSegmentsSum
- Removed experimental SparseWeightedSum layer
- Implemented tests for the transformation
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Fix EmbeddingBag shape infer
* Fix EmbeddingSegmentsSum transformation for Wide and Deep
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Fix EmbeddingSegmentSum replacer after ports swap
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Update package_BOM.txt
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Add unit tests for EmbeddingXXX shape infer
* Fix ATen resolver
* Remove deleted files from BOM
* Add opset version to embedding_bag
* Use base class for EmbeddingBag
* Fix per_sample_weights case
* Fix EmbeddingSegmentsSum transformation
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Fix EmbeddingBag checks
* Fix ATen front transformation and merge conflicts
* Fix BOM
* Work around limitation for I64 input of W&D model
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Cleanup where operation to fix affect of WhereDecomposition transform
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Fix BOM
* Correct EmbeddingSegmentSum transform for Wide and Deep
Add casting segment ids to i32 and remove ConstToResult sub-graph.
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Update BOM with RemoveConstToResult transform
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Add more comments for RemoveConstToResult transformation
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Remove useless logging in EmbeddingSegmentsSum transformation
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Small fixes
* Move EmbeddingBag resolving back to front phase
* Improve error messages
* Fix typo in unittests
* Reimplement sparse_reshape middle transform
Avoid deprecated API.
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Clean-up graph after sparse_reshape and ConstToResult transformation
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Fix clean-up for transformations
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Fix clean-up for transformation #2
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
Co-authored-by: Roman Kazantsev <roman.kazantsev@intel.com>
WhereDecomposition transform is applied to Where operation in for-garbage sub-graph remained after SparseWeightedSum transform.
Signed-off-by: Roman Kazantsev <roman.kazantsev@intel.com>
* Small fix in the transformation ConvertGroupedStridedSlice. Now VariadicSplit is generated only in the case when node has at least 2 output nodes.
* Added unittests for the case when there is only one StridedSlice.
* Fixed ONNX Mask-RCNN conversion
* Fixed validate_and_infet_types for NMS ops: added check for number of connected inputs
* Updated NMS ops to properly handle optional input with index 2
* Fixed typo in the implementation
When stdout is not a terminal, Python will buffer it by default. This
means that a consumer of MO's output will not see the argument information
until the buffer is flushed, which will normally only happen once MO
finishes (which might take a while).
Flushing stdout explicitly allows the consumer to see this info as soon
as it's printed.
