IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ONNX Loop operation support (#2756)

Browse Source 
* Generate TensorIterator without back edges from TensorFlow models

* Added a check in the MarkSubgraphsWithCorrectLayout to not fail when port is not connected

* Updated the 'protobuf2nx' to consume the graph protobuf message

* Cleanup TI from the IRv7 specific code

* Do not run some front transformations recursively

* Draft support for the ONNX Loop operation when 'cond' = True

* LoopToTI transformation changes

* Added draft of Loop operation and parser for ONNX Loop operation body

* Updated Loop body parser + added shape and type infer for the Loop operation

* Fixes for ONNX Loop operation parser

* Moved Loop parsing to Loop op extractor. Added generation of external edges for the Loop body ops

* Added support for ThresholdedRelu using decomposition

* Added support for Min ONNX operation

* Draft fixes for port_map generation for the Loop

* Rename transformation file and fix BOM

* Fixed shape inference for Loop scan outputs (axis is not None)

* Fixed shape inference for ONNX Loop operation

* Refactor checks in the TensorIteratorMerge transformation

* Code refactoring. Enabled commented transformations

* Documentation update for ONNX Loop, ThresholdedRelu and Min

* Fixed typo in the Loop front transformation where execution condition input is connected. Other refactorings

* Fixed in the Loop extractor

* Added printing 'internal_layer_id' attribute in the graph dumper

* Updated calculation of iterations number for the Loop

* Added missing code

* Fixed output port shapes and types generation for Loop operation

* Update function names and variable names in the Loop operation

* Fixed type inference for iteration count input

* Added removal of input/output ports of the Loop if they are not used

* Fixed renumbering Loop operations input/output ports to keep mandatory

* Fixed ThresholdedReluDecomposition transformation

* Updated MO IR Reader to know about Loop operation. But it is still not supported by the MO IR Reader

* Added unit test for Slice op shape infer (reverse the sequence of elements)

* Reverted changes in the ONNX loader function call to protobuf2nx

* Enable Reshape0DToSqueeze transformation recursively

* Refactored Loop operation support implementation

* Changed ThresholdedReluDecomposition to generate Const with shape [1] instead of scalar

* Code style and wording fixes

* Restored accidentally removed 'return' statement in the TI shape infer function

* Fixed comments

* Fixed comment

Co-authored-by: Evgeny Lazarev <elazarev.nnov@gmail.com>
This commit is contained in:
Evgeny Lazarev
2020-10-27 23:04:43 +03:00
committed by GitHub
parent f6842863c7
commit cabf8d8534
28 changed files with 1022 additions and 110 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/ops/infrastructure/Loop_5.md
View File

@@ -85,7 +85,7 @@ Loop operation description in the IR also has several special sections: `body`,
* **Port map attributes**:
* *external_port_id*
* **Description**: *external_port_id* is a port ID of the `Loop` operation.
* **Description**: *external_port_id* is a port ID of the `Loop` operation. The value `-1` means that the body node is not connected to the `Loop` operation.
* **Range of values**: IDs of the *Loop* outputs
* **Type**: `int`
* **Default value**: None