[MO] StridedSlice improvements

[pavel-esir]

Description: Fixed some issues with StridedSlice permutation and processing ellipsis masks. Now infer function do not change attributes and inputs of StridedSlice, moved normalization into a separate transformation, also reused general existing permute mechanism. Now complex StridedSlices are correctly inferred in ShapeOf subgraphs.

tickets 40372, 43072, 39311

Code:

• Comments
• Code style (PEP8)
• Transformation generates reshape-able IR
• Transformation preserves original framework node names

Validation:

• Unit tests
• Framework operation tests
• Transformation tests
• e2e model test with an update of ./tests/e2e_oss/utils/reshape_utils.py
• Model Optimizer IR Reader check

Documentation:

• Supported frameworks operations list: N/A
• Supported public models list: N/A
• New operations specification: N/A
• Guide on how to convert the public model: N/A
• User guide update: N/A

[lazarevevgeny]

In general looks good, but still have some comments.

Also, please, run unit test with code coverage to see how new functions are covered with unit tests: coverage run -m unittest discover -p "*_test.py"

[pavel-esir]

@lazarevevgeny thanks for your comments! Resolved them all.
coverage run -m unittest discover -p "*_test.py" - successfully done

There are 3 sets of automatically unit-tests that cover almost all cases.
ShapeInfer tests with shrink_axis, newaxis and ellipsis placed on almost every possible positions of 2, 3, 4 rank slices with different input_ranks.

openvino/model-optimizer/mo/ops/strided_slice_test.py

Line 511 in 73046d7

# automatically generated the whole range of 2d slices over 2d, 3d and 4d input tensors

The same ShapeInfer tests after Normalizer

openvino/model-optimizer/extensions/middle/StridedSliceNormalizer_test.py

Line 395 in 73046d7

# automatically generated the whole range of 2d slices over 2d, 3d and 4d input tensors

Permute tests

openvino/model-optimizer/extensions/middle/StridedSliceNormalizer_test.py

Line 1660 in 73046d7

# automatically generated permutation tests

dl-benchmark finished without regressions on a previous run. @jane-intel, @sadolini, @rkazants please review

[jane-intel]

First part 14/20 files

[lazarevevgeny]

Overall looks good. But left several minor comments to be fixed.

[likholat]

According to the documentation, the stride input is optional: https://docs.openvinotoolkit.org/latest/openvino_docs_ops_movement_StridedSlice_1.html
But after these changes, the absence of the stride input causes the error:

[ ERROR ]  Traceback (most recent call last):
  File "/home/anna/projects/openvino/model-optimizer/mo/utils/class_registration.py", line 290, in apply_transform
    for_graph_and_each_sub_graph_recursively(graph, replacer.find_and_replace_pattern)
  File "/home/anna/projects/openvino/model-optimizer/mo/middle/pattern_match.py", line 59, in for_graph_and_each_sub_graph_recursively
    func(graph)
  File "/home/anna/projects/openvino/model-optimizer/extensions/middle/StridedSliceNormalizer.py", line 107, in find_and_replace_pattern
    StridedSliceNormalizer.normalize_strided_slice(graph, node)
  File "/home/anna/projects/openvino/model-optimizer/extensions/middle/StridedSliceNormalizer.py", line 129, in normalize_strided_slice
    StridedSliceNormalizer.normalize_slices_attr(node)
  File "/home/anna/projects/openvino/model-optimizer/extensions/middle/StridedSliceNormalizer.py", line 249, in normalize_slices_attr
    res_slices[-1] = slice(*res_slices[-1].indices(data_shape[in_idx]))  # convert negative begins/ends
TypeError: slice indices must be integers or None or have an __index__ method

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/home/anna/projects/openvino/model-optimizer/mo/main.py", line 305, in main
    ret_code = driver(argv)
  File "/home/anna/projects/openvino/model-optimizer/mo/main.py", line 269, in driver
    ret_res = emit_ir(prepare_ir(argv), argv)
  File "/home/anna/projects/openvino_contrib/modules/mo_pytorch/mo_pytorch.py", line 106, in _prepare_ir
    graph = unified_pipeline(argv)
  File "/home/anna/projects/openvino/model-optimizer/mo/pipeline/unified.py", line 29, in unified_pipeline
    class_registration.ClassType.BACK_REPLACER
  File "/home/anna/projects/openvino/model-optimizer/mo/utils/class_registration.py", line 340, in apply_replacements
    apply_replacements_list(graph, replacers_order)
  File "/home/anna/projects/openvino/model-optimizer/mo/utils/class_registration.py", line 330, in apply_replacements_list
    num_transforms=len(replacers_order))
  File "/home/anna/projects/openvino/model-optimizer/mo/utils/logger.py", line 124, in wrapper
    function(*args, **kwargs)
  File "/home/anna/projects/openvino/model-optimizer/mo/utils/class_registration.py", line 318, in apply_transform
    )) from err
Exception: Exception occurred during running replacer "REPLACEMENT_ID (<class 'extensions.middle.StridedSliceNormalizer.StridedSliceNormalizer'>)": slice indices must be integers or None or have an __index__ method

Is it a bug or the stride input is now required?

[pavel-esir]

Is it a bug or the stride input is now required?

Specification haven't been changed. strides are still optional. Looks like a bug. I will take a look