[TF FE]: Support complex tensors for Gather, GatherV2 operations

[NingLi670]

Details:

• Support complex tensors for Gather, GatherV2 operations
• Add corresponding layer test

Tickets:

• [Good First Issue][TF FE]: Support complex tensors for Gather, GatherV2 operations #22951

[rkazants]

build_jenkins

[NingLi670]

@rkazants I'm not quite familiar with the error [TensorFlow Frontend] Internal error, no translator found for operation(s): ComplexTypeMark. Do you know how I can locate the code where the bug occurs?

FAILED install/tests/layer_tests/tensorflow_tests/test_tf_Gather.py::TestComplexGather::test_gather[ ie_device:CPU - precision:FP16 - indices_type:<class 'numpy.int32'> - params_type:<class 'numpy.float32'> - params:{'params_shape': [3, 2, 6, 4], 'indices_shape': [3, 2, 1, 3], 'axis_value': -1, 'batch_dims': -2, 'operation_type': 'GatherV2'} ] - openvino._pyopenvino.OpConversionFailure: Check 'is_conversion_successful' failed at src/frontends/tensorflow/src/frontend.cpp:478:
FrontEnd API failed with OpConversionFailure:
[TensorFlow Frontend] Internal error, no translator found for operation(s): ComplexTypeMark
To facilitate the conversion of unsupported operations, refer to Frontend Extension documentation: https://docs.openvino.ai/latest/openvino_docs_Extensibility_UG_Frontend_Extensions.html

[NingLi670]

@rkazants
Using inputsparams.shape=[3,2,6,4], indices.shape=[3,2,1,3], axis=-1, batch_dims=-2 gets a resulting shape of [3, 2, 6, 1, 3].
When params is complex tensors:
Using inputsparams.shape=[3,2,6,4,2], indices.shape=[3,2,1,3], axis=-2, batch_dims=-3 gets a resulting shape of [3, 2, 6, 2, 1, 3, 2], which is wrong. While using inputsparams.shape=[3,2,6,4,2], indices.shape=[3,2,1,3], axis=3, batch_dims=2 gets a resulting shape of [3, 2, 6, 1, 3, 2]. The later is right. It seems that simplily minusing 1 when axis and batch_dims are negative values is not always right. Do you have any suggestions?

[NingLi670]

@rkazants Using inputsparams.shape=[3,2,6,4], indices.shape=[3,2,1,3], axis=-1, batch_dims=-2 gets a resulting shape of [3, 2, 6, 1, 3]. When params is complex tensors: Using inputsparams.shape=[3,2,6,4,2], indices.shape=[3,2,1,3], axis=-2, batch_dims=-3 gets a resulting shape of [3, 2, 6, 2, 1, 3, 2], which is wrong. While using inputsparams.shape=[3,2,6,4,2], indices.shape=[3,2,1,3], axis=3, batch_dims=2 gets a resulting shape of [3, 2, 6, 1, 3, 2]. The later is right. It seems that simplily minusing 1 when axis and batch_dims are negative values is not always right. Do you have any suggestions?

Maybe we can use the shape of params or indices to solve this problem. Can we obtain the shape of any of them when operating Gather or GatherV2?

[NingLi670]

In tf.Gather, if batch_dims < 0, it will be converted to batch_dims + rank(indices). So the batch_dims should not minus 1 when the input is complex tensor. The code is updated.

[NingLi670]

@rkazants Can you build the code?

[rkazants]

build_jenkins

[NingLi670]

@rkazants Can you build the code?

[github-actions]

This PR will be closed in a week because of 2 weeks of no activity.

[NingLi670]

@rkazants @mlukasze Can you build the code?

[mlukasze]

build_jenkins

[github-actions]

This PR will be closed in a week because of 2 weeks of no activity.

[github-actions]

This PR was closed because it has been stalled for 2 week with no activity.