[TF FE] Support MatrixDiagV3 operation

[duydl]

Details:

• WIP, will update and request review when finished.
• Parse logic from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/linalg/matrix_diag_op.cc to openVINO operations
• Initial examples from https://www.tensorflow.org/api_docs/python/tf/raw_ops/MatrixDiagV3#for_example

Tickets:

• [Good First Issue][TF FE]: Support MatrixDiagV3 operation for TensorFlow #23248

[rkazants]

Good start, keep this way!

[rkazants]

build_jenkins

[rkazants]

Hi @duydl, any progress on the next steps? Please finalize PR.

[duydl]

@rkazants Hi, I have some problem with retrieving attribute from tf graph: attribute always get the fallback value.

    auto k = node.get_attribute<int64_t>("k", 0); // get an int instead of dynamic size node
    auto num_rows = node.get_attribute<int64_t>("num_rows", -1); 
    auto num_cols = node.get_attribute<int64_t>("num_cols", -1); 
    auto padding_value_ = node.get_attribute<int64_t>("padding_value", 0);

I check other ops using this but couldn't see the difference in how attribute passed into tf graph in the pytest file.

[rkazants]

@rkazants Hi, I have some problem with retrieving attribute from tf graph: attribute always get the fallback value.

I think these are not attributes and should be treat as input tensors and accessed with node.get_input(i). You just need to check get_input_size that will tell how many inputs and what are inputs are not passed and we need to use default value for this.

[duydl]

@rkazants Hi, I have some problem with retrieving attribute from tf graph: attribute always get the fallback value.

I think these are not attributes and should be treat as input tensors and accessed with node.get_input(i). You just need to check get_input_size that will tell how many inputs and what are inputs are not passed and we need to use default value for this.

So params could only be either attribute or input and it depend on tf internal? I originally use get_input to get k but get this error:

E       FrontEnd API failed with OpConversionFailure:
E       [TensorFlow Frontend] Internal error, conversion is failed for MatrixDiagV3 operation with a message:
E       Exception from src/core/src/partial_shape.cpp:268:
E       to_shape was called on a dynamic shape.

from this snippet.

    auto axes = make_shared<v0::Constant>(element::i64, Shape{1}, std::vector<int64_t>{0});
    auto k_unsqueezed = make_shared<v0::Unsqueeze>(k_, axes);
    auto updated_shape = make_shared<v3::ScatterElementsUpdate>(
      zero_padded_diag_shape, // data input
      make_shared<v0::Constant>(element::i64, Shape{1}, std::vector<int64_t>{-1}), // indices
      k_unsqueezed, // updates 
      make_shared<v0::Constant>(element::i64, Shape{}, std::vector<int64_t>{0}) // axis 
    );

I expected it is because k is from node input so change to attribute instead. But as it is not possible, do you have any insight?

Edit: ie, openVINO ops nodes need to be config with value of k (and some other params of the tf ops) but seem like it is not possible to access the value as it is considered dynamic.