[Good First Issue]: Support ConjugateTranspose operation for TensorFlow models

[rkazants]

Context

OpenVINO component responsible for support of TensorFlow models is called as TensorFlow Frontend (TF FE). TF FE converts a model represented in TensorFlow opset to a model in OpenVINO opset.

In order to infer TensorFlow models with ConjugateTranspose operation by OpenVINO, TF FE needs to be extended with this operation support.

What needs to be done?

For ConjugateTranspose operation support, you need to add the corresponding loader into TF FE op directory and to register it into the dictionary of Loaders. One loader is responsible for conversion (or decomposition) of one type of TensorFlow operation.

Here is an example of loader implementation for TensorFlow Einsum operation:

OutputVector translate_einsum_op(const NodeContext& node) { 
     auto op_type = node.get_op_type(); 
     TENSORFLOW_OP_VALIDATION(node, op_type == "Einsum", "Internal error: incorrect usage of translate_einsum_op."); 
     auto equation = node.get_attribute<std::string>("equation"); 
  
     OutputVector inputs; 
     for (size_t input_ind = 0; input_ind < node.get_input_size(); ++input_ind) { 
         inputs.push_back(node.get_input(input_ind)); 
     } 
  
     auto einsum = make_shared<Einsum>(inputs, equation); 
     set_node_name(node.get_name(), einsum); 
     return {einsum}; 
 } 

In this example, translate_einsum_op converts TF Einsum into OV Einsum. NodeContext object passed into the loader packs all info about inputs and attributes of Einsum operation. The loader retrieves an attribute of the equation by using the NodeContext::get_attribute() method, prepares input vector, creates Einsum operation from OV opset and returns a vector of outputs.

Responsibility of a loader is to parse operation attributes, prepare inputs and express TF operation via OV operations sub-graph. Example for Einsum demonstrates the resulted sub-graph with one operation. In PR #19007 you can see operation decomposition into multiple node sub-graph.

Hint

OpenVINO does not support directly complex tensors. For supporting this, we use ComplexTypeMark that loader for Conj should propagate forward by representing complex tensor as floating-point tensor having additional dimension in the tail to contain real and imaginary parts. See examples of support complex tensors by ComplexAbs loader from #20860.

Example Pull Requests

• [TF FE] Support Complex Tensors #20860
• [TF FE][TF Hub] Support DivNoNan and EnsureShape operations #19007

Resources

• How to Build OpenVINO
• Developer documentation for TensorFlow Frontend
• What is OpenVINO?
• Contribution guide
• Blog post on contributing to OpenVINO
• User documentation

Contact points

@openvinotoolkit/openvino-tf-frontend-maintainers

[Abdulrahman-Adel]

Hi @rkazants, could you please assign this issue to me? I'd like to work on it.

[rkazants]

Hi @rkazants, could you please assign this issue to me? I'd like to work on it.

Hi @Abdulrahman-Adel, sure. The task is yours:)

Best regards,
Roman

[Abdulrahman-Adel]

Hello @rkazants, Please review my pull request. #21586

[mlukasze]

Hey @Abdulrahman-Adel
Roman is on short vacations but I asked Maxim to review you PR.
Thank you for your contribution!

[rkazants]

@Abdulrahman-Adel, congratulations. PR was merged!

[Abdulrahman-Adel]

@Abdulrahman-Adel, congratulations. PR was merged!

thank you @rkazants for your guidance and patience. and I hope to collaborate more in the future.