[Good First Issue][TF FE] Support Conj 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 Conj operation by OpenVINO, TF FE needs to be extended with this operation support.

What needs to be done?

For Conj 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

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

Contact points

@openvinotoolkit/openvino-tf-frontend-maintainers

[jvr0123]

.take

[github-actions]

Thank you for looking into this issue! Please let us know if you have any questions or require any help.

[p-wysocki]

Hello @rupeshs! Do you have any questions or require any help?

Just yesterday our CONTRIBUTING.md has been updated with a technical guide - I highly recommend checking it out. :)

[rkazants]

Hi @jvr0123,

do you have any update on this task?

Best regards,
Roman

[jvr0123]

Hi @rkazants,

Will have a pull request by the weekend. Was having some issues with the testing framework that I think I have mostly figured out.

[rkazants]

Hi @jvr0123, please use this PR #21586 as a reference. ConjugateTranspose is quite similar to Conj.

[rkazants]

Hi @jvr0123, is there any update on this task? Please use PR #21586 it should simplify your task;)

Best regards,
Roman

[p-wysocki]

I am happy to announce that we have created a channel dedicated to Good First Issues support on our Intel DevHub Discord server! Join it to receive support, engage in discussions, ask questions and talk to OpenVINO developers.

[jvr0123]

Hello, extreme apologies for the wait. I have just created a draft pull request. Thanks for linking the other PR, was extremely helpful!