[Good First Issue][TF FE]: Support Bincount operation for TensorFlow Hub 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 Hub models with Bincount operation by OpenVINO, TF FE needs to be extended with this operation support.

We have a couple of models from TF Hub containing this operation:

• mil-nce/s3d
• mil-nce/i3d

What needs to be done?

For Bincount operation support, you need to implement 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.

Once you are done with implementation of the translator, you need to implement the corresponding layer tests test_tf_Bincount.py and put it into layer_tests/tensorflow_tests directory. Example how to run some layer test:

export TEST_DEVICE=CPU
cd openvino/tests/layer_tests/tensorflow_tests
pytest test_tf_Shape.py -m use_new_frontend

Hint

Bincount operation requires decomposition using OV opset. The following idea is proposed to use for the decomposition but you may implement your more elegant and performance efficient idea:

1. Generate range tensor [0, size) using Range operation
2. Reshape tensors of arr and weights to 1D tensors
3. Unsqueeze range to [size, 1] shape and unsqueeze arr and weights to shapes [1, num]
4. Generate a mask of equalness range == arr where to set 0 or weight from weights tensor
5. After we use operations Select, ReduceSum, and ScatterUpdate for result generation

Example Pull Requests

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

Resources

• What is OpenVINO?
• How to Build OpenVINO
• Developer documentation for TensorFlow Frontend
• Contribution guide - start here!
• Intel DevHub Discord channel - engage in discussions, ask questions and talk to OpenVINO developers

Contact points

@openvinotoolkit/openvino-tf-frontend-maintainers

[TheNobody-12]

Hello @rkazants , I would like to contribute to this program. Can you please assign me this task. ?

[rkazants]

Fyi, please use new link to join Intel DevHub Discord server, if you have not yet joined. The previous one is outdated.

Best regards,
Roman

[rkazants]

Hi @TheNobody-12, do you have any update on this task? Please let me know in case questions and you can create a preliminary PR to discuss.

Best regards,
Roman

[rkazants]

Hi @TheNobody-12, any update on this task?

[TheNobody-12]

If possible i require some guidance, how i should start working on this issue?

[ViswanathBalla22]

hi @rkazants I'm a beginner excited to contribute. I'm looking to solve issue can you assign to me. and can you help me if I stuck with any problem

[rkazants]

hi @rkazants I'm a beginner excited to contribute. I'm looking to solve issue can you assign to me. and can you help me if I stuck with any problem

Hi @ViswanathBalla22, sorry but task has been already assigned. Please track new (available) good first issues in OpenVINO. We try to post new GFIs almost every day.

Best regards,
Roman

[rkazants]

If possible i require some guidance, how i should start working on this issue?

Hi @TheNobody-12, try to start with completing the first two steps of the hint in the ticket description and post PR with these changes so I will check and give the next guidance. You can see PR with example for another operation: #19007

Best regards,
Roman

[rkazants]

Hi @ViswanathBalla22, here is the task that just gets available: #22082.
You you wish to contribute there, just type a comment .take in that ticket.

[ViswanathBalla22]

ok @rkazants ,I can assign that issue. and btw can you provide any community channel or discord for this project

[mlukasze]

hey @ViswanathBalla22
Here is our Discord
Have fun :)

[p-wysocki]

Hello @ViswanathBalla22, are you still working on that issue?

[Aryan8912]

assign it to me so I can work on this issue

[mlukasze]

as you wish @Aryan8912 :)

[rkazants]

Hi @Aryan8912, any progress on it?

[Aryan8912]

sorry sir for the late reply I am traveling, give me one or two days, and I will submit a PR

[rkazants]

Hi @Aryan8912, I return this GFI to available status since no update.

[chux0519]

Hi developers, I'd like to take this one!

[mlukasze]

be our guest, have fun, it's yours now :)

[chux0519]

Hi, developers

I'm still work in progress, following some example code and I want to check if my implementation works so far, I added a test case (test_tf_Bincount.py) in python side, run pytest test_tf_Bincount.py and I hit

Internal error, no translator found for operation(s): Bincount

(but I can run other test cases)

here is the diff(implementation is still incomplete, and I just want to make sure test case works:

https://github.com/openvinotoolkit/openvino/compare/master...chux0519:openvino:feat-bincount?expand=1

I'm using a intel mac and my compile command is

cmake \
-G "Ninja Multi-Config" \
-DENABLE_CPPLINT=OFF \
-DENABLE_NCC_STYLE=OFF \
-DENABLE_TESTS=ON \
-DCMAKE_COMPILE_WARNING_AS_ERROR=OFF \
-DENABLE_STRICT_DEPENDENCIES=OFF \
-DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-DCMAKE_C_COMPILER_LAUNCHER=ccache \
-DENABLE_SYSTEM_PUGIXML=ON -DENABLE_SYSTEM_SNAPPY=ON -DENABLE_SYSTEM_PROTOBUF=ON -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DENABLE_TESTS=ON -DCMAKE_BUILD_TYPE=Debug -DENABLE_PYTHON=ON -DPython3_EXECUTABLE=/Users/chux0519/venv/bin/python -DENABLE_WHEEL=ON ..

python version is 3.11, and I reinstalled wheels after rebuild, did I miss anything

update

it seems the wheel is not rebuilt(timestamp shows it's not updated), i'll clean and rebuild and see if it works

cool, seems so, just delete wheels and regenerate one fixed this.

[chux0519]

PR sent: #23418