CONTRIBUTING_MISSING_OP.md

How to Contribute a New op to TF.js Repository

Step 1. Add the new op to tfjs-core ops directory.

Prerequisite

1. The op should already be supported in TF api.
2. The op should not have been supported in TF.js, see op support list.

Implementation Details

1. Create a new op in tfjs-core/ops directory.

   The op file should have the following elements: 1. License information. 2. JSDoc comment. 3. Input validations, if any. 4. Delegate the execution to the right kernel through ENGINE.runKernel().

   In addition, for the kernel delegation to work properly, in kernel_names.ts file, define: 1. kernel name. 2. input type. 3. attribute type.

2. Export the op file in ops.ts in the same directory.

3. Add tests for the op in the same directory. Test file name should be the same as the op file’s name with _test suffix.

4. Exclude the test in all the backends, and add annotation “Not implemented yet.”. See below for where to exclude the test in each backend:

cpu backend

In run_tests.ts, customInclude method, add:

    // Not implemented yet.
    if (testName.includes(test_name)) {
      return false;
    }

webgl backend

In setup_test.ts, customInclude method, add:

    // Not implemented yet.
    if (testName.includes(test_name)) {
      return false;
    }

node backend

In run_tests.ts, IGNORE_LIST, add test_name to the list.

Step 2. Add a new kernel in a backend.

Implementation Details

1. Create a new kernel in the kernels directory of a backend.

   The kernel file should have the following elements:

   1. License information.
   2. The kernel implementation, and export it.
   3. Export a kernelConfig.

2. Register the kernel in register_all_kernels.ts in the corresponding backend.

3. Remove the op from test exclusion list in the corresponding backend. For wasm backend, add the test to the inclusion list.

Step 3. Add the op to Converter’s executor.

1. Add op mapping in the op list ts file, use your best judgement to assign an op category: tfjs-converter/src/operations/op_list/{corresponding_op_category}.ts. Use the TF C++ op api as reference for tfOpName, inputs, and attrs.

2. Auto generates corresponding json file by running following command in directory tfjs-converter: yarn gen-json. Check that the json object is generated in python/tensorflowjs/op_list/{corresponding_op_group}.json

3. Find the corresponding executor for the op and add the op to the switch, the executors are in tfjs-converter/src/operations/executors.

4. Add a test to the corresponding executor test file.

5. Update the supported op doc in tfjs-converter/docs/supported_ops.md.

6. Add a mapping in kernel2op.json.