There are several ways to contribute to tensor-js:

• You can propose new features. For this open an issue
• You can request new onnx operators. Like above you can open an issue
• You can help implementing new features. For this, see the paragraphs below
• You can implement new examples. While there are some in the examples folder, there could definitely be more!

Implementing new features

Before starting with a feature, please check the github issues and see if a corresponding issue exists.

For a description of the development setup, check Development.md.

You should generally implement new features by forking the repository and opening a Pull request against the develop branch. Pull requests will only be merged if all tests pass and no bad linting errors are added. If you are using VSCode, files should be automatically formatted on save, which will prevent many linting errors.

Implementing a new operator

If you want to implement a new operator, the process is typically the following:

• Add the function declaration to the Tensor class in lib/types.ts
• Implement the operator on some (or in the best case all) backends:
  • CPU backend:
    • Its probably wise to start with this backend, since no special knowledge is required for this.
    • Implement the operator in a single file in lib/ops/cpu
    • Bind the implementation in lib/tensor/cpu/tensor.ts. This should ideally be only very few lines of code.
  • WASM backend:
    • Implement the operation in src/tensor.rs.
    • Then build the wasm backend with npm run build:wasm
    • Bind the implementation in lib/tensor/wasm/tensor.ts. This should ideally be only very few lines of code.
  • GPU backend:
    • Implement an operation in lib/ops/gpu. Check lib/ops/gpu/operation.ts and other GPU implementations to understand how this works
    • Bind the implementation in lib/tensor/gpu/tensor.ts.
      • For this, create a new operation dispatcher at the bottom (where all the dispatchers live)
      • Call the calc method in the GPUTensor implementation
• If a matching onnx operator exists, you can add a layer definition in lib/onnx/nodes and add it to lib/onnx/resolve.ts
  • In this case, enable the correct test case in test/data/enabledTests.ts
• If you want to, you can implement the respective backward operation in lib/autograd/variable.ts for autograd support.