WASI-NN milestones

[tonibofarull]

Motivation

We'd like to implement WASI-NN into WebAssembly runtime so that all WASM applications that we develop have a standard API to manage neural networks.

Milestones

MVP milestone

Neural networks can be exported into multiple formats, have multiple tensor types, and run in multiple types of targets. We define the Minimum Viable Product (MVP) as a solution that implements all the functionalities for a particular format, tensor type, and target.

For instance, the standard specifies the following formats,

enum graph-encoding {
    openvino,
    onnx,
    tensorflow,
    pytorch
}

targets,

enum execution-target {
    cpu,
    gpu,
    tpu
}

and tensor types,

enum tensor-type {
    fp16,
    fp32,
    up8,
    ip32
}

Check here for more information.

For the MVP, we have selected:

• TensorFlow
• CPU
• FP32

The criteria are to select the most usage of each of the options. In the case of TensorFlow, the developer survey of Stack Overflow.

GPU capability

Add compatibility of execution-target = gpu for Tensorflow.

Support of all tensor types

Support all tensor types for Tensorflow.

What’s Next

• Add support for other frameworks and devices.
• Frequently check WASI-NN to keep the interface and new features up to date.