I’m looking to invoke tflite models in a Go environment without using cgo. The environment is off the grid so it’s not possible for me to send requests to some API to do run the model. Are there any good solutions to this problem today?
I’ve looked at writing my own wasm runtime bindings using Wazero but it looks like the TensorFlow wasm module may not be inline with the wasi spec: TensorFlow Lite Bindings: import func[wasi_snapshot_preview1.fd_seek]: signature mismatch · Issue #1461 · tetratelabs/wazero · GitHub. Edit, it also looks like this is just the TensorFlow backend, which is only half the problem so this doesn’t really solve my issue either.
I’ve looked at using ONNX but it doesn’t support some operators used by embedding layers like the
Is my use-case (running inference natively in Go) actually that obscure?