Named Model Example

This example is mostly the same as the classification-example but uses wasi-nn's "named model" API to load the ML model before the WebAssembly program executes. Instead of loading the model bytes during program execution (which may be prohibitively slow for large models), this wasi-nn extension allows the model to be loaded by the engine prior to execution.

Pre-requisites

The example pre-requisites are mostly the same as that of the classification-example, except that we separate the model files that the engine pre-loads (mobilenet/*) from the image file read via WASI from the host (fixture/*):

wget https://download.01.org/openvinotoolkit/fixtures/mobilenet/mobilenet.bin -O mobilenet/model.bin
wget https://download.01.org/openvinotoolkit/fixtures/mobilenet/mobilenet.xml -O mobilenet/model.xml
wget https://download.01.org/openvinotoolkit/fixtures/mobilenet/tensor-1x224x224x3-f32.bgr -O fixture/tensor.bgr

As before, this Rust example compiles to a WebAssembly program using the wasm32-wasip1 target: cargo build --target=wasm32-wasip1.

Run

The program is invoked with slightly different flags than the classification-example:

<path>/<to>/wasmtime run --dir=fixture --wasi=nn --wasi=nn-graph=openvino::mobilenet target/wasm32-wasip1/debug/wasi-nn-example-named.wasm

The primary difference is the addition of --wasi=nn-graph=openvino::mobilenet: this informs Wasmtime's wasi-nn implementation to pre-load a named model from the mobilenet directory using the openvino encoding. Note that, in this implementation, the model name (i.e., mobilenet) is derived from the base name of the path passed in --nn-graph=<encoding>::<path>.