README

This example project demonstrates using the wasi-nn API to perform WinML-based inference. We first build Wasmtime, a fast and secure runtime for WebAssembly, and then build a WebAssembly example, which:

• reads an input image from fixture/kitten.png,
• converts it to the correct tensor format,
• and then classifies the image using fixture/mobilenet.onnx

To run this example, perform the following steps on Windows 10 v1803 and later:

1. Build Wasmtime according to the build guide, but enable the winml feature:
   consoleCopy

   cargo build --release --features wasmtime-wasi-nn/winml
   

2. Navigate to this directory from Wasmtime's top-level directory (referred to later as `%PROJECT_DIR%).
   Copy

   set PROJECT_DIR=%CD%
   cd crates\wasi-nn\examples\classification-example-winml
   

3. Install the wasm32-wasip1 Rust target:
   Copy

   rustup target add wasm32-wasip1
   

4. Compile this example; the wasm32-wasip1 output is a WebAssembly file:
   Copy

   cargo build --release --target=wasm32-wasip1
   

5. Run the sample; the fixture directory containing the model and image must be mapped in to be accessible to WebAssembly.
   Copy

   %PROJECT_DIR%\target\release\wasmtime.exe --dir fixture::fixture -S nn target\wasm32-wasip1\release\wasi-nn-example-winml.wasm
   

6. The example will print the top 5 classification results. To run with a different image or ONNX model, modify the files in the fixture directory along with any path changes this may cause src/main.rs.