Enable unit test for wasi-nn WinML backend.

This test was disabled because GitHub Actions Windows Server image
doesn't have desktop experience included. But it looks like we can have
a standalone WinML binary downloaded from ONNX Runtime project.

Wasi-nn WinML backend and ONNX Runtime backend now share the same test
code as they accept the same input, and they are expected to produce the
same result.

This change also make wasi-nn WinML backend as a default feature.

prtest:full

.github/workflows/main.yml

@@ -678,6 +678,9 @@ jobs:
       matrix:
         feature: ["openvino", "onnx"]
         os: ["ubuntu-latest", "windows-latest"]
+        include:
+          - os: windows-latest
+            feature: winml
     name: Test wasi-nn (${{ matrix.feature }}, ${{ matrix.os }})
     runs-on: ${{ matrix.os }}
     needs: determine
@@ -692,6 +695,15 @@ jobs:
     - uses: abrown/install-openvino-action@v8
       if: runner.arch == 'X64'
 
+    # Install WinML for testing wasi-nn WinML backend. WinML is only available
+    # on Windows clients and Windows Server with desktop experience enabled.
+    # GitHub Actions Window Server image doesn't have desktop experience
+    # enabled, so we download the standalone library from ONNX Runtime project.
+    - uses: nuget/setup-nuget@v2
+      if: (matrix.os == 'windows-latest') && (matrix.feature == 'winml')
+    - run: nuget install Microsoft.AI.MachineLearning
+      if: (matrix.os == 'windows-latest') && (matrix.feature == 'winml')
+
     # Install Rust targets.
     - run: rustup target add wasm32-wasi
 

crates/test-programs/src/bin/nn_image_classification_onnx.rs

@@ -8,10 +8,13 @@ pub fn main() -> Result<()> {
         .expect("the model file to be mapped to the fixture directory");
     let graph =
         GraphBuilder::new(GraphEncoding::Onnx, ExecutionTarget::CPU).build_from_bytes([&model])?;
-    let tensor = fs::read("fixture/tensor.bgr")
+    let tensor = fs::read("fixture/000000062808.rgb")
         .expect("the tensor file to be mapped to the fixture directory");
     let results = classify(graph, tensor)?;
     let top_five = &sort_results(&results)[..5];
+    // 963 is meat loaf, meatloaf.
+    // https://github.com/onnx/models/blob/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/synset.txt#L963
+    assert_eq!(top_five[0].class_id(), 963);
     println!("found results, sorted top 5: {:?}", top_five);
     Ok(())
 }

crates/test-programs/src/nn.rs

@@ -42,12 +42,15 @@ pub fn classify(graph: Graph, tensor: Vec<u8>) -> Result<Vec<f32>> {
 /// placing the match probability for each class at the index for that class
 /// (the probability of class `N` is stored at `probabilities[N]`).
 pub fn sort_results(probabilities: &[f32]) -> Vec<InferenceResult> {
+    let probabilities_iter = probabilities.iter();
+
     // It is unclear why the MobileNet output indices are "off by one" but the
     // `.skip(1)` below seems necessary to get results that make sense (e.g. 763
     // = "revolver" vs 762 = "restaurant").
-    let mut results: Vec<InferenceResult> = probabilities
-        .iter()
-        .skip(1)
+    #[cfg(feature = "openvino")]
+    probabilities_iter.skip(1);
+
+    let mut results: Vec<InferenceResult> = probabilities_iter
         .enumerate()
         .map(|(c, p)| InferenceResult(c, *p))
         .collect();

crates/wasi-nn/Cargo.toml

@@ -50,7 +50,7 @@ wasi-common = { workspace = true, features = ["sync"] }
 wasmtime = { workspace = true, features = ["cranelift"] }
 
 [features]
-default = ["openvino"]
+default = ["openvino", "winml"]
 # openvino is available on all platforms, it requires openvino installed.
 openvino = ["dep:openvino"]
 # onnx is available on all platforms.

crates/wasi-nn/src/testing.rs

@@ -8,7 +8,12 @@
 
 #[allow(unused_imports)]
 use anyhow::{anyhow, Context, Result};
-use std::{env, fs, path::Path, path::PathBuf, process::Command, sync::Mutex};
+use std::{
+    env, fs,
+    path::{Path, PathBuf},
+    process::Command,
+    sync::Mutex,
+};
 
 #[cfg(all(feature = "winml", target_arch = "x86_64", target_os = "windows"))]
 use windows::AI::MachineLearning::{LearningModelDevice, LearningModelDeviceKind};
@@ -50,13 +55,12 @@ pub fn check() -> Result<()> {
     #[cfg(feature = "openvino")]
     check_openvino_artifacts_are_available()?;
 
-    #[cfg(feature = "onnx")]
+    #[cfg(any(feature = "onnx", all(feature = "winml", target_os = "windows")))]
     check_onnx_artifacts_are_available()?;
 
-    #[cfg(all(feature = "winml", target_arch = "x86_64", target_os = "windows"))]
+    #[cfg(all(feature = "winml", target_os = "windows"))]
     {
         check_winml_is_available()?;
-        check_winml_artifacts_are_available()?;
     }
     Ok(())
 }
@@ -108,7 +112,7 @@ fn check_openvino_artifacts_are_available() -> Result<()> {
     Ok(())
 }
 
-#[cfg(all(feature = "winml", target_arch = "x86_64", target_os = "windows"))]
+#[cfg(all(feature = "winml", target_os = "windows"))]
 fn check_winml_is_available() -> Result<()> {
     match std::panic::catch_unwind(|| {
         println!(
@@ -121,59 +125,34 @@ fn check_winml_is_available() -> Result<()> {
     }
 }
 
-#[cfg(feature = "onnx")]
+#[cfg(any(feature = "onnx", all(feature = "winml", target_os = "windows")))]
 fn check_onnx_artifacts_are_available() -> Result<()> {
     let _exclusively_retrieve_artifacts = ARTIFACTS.lock().unwrap();
 
-    const OPENVINO_BASE_URL: &str =
-        "https://github.com/intel/openvino-rs/raw/main/crates/openvino/tests/fixtures/mobilenet";
     const ONNX_BASE_URL: &str =
-        "https://github.com/onnx/models/raw/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/mobilenet/model/mobilenetv2-7.onnx?download=";
+        "https://github.com/onnx/models/raw/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/mobilenet/model/mobilenetv2-10.onnx?download=";
 
     let artifacts_dir = artifacts_dir();
     if !artifacts_dir.is_dir() {
         fs::create_dir(&artifacts_dir)?;
     }
 
-    for (from, to) in [
-        (
-            [OPENVINO_BASE_URL, "tensor-1x224x224x3-f32.bgr"].join("/"),
-            "tensor.bgr",
-        ),
-        (ONNX_BASE_URL.to_string(), "model.onnx"),
-    ] {
+    for (from, to) in [(ONNX_BASE_URL.to_string(), "model.onnx")] {
         let local_path = artifacts_dir.join(to);
         if !local_path.is_file() {
             download(&from, &local_path).with_context(|| "unable to retrieve test artifact")?;
         } else {
             println!("> using cached artifact: {}", local_path.display())
         }
     }
-    Ok(())
-}
 
-#[cfg(all(feature = "winml", target_arch = "x86_64", target_os = "windows"))]
-fn check_winml_artifacts_are_available() -> Result<()> {
-    let _exclusively_retrieve_artifacts = ARTIFACTS.lock().unwrap();
-    let artifacts_dir = artifacts_dir();
-    if !artifacts_dir.is_dir() {
-        fs::create_dir(&artifacts_dir)?;
-    }
-    const MODEL_URL: &str="https://github.com/onnx/models/raw/5faef4c33eba0395177850e1e31c4a6a9e634c82/vision/classification/mobilenet/model/mobilenetv2-12.onnx";
-    for (from, to) in [(MODEL_URL, "model.onnx")] {
-        let local_path = artifacts_dir.join(to);
-        if !local_path.is_file() {
-            download(&from, &local_path).with_context(|| "unable to retrieve test artifact")?;
-        } else {
-            println!("> using cached artifact: {}", local_path.display())
-        }
-    }
-    // kitten.rgb is converted from https://github.com/microsoft/Windows-Machine-Learning/blob/master/SharedContent/media/kitten_224.png?raw=true.
-    let tensor_path = env::current_dir()?
+    // Copy image from source tree to artifact directory.
+    let image_path = env::current_dir()?
         .join("tests")
         .join("fixtures")
-        .join("kitten.rgb");
-    fs::copy(tensor_path, artifacts_dir.join("kitten.rgb"))?;
+        .join("000000062808.rgb");
+    let dest_path = artifacts_dir.join("000000062808.rgb");
+    fs::copy(&image_path, &dest_path)?;
     Ok(())
 }
 

crates/wasi-nn/tests/all.rs

@@ -97,10 +97,10 @@ fn nn_image_classification_named() {
 #[cfg_attr(not(all(feature = "winml", target_os = "windows")), ignore)]
 #[test]
 fn nn_image_classification_winml() {
-    #[cfg(feature = "winml")]
+    #[cfg(all(feature = "winml", target_os = "windows"))]
     {
         let backend = Backend::from(backend::winml::WinMLBackend::default());
-        run(NN_IMAGE_CLASSIFICATION_WINML, backend, true).unwrap()
+        run(NN_IMAGE_CLASSIFICATION_ONNX, backend, true).unwrap()
     }
 }
 

crates/wasi-nn/tests/fixtures/readme.md

@@ -0,0 +1,9 @@
+The original image of 000000062808.rgb is 000000062808.jpg from ImageNet
+database. It processed by following Python code with
+https://github.com/onnx/models/blob/bec48b6a70e5e9042c0badbaafefe4454e072d08/validated/vision/classification/imagenet_preprocess.py
+
+```
+image = mxnet.image.imread('000000062808.jpg')
+image = preprocess_mxnet(image)
+image.asnumpy().tofile('000000062808.rgb')
+```