InstanceNormalization op works with cpu backend, but fails with wasm backend.

[gnsmrky]

Hi @hariharans29, I just tried out the newly InstanceNormalization op, as in #18 . With cpu backend, it works good.

But when using wasm backend, it gives the following RuntimeError error when running with Chrome browser (72.0.3626.119) on Windows 10:

Uncaught (in promise) RuntimeError: memory access out of bounds
    at wasm-function[49]:376
    at t._instance_normalization_f32 (https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js:21:64093)
    at e.t.func (https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js:21:48678)
    at e.t.ccall (https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js:21:48067)
    at e.run (https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js:21:81058)
    at t.<anonymous> (https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js:21:223361)
    at https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js:21:220510
    at Object.next (https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js:21:220615)
    at https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js:21:219528
    at new Promise (<anonymous>)
wasm-function[49]	@	wasm-0000006e-49:205
t._instance_normalization_f32	@	onnx-wasm.js:8
t.func	@	wasm-binding-core.ts:154
t.ccall	@	wasm-binding-core.ts:116

Here is the code:

<script src="https://cdn.jsdelivr.net/npm/onnxjs@0.1.4/dist/onnx.min.js"></script>
<script>
    const sess = new onnx.InferenceSession({backendHint: 'wasm'});

    sess.loadModel("./onnx_models/candy_128x128_wasm.onnx").then(()=>{

    const x = new Float32Array(1*3*128*128).fill(1);
    const inputT = new onnx.Tensor(x, 'float32', [1,3,128,128]);

    sess.run([inputT]).then(output=>{
        const outputT = output.values().next().value;

        //console.log(`model output tensor: ${outputT.data}.`);
        console.log(`model output tensor size: ${outputT.size}`);
        });
    });
</script>

I've setup a GitHub page to repro this issue. Both onnx-wasm.wasm and onnx-worker.js are in the same directory as insnorm_test.html as instructed.
https://gnsmrky.github.io/pytorch-fast-neural-style-onnxjs/insnorm_test.html

Please let me know what to do to help out.

[hariharans29]

Hi,

Thanks for reporting this. This is strange. Let me take a look tomorrow.

Thanks

[fs-eire]

This issue is fixed in v0.1.5.