HandPoseDetection demo error on wasm backend

System information

• I used hand-pose-detection from GitHub but added a tfjs-backend-wasm
• TensorFlow.js installed: npm
• TensorFlow.js version: tried 3.09 - 3.13

Describe the current behavior Trying to run HandPoseDetection lib on TFJS WASM backend. It works well with WebGL backend, but once I am switching to WebAssembly I get an error:

RuntimeError: memory access out of bounds
    at tfjs-backend-wasm-simd.wasm:0x8674
    at tfjs-backend-wasm-simd.wasm:0x3c25
    at tfjs-backend-wasm-simd.wasm:0x2444
    at BackendWasm.move (backend_wasm.ts:84)
    at BackendWasm.write (backend_wasm.ts:55)
    at BackendWasm.makeOutput (backend_wasm.ts:193)
    at Object.kernelFunc (binary_kernel.ts:54)
    at kernelFunc (engine.ts:644)
    at engine.ts:711
    at Engine.scopedRun (engine.ts:478)

Seems like it happens after trying to detect palm but I am not sure, to be honest.

I tried multiple TFJS versions (3.09 - 3.13), and see an error on each.

Describe the expected behavior The expected behavior is hand-pose-detection works with WASM like it works on WEBGL, CPU backends or old hand pose library (which is also works as expected on all backends)

Standalone code to reproduce the issue I used this demo, the stock source code is here just added tf-js-wasm-backend.

[rthadur]

I was able to reproduce the same but it happens as soon as i change to tfjs-wasm and I get this error

RuntimeError: memory access out of bounds

[google-ml-butler[bot]]

Are you satisfied with the resolution of your issue? Yes No