jianjunz opened PR #8442 from jianjunz:winml-ci
to bytecodealliance:main
:
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 since they accept the same input, and they are expected to produce the same result. Pre-processing and post-processing are added to nn_image_classification_onnx for improving its accuracy.
This change also make wasi-nn WinML backend as a default feature as it's covered by test now.
Fixes #8391.
jianjunz requested fitzgen for a review on PR #8442.
jianjunz requested wasmtime-core-reviewers for a review on PR #8442.
jianjunz requested wasmtime-default-reviewers for a review on PR #8442.
jianjunz updated PR #8442.
fitzgen requested abrown for a review on PR #8442.
devigned submitted PR review:
Before a detailed review, we need to determine if adding these dependencies are ok.
devigned submitted PR review:
Before a detailed review, we need to determine if adding these dependencies are ok.
devigned created PR review comment:
We had avoided adding image and ndarray dependencies in favor of using the raw image used in the openvino test. What you did with the dog image is pretty close to what I had originally. @abrown, WDYT?
jianjunz submitted PR review.
jianjunz created PR review comment:
Image processing and post-processing are copied from your classification-component-onnx example. If new dependencies are not allowed, we can do preprocessing offline, just put a RGB file here.
devigned submitted PR review.
devigned created PR review comment:
Personally, I like showing the image prep / processing as it provides guidance to folks learning how to get started. In fact, it'd probably be helpful to add more comments / documentation to samples to explain what is happening. I think we, as practitioners, may be too accustomed to the transformations we regularly do to realize how odd they may appear to the naive observer.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz created PR review comment:
My original idea was to use an image as input for all backends, so we can cross check the results for correctness.
To avoid introducing extra dependencies, the image is replaced with processed tensor data (000000062808.rgb). I believe this is the same image for openvino backend (tensorization for openvino).
jianjunz submitted PR review.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz updated PR #8442.
jianjunz submitted PR review.
jianjunz created PR review comment:
Removing this line because it's likely to be a workaround for this specific openvino model only. If mobilenet-v1-0.25-128 is the model for openvino test, it may have an additional class 0 for background. The shape (1, 1001) also shows it has one more value than ONNX model (1, 1000).
jianjunz submitted PR review.
devigned submitted PR review:
LGTM! Tests results are the same across the impls.
Thank you for documenting how the raw image was created. Great work!
@abrown tag, you're it
abrown submitted PR review:
LGTM!
abrown merged PR #8442.
Last updated: Nov 22 2024 at 17:03 UTC