[Question] Blob Tensor Format

[wuassassin]

Hello @abrown, I'm interested in contributing example codes to this repo, especially for issue #7. But I'm not familiar with the .bgr file in the fixtures folder. Is it compatible with https://github.com/bytecodealliance/wasi-nn/tree/main/image2tensor?

[abrown]

Thanks for taking a look at #7! Your contributions would be appreciated. Now, about compatibiliy, I am not completely sure but let me share what I think. The .bgr files in this repository were created with the openvino-tensor-converter tool which I developed as a helper for the test fixtures you mention. Here's an example of how it is used in one of the build.sh scripts that attempts to regenerate the fixture artifacts:

openvino-rs/crates/openvino/tests/fixtures/alexnet/build.sh

Line 39 in 7bea5f4

cargo run -p openvino-tensor-converter -- $TMP_DIR/val2017/000000062808.jpg $FIXTURE_DIR/tensor-1x3x227x227-f32.bgr 227x227x3xfp32

Images need to be decoded to perform the inference and the openvino-tensor-converter uses OpenCV as the library to do the conversions--IIRC, I ran into issues where I could not seem to generate the right raw bytes and using OpenCV fixed this (I think this is also what OpenVINO uses internally if one asks it to do the conversions). I found it more clear to create separate .bgr files for these decoded bytes but I could have just as well decoded the bytes in memory using the lib.rs functions.

The image2tensor library does essentially the same thing: it decodes images into the right bytes for a tensor. The differences are that it uses the image library instead of OpenCV and it only handles f32 precision at the moment, I believe. And the purpose of the tools is slightly different: openvino-tensor-converter is made for this repository's tests (a small scope!) whereas image2tensor is intended for anyone who might need to convert images to tensors in WebAssembly-compiled code. @brianjjones, any other comments on differences or similarities?

Bottom line, in my opinion, is that you're better off taking a crack at this with openvino-tensor-converter at the moment. It's probably the path of least resistance for contributing to this repository.

[wuassassin]

Bottom line, in my opinion, is that you're better off taking a crack at this with openvino-tensor-converter at the moment. It's probably the path of least resistance for contributing to this repository.

Cool will do 👍🏼

Anyway, can we close this issue after I tried image2tensor? (or perhaps after @brianjjones clarify the inquiry)

[abrown]

It's been a while since this has opened and the situation has not changed much. @wuassassin, I'll close it for now but feel free to reopen it if there are further questions along these lines.