Setting floating-point precision for cout

[robinvanemden]

Hi @kraiskil ,

Thanks for making onnx2c available! The onnx-to-c conversion works like a charm 😄

As a result of cout's limited default floating-point precision, some of my test models did fail the backend runner's result comparisons at first.

I was able to resolve this by simply raising cout's floating-point precision in main.cc and onnx_backend_tests_runner.cc using

std::cout.precision(20);

[kraiskil]

Hello Robin, glad you like it.

Thanks for the heads up about the precision. Hadn't thought about the possibility that it would be the printing that limits the testing accuracy. I had to at some point loosen then testing accuracy so that tests especially bigger networks would match the reference from Tensorflow. But I thought that was just errors accumulating from calculations, but could be it was also this same printing error that got those calculations off to a bad start.

I can just add that line to main and test runner of course, but would you have a simple test showcasing this error that you could do a pull request for?

[robinvanemden]

Sure! See pull request #2 . I left out "std::cout.precision(20)" in the test runner and main code to make sure the showcase test will still fail.

The matmul test in the pull request fails (for me) when not also adding std::cout.precision(20) to onnx_backend_tests_runner.cc (I also added it to main.cc).

The data is generated by a basic Python script that I use to create simple models together with their input and onnxruntime generated output.

It is probably a good idea to double check whether a precision of 20 offers the right balance between c file size and model precision.

[kraiskil]

Thanks.
I'd say 20 is good for now. In case some future user has problems with the generated file size, then maybe they will contribute with a command line option to onnx2c to allow run-time selection :)

[kraiskil]

#2 merged. Thanks!