-
Notifications
You must be signed in to change notification settings - Fork 2.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
single-image-test: added accuracy workaround for computing CPU reference and adjust the NRMSE and PSNR metrics #26020
single-image-test: added accuracy workaround for computing CPU reference and adjust the NRMSE and PSNR metrics #26020
Conversation
build_jenkins |
2 similar comments
build_jenkins |
build_jenkins |
492baa9
to
e415906
Compare
build_jenkins |
@PatrikStepan @pereanub can you please have a look? Thank you! |
bc11368
to
eb73a8c
Compare
build_jenkins |
17080d7
to
7dd1277
Compare
build_jenkins |
1 similar comment
build_jenkins |
@Maxim-Doronin could you have a look please? |
@@ -1302,7 +1304,7 @@ bool testPSNR(const TensorMap& outputs, const TensorMap& references, const int d | |||
|
|||
auto result = utils::runPSNRMetric(actOutput, refOutput, dstHeight, dstWidth, scaleBorder, normalizedImage); | |||
|
|||
if (std::fabs(result - FLAGS_psnr_reference) > FLAGS_psnr_tolerance) { | |||
if (FLAGS_psnr_reference - result > FLAGS_psnr_tolerance) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do we expect a negative value here?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, it can happen that PSNR metric to provide value greater then the FLAGS_psnr_reference which means that the NPU actual result beats the reference. For this metric, the larger the result is the better accuracy is.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for the explanation!
It's not clear from the PR description how this tiny change is related to dynamic quantization |
I updated the PR description. We disable dynamic quantization for CPU plugin. |
build_jenkins |
…nce and adjust the NRMSE and PSNR metrics (#26020) ### Details: Small improvements for SIT: - Disable [dynamic quantization feature](https://github.com/openvinotoolkit/openvino/blob/master/src/inference/include/openvino/runtime/properties.hpp#L574), which is used by default by CPU pipeline to generate the reference outputs and affects accuracy for some particular models - Increase number of decimals for reported NRMSE metric to improve - For PSNR metric report fails only when the accuracy decrease not when it increase above the target ### Tickets: - *[CVS-143420](https://jira.devtools.intel.com/browse/CVS-143420)*
4bc5fae
to
03f6dfc
Compare
build_jenkins |
build_jenkins |
Details:
Small improvements for SIT:
Tickets: