Hi，
Due to the final formula having two elements，“runtime” and “result's fidelity”（PSNR）,
So, is the result of PSNR calculation obtained from FP16 model or FP32 model inference?
I have read the code of sr_demo.py and feel that it is derived from FP32, but I am not sure if I understand it correctly?
So I want to know which precision, FP16 or FP32, is used in the evaluation formula that leads to the final score for the two metrics of PSNR and runtime.?
Thanks :)
Posted by: loveletter @ Feb. 20, 2023, 7:14 a.m.Hello, thanks for your question.
Models will run as FP16 , you can find here https://github.com/eduardzamfir/NTIRE23-RTSR IMDN and RFDN basline times using our setup.
We will check that script in the repo to make sure everything runs FP16. Feel free to open git issues there if you find something confusing :)
Thank you very much for your answer.
I still want to make sure about one question:
In the git, you said
"We compute our metrics using 'calc_metrics.py' and the SR outputs you provide in 'results/' and report average PSNR/SSIM on RGB and Y-Channel first."
I would like to know if the images in the 'results/' we submitted were obtained from an FP32 or FP16 model？
Or if we need to provide two version images('results/') including FP32 and FP16，which is convenient to calculate the PSNRs of two versions under different precisions(FP32 /FP16)?
Looking forward to your reply, thank you very much:)
Posted by: loveletter @ Feb. 21, 2023, 2:17 a.m.Hello, we assume there are not going to be notable differences when running on MP FP16, therefore the output images should be generated by running the model on FP16 directly.
That is the way we will evaluate internally, your submission can be done with FP32 models, but consider how we will evaluate internally just in case there are differences.