Closed jinxiqinghuan closed 3 years ago
Hi @jinxiqinghuan Can you please provide a minimal working example that demonstrates this bug?
Can you also specify how large is the difference between runs with different batch sizes?
It may be possible that due to reduction='mean'
flag values are reduced slightly differently for each batch and thus result in non-identical metric values.
Try to run PSNR with reduction='none'
and verify that individual values are the same.
Hi, @zakajd Thanks for your answer. I have found the problem, it is not caused by piq. Thank you again for your kindness.
I found that when I test the same data set, my batch-size is different, which may cause the final index to be different. For example the PSNR.