search

DeokyunKim / Progressive-Face-Super-Resolution

Official Pytorch Implementation of Progressive Face Super-Resolution (BMVC 2019 Accepted)
260 stars 60 forks source link

Test Batchsize #25

Open cocoxw opened 2 years ago

cocoxw commented 2 years ago

Thank you for sharing! Recently, I encountered a problem during the test. When the test batchsize changes, the final results will be a little different. Could you give me some advice, thanks. ^_^

cocoxw commented 2 years ago

The final result means the average PSNR generated by eval.py

DeokyunKim commented 2 years ago

Hi, cocoxw. Thank you for your interest in our work.

Your question is that the PSNR values are different along with batch size. The PSNR values are slightly different since the batch norm layer of our face SR model calculates the mean and variance of each batch. To resolve the problem, there are some options such as using another normalization layer or removing batch norm layers.

We note that the uploaded trained model is re-trained best version model through extensive experiments.

If you have more questions, please comment here.

Best regards, Deokyun Kim

cocoxw commented 2 years ago

I got it. Thank you for your reply.