Open vinbhaskara opened 5 years ago
We also observed this on our Titan-V GPU but not on Titan-Xp. Some users pointed out setting pin_memory=False
can partially reduce this problem (https://github.com/xingyizhou/CenterNet/issues/247). I am still not sure about the root cause.
Thank you!
When testing the Pascal models, for some strange reason, using --flip_test flag speeds up the evaluation a lot. But when testing without flip_test, the inference is slower. It seems counterintuitive to me. Do you know what's the difference here other than the 2 forward passes per image and then averaging the predictions? Thanks!