ceccocats/tkDNN

not matching Darknet

Opened this issue · 1 comments

Hi,

Recently I updated all the libraries(Darknet and tkDNN) and noticed that there are incorrect object recognitions in tkDNN that were not there on prior tkDNN version, and also, Darknet doesn't detect them incorrectly. I expect precision differences because of FP16, but object detection should be identical?

I've no clue where to begin to look to learn more. Any advice?

can you share more details on your issue like a screenshot of the wrong predictions ,with what models are you experiencing this issue and your setup (windows or linux) ?