TrainCascade: What happened to -sym and -nonsym? [closed]

asked 2014-08-25 16:51:16 -0500

Troup gravatar image


I see -sym and -nonsym arguments in tutorials for the haartraining utility, but I don't see it mentioned in the new traincascade. I checked the documentation and didn't see it deprecated.

What should we know? Does traincascade now assume symmetry in positive samples, as it was the default in the old utility? Or does it assume non-symmetry by working in a different manner?


edit retag flag offensive reopen merge delete

Closed for the following reason the question is answered, right answer was accepted by StevenPuttemans
close date 2014-08-27 03:18:43.473288


It is simple, the new models don't support symmetry, they just assume that samples are non symetric since the symetric measurement was kind of giving bad results. It is dumb to use the symmetry to train only a half model and copy it, because even small changes will destroy performance then.

StevenPuttemans gravatar imageStevenPuttemans ( 2014-08-26 04:15:40 -0500 )edit

Thank you, Steven.

Matter closed. (uh, when I get enough reputation, I guess)

Troup gravatar imageTroup ( 2014-08-26 21:24:11 -0500 )edit