Multithreading Face recognition train

asked 2014-05-17 14:04:42 -0500

Hi, I'm training each available FaceRecognition model, and I would like to do it parallel. To be exact, I'm processing csv file, and then call 3 models in a row, although it looks like ideal example for multithreading.

Also, I would like to do recognition in parallel as well, but first things first...


edit retag flag offensive close merge delete


if you're re-training the whole schlepp each and every time, you're doing it wrong anyway.

berak gravatar imageberak ( 2014-05-17 15:49:45 -0500 )edit

sorry for not being too constructive before, but instead of a multithreading approach and the problems related to that(that's beyond the scope of this site) - consider training once and use a alternative/better/faster serialization than the builtin xml/yaml FileStorage.

berak gravatar imageberak ( 2014-05-17 17:22:35 -0500 )edit

Thanks a lot for info anyways, no problem. It works fine right now, I'm just curious how to speed it up somehow. First, I'm doing research on huge database, with different combination of training/test data, so I have to calculate at least 50 training sets * 3 algorithms... And looking at system usage with just one core used makes me wanna try everything! :-)

markoub gravatar imagemarkoub ( 2014-05-17 17:49:38 -0500 )edit

yea. true. leave no stone unturned.

still, look at the ideas behind the 3 different models.

anything, that requires training a global model(eigen,fisher,svm,ann), will need all available data at the same time and the same place to train.

the lbph idea is different from that, it's just a local transformation to lbp space per image, independantly of all others. so you could even have 20 independant servers with 5000 images each, and in the end just run them in parallel and use a simple nearest neighbour seach between the 20 independant results.

berak gravatar imageberak ( 2014-05-17 18:18:43 -0500 )edit