search

amitshaked / resmatch

Implementation of "Improved Stereo Matching with Constant Highway Networks and Reflective Confidence Learning"
BSD 2-Clause "Simplified" License
75 stars 27 forks source link

while training on mb dataset, err will be nan #1

Closed Sarah20187 closed 7 years ago

Sarah20187 commented 7 years ago

After preparing data, I use th main.lua -ds mb -a train_mcn -mc mc-cnn -m slow -gpu 2 to train a matching cost net and error occurred.

===> Loading matching cost network...
5
===> Loaded! Network: mb_mc-cnn_slow_LL_rgb
===> training matching cost network
{[........................................ 2/402508 ....................................]  ETA: 1D8h | Step: 289ms
  1 : nan
}
2
WARNING! err=nan

Could you help me?

amitshaked commented 7 years ago

Thanks for bringing up the issue. It's working now. Please make sure that after you compile the MiddEval3-SDK for Middlebury dataset, you end up with the computemask binary in one of the directories listed in your PATH enviromential variable.

Sarah20187 commented 7 years ago

hi, Thank you for your reply. after I applied your modified code, the problem remains randomly. I'm sure that the mask file is correct as it works while I train mb slow net on mc-cnn source code? Have you figure out what leads the nan error?

2017-01-20 20:36 GMT+01:00 amitshaked notifications@github.com:

Thanks for bringing up the issue. It's working now. Please make sure that after you compile the MiddEval3-SDK for Middlebury dataset, you end up with the computemask binary in one of the directories listed in your PATH enviromential variable.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/amitshaked/resmatch/issues/1#issuecomment-274160208, or mute the thread https://github.com/notifications/unsubscribe-auth/APZ2132jK6tZ-WZfQ4_gseadAALRirilks5rUQy8gaJpZM4LnLgZ .