search

yunlongyu / EPGN

26 stars 6 forks source link

Does anyone have the problem of "nan" loss and "0.0" accurracy? #5

Open Lothakim opened 3 years ago

Lothakim commented 3 years ago

[epoch 1/15, episode 50/100] => loss:nan [epoch 1/15, episode 100/100] => loss:nan ('Accuracy: unseen class', 0.05) ('accuracy: unseen class:', 0.05, '| seen class:', 0.0, '| harmonic:', 0.0)

[epoch 2/15, episode 50/100] => loss:nan [epoch 2/15, episode 100/100] => loss:nan ('Accuracy: unseen class', 0.05) ('accuracy: unseen class:', 0.05, '| seen class:', 0.0, '| harmonic:', 0.0)

I am running the script using the FLO dataset, and the output looks like above. And tensorflow v1's stupid placeholder makes it totally impossible to debug, I couldn't even locate which part of the code goes wrong.

WilliamYi96 commented 3 years ago

I have the same problem as you. @Lothakim Have you resolved this problem?

WilliamYi96 commented 3 years ago

@yunlongyu Any feedback on this?