search

ZJULearning / ttfnet

Apache License 2.0
485 stars 99 forks source link

If set wh_gaussian=False, the wh_loss is nan #7

Open Ranchentx opened 5 years ago

Ranchentx commented 5 years ago

Thanks for sharing your work. I have one question: in tab 1 of the paper, if not use wh_gaussian, the accuracy reduce from 32.9 to 32. I want to reproduce the result of 32. If set wh_gaussian=False, the wh_loss is nan. What other setting I should change.

liuzili97 commented 5 years ago

Thanks for your interest. This bug has been fixed in the latest commit, and you can reproduce the result now.

Ranchentx commented 5 years ago

@liuzili97 Thanks for your work. I set wh_gaussian=False in ttfnet_1x, it reproduces 24.8. But if I set wh_gaussian=True, it reproduces only 18.3. I try twice and I do not know why.

liuzili97 commented 5 years ago

It seems that everything works fine according to our experiments. How many GPUs have you used, and have you changed other settings?