search

yijingru / BBAVectors-Oriented-Object-Detection

[WACV2021] Oriented Object Detection in Aerial Images with Box Boundary-Aware Vectors
MIT License
462 stars 87 forks source link

hm loss is nan;wh loss and off loss is inf #83

Open zrx0311 opened 3 years ago

zrx0311 commented 3 years ago

2*2080ti, batchsize=8; I've used the BN layer, but after a few epochs loss still becomes NaN

yijingru commented 3 years ago

Maybe the batch size is too small.

zrx0311 commented 3 years ago

thank！Now i have 4 Tesla.And would you get a better result if you didn't use BN?

yijingru commented 3 years ago

I haven't tried. But BN may enable more stable training or improve the generalization ability.