interesting countered with nan loss too

leeyoshinari / YOLO_v2

The implementation of YOLO v2 with TensorFlow

GNU General Public License v3.0

80 stars 38 forks source link

interesting countered with nan loss too #26

Closed zhhtu closed 4 years ago

zhhtu commented 5 years ago

It seems training loss encountered with NAN, then loss > 1e04 script stopped with "successful training". I've also trained yolo with MobileNet V1 also encountered with same NAN question. Maybe problem of feature extraction structure itself? Anyone tried successfully training with converge output?

guker commented 5 years ago

About Nan, you can set batch_size=64, and use SGD optimizer