train loss is always unchanged

tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)

https://arxiv.org/abs/1904.01355

Other

3.28k stars 630 forks source link

train loss is always unchanged #149

Closed dreamhighchina closed 5 years ago

dreamhighchina commented 5 years ago

19-09-21 00:51:36,413 fcos_core.trainer INFO: eta: 1 day, 0:04:04 iter: 69740 loss: 0.8339 (0.9184) loss_centerness: 0.5668 (0.5709) loss_cls: 0.0916 (0.1203) loss_reg: 0.1802 (0.2272) time: 0.7790 (0.7858) data: 0.0104 (0.0146) lr: 0.010000 max mem: 8322

from the begining to the iter: 69740，the total loss doesnot change （keep 0.7956 -0.9161）。

tianzhi0549 commented 5 years ago

@dreamhighchina Please make sure your code for computing training targets is correct.