nan loss.
Opened this issue · 1 comments
feifeiwei commented
[epoch 1][iter 0/5095] loss: 357.1830, lr: 1.00e-04
fg/bg=(13/1011), time cost: 1.175037
rpn_cls: 237.8793, rpn_box: 22.8905, rcnn_cls: 87.1240, rcnn_box 9.2892
[epoch 1][iter 20/5095] loss: nan, lr: 1.00e-04
fg/bg=(58/966), time cost: 21.597601
rpn_cls: 0.6926, rpn_box: 0.0870, rcnn_cls: nan, rcnn_box nan
i want rto train a few-shot detector with DOTA dataset, but nan loss occured during the second iter,
Tung-I commented
Seems like a hyperparameter problem. Try to follow the paper to set appropriate hyperparameters.