train-loss = Nan

[xiyufeng2]

2021-12-20 17:13:30,133 - mmdet - INFO - Epoch [12][150/4867] lr: 5.000e-05, eta: 0:29:00, time: 0.400, data_time: 0.005, memory: 4601, loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_cls: nan, acc: 11.5102, loss_bbox: nan, loss: nan, grad_norm: nan all loss is always nan.what's wrong?

[jbwang1997]

Hello, can you provide more information?

[xiyufeng2]

I didn't load pretrained weight to train the model , leading the loss to be large. I think the weight init method is not suitable