环境:win7+python3.5+FasterRCNN-tensorflow
源码:https://github.com/dBeker/Faster-RCNN-TensorFlow-Python3.5
问题:FasterRCNN训练自己的数据前40个iter正常后面突然出现rpn_loss_cls: nan rpn_loss_box: nan,log如下:
iter: 10 / 40000, total loss: 1.628789
>>> rpn_loss_cls: 0.594060
>>> rpn_loss_box: 1.034645
>>> loss_cls: 0.000085
>>> loss_box: 0.000000
speed: 0.315s / iter
iter: 20 / 40000, total loss: 4.250577
>>> rpn_loss_cls: 1.961701
>>> rpn_loss_box: 2.258930
>>> loss_cls: 0.029946
>>> loss_box: 0.000000
speed: 0.228s / iter
iter: 30 / 40000, total loss: 2.831758
>>> rpn_loss_cls: 0.598376
>>> rpn_loss_box: 0.246269
>>> loss_cls: 1.368618
>>> loss_box: 0.61