search

albanie / mcnSSD

A matconvnet implementation of the Single Shot Detector
MIT License
36 stars 23 forks source link

the class_loss loc_loss mbox_loss are 'NaN' #5

Closed daibu0 closed 7 years ago

daibu0 commented 7 years ago

hello, when I trained the ssd by your code, in the iteration 1, the class_loss loc_loss mbox_loss are 21.07 2.914 23.989, but after the iteration 1 , all of the loss turned to 'NaN', Do you know how to solve it? thank you!

albanie commented 7 years ago

Hi @daibu0, it sounds as though the learning rate is too high. Normally the first epoch uses a lower "warmup" learning rate (it is set here). You could try reducing the learning rate that follows.

daibu0 commented 7 years ago

@albanie thanks for your reply, i have solved this problem. i use your default code to train the ssd and i use cpu. so the derOutputs of the mbox_loss is inf, and now i have modified it.

albanie commented 7 years ago

Great, glad to hear it!