NaN Loss

[chenjiachengzzz]

"Dear author, you have done a fantastic job and made it open source. When I was reproducing your work, the network suddenly collapsed when epoch > 50 (I repeated it four times and couldn't change the situation). Have you encountered this situation before?"

[laoyangui]

Hi, @chenjiachengzzz. Based on the log, you may have missed the following parameter: -- rgb_ range 1. (An example command is in the file 'demo.sh'.)

[chenjiachengzzz]

Hi, @chenjiachengzzz. Based on the log, you may have missed the following parameter: -- rgb_ range 1. (An example command is in the file 'demo.sh'.)

Thank you very much for your reply, will training in the interval 0-255 cause the model to collapse?

[laoyangui]

Hi, @chenjiachengzzz. Based on the log, you may have missed the following parameter: -- rgb_ range 1. (An example command is in the file 'demo.sh'.)

Thank you very much for your reply, will training in the interval 0-255 cause the model to collapse?

Training in the interval 0-255 may cause the model to collapse. If you insist on using this interval , you can try to further reduce the parameter --res_scale.

[chenjiachengzzz]

Hi, @chenjiachengzzz. Based on the log, you may have missed the following parameter: -- rgb_ range 1. (An example command is in the file 'demo.sh'.)

Thank you very much for your reply, will training in the interval 0-255 cause the model to collapse?

Training in the interval 0-255 may cause the model to collapse. If you insist on using this interval , you can try to further reduce the parameter --res_scale.
I got it. Thank you very much!