Training loss is nan when using ssd to train deep learning model

[Ajimu]

hello,  I am trying to train a ssd model following the guide of this lesson "Use deep learning to assess palm tree health"(https://learn.arcgis.com/en/projects/use-deep-learning-to-assess-palm-tree-health/) .  I tried many times with different tile size like 448/128/256/224, different argis pro version 2.6/2.8, on different computers, but met with the sample problem. Every Epoch finished very quickly and the train loss remained 'nan'. Here is my samples and parameters. I have no idea what cause the problem. Please help me, thanks a lot.

 

 

Dear @Anonymous User,  I found you have answered a question similar with mine, I wish you could see this.