I have a model. After training, the result of running out of caffe is correct. But the results of running out of ncs are all nan. It's weird. I'd like to ask who has encountered it. What's a good suggestion? my test is herehttps://drive.google.com/drive/folders/1FhJhBO-bIHvvm5Kqe5acAyBhdWcutYOd, Thank you.
@bai I'm using NCSDK 2.04.00.06 and everything runs fine until the last layer. If possible maybe you could run the last layer on the CPU? I was not able to experience the nan issue that you reported. Which version of the NCSDK are you using?
Thank you, I have solved this problem, NCS does not support in-place prelu, modify the network definition can solve the problem, refer to link[https://ncsforum.movidius.com/discussion/comment/2117/#Comment_2117]