The reason why float 16 yields nan is because the number of classes exceeds the range of float16. Thus, some of the labels are inf and cause the network to diverge. Some of the predictions are also in…
hello,I run your code, find that the highest accuracy in lfw is about 98.9%. But the insightface code can easily get 99.5% accuracy in lfw. Is there anying wrong in the training code? or the evaluati…