fix bug for multilabel training

[X00123]

I tried training a multiclass model to classify real and anime faces but failed, mainly due to the index confilcts between face keypoints and face labels , I check the code and fix this bug. hope it can help those in need.

By the way, there is still a bug for AWS Inferentia, i'm not familiar with AWS Neuron so i ignore it, this will not affect running code on Nvidia devices.