my xnorm in 'cfp_fp' is too big

Hello, when I train vargfacenet with arcface, the problem always exists. The value of xnorm is very big and accuracy is very low in 'cfp_fp'. But in others validation datasets, it is normal. Could you help me? Thanks a lot.
My command line is below:
``` shell
python -u train.py --network vargfacenet --loss arcface --dataset retina --lr 0.01
```
The Xavier method is below:
```python
initializer = mx.init.Xavier(rnd_type='gaussian', factor_type='out', magnitude=2)
```