my model performed poor

[sdpscnc]

I tested my self trained model with epoch=25 step=398632.ckpt on the LFW dataset and found poor performance, with a maximum similarity score of only 0.1-0.2. However, the pre trained model adaface_ir50_ms1mv2.ckpt performed well with a maximum similarity score of 0.99. May I ask where the problem may be?
I used run_ir50_ms1mv2.sh for training:
python main.py
--data_root /home/xxx/PycharmProjects/AdaFace/data
--train_data_path faces_webface
--val_data_path faces_webface
--prefix ir50_ms1mv2_adaface
--gpus 1
--use_16bit
--arch ir_50
--batch_size 32
--num_workers 16
--epochs 26
--lr_milestones 12,20,24
--lr 0.1
--head adaface
--m 0.4
--h 0.333
--low_res_augmentation_prob 0.2
--crop_augmentation_prob 0.2
--photometric_augmentation_prob 0.2

[afm215]

Hello, on which dataset are you training on exactly? Be aware that you should set the parameter --custom_num_class as well