Results of experimental reproduction are lower than reported

[Qianshaowei]

Hi, thank you for your excellent work! During the experiment, I can't reproduce the results reported in the paper using the provided code. The table shows the experimental results. The datasets and experimental configurations used are consistent with those in the provided code. We used the 2xV100 32G to train the model. The script used is as follows：

python main.py \
    --data_root /mnt/home/kimminc2/data \
    --train_data_path faces_emore \
    --val_data_path ms1m_val/faces_emore \
    --prefix ir101_ms1mv2_adaface \
    --gpus 2 \
    --use_16bit \
    --arch ir_101 \
    --batch_size 512 \
    --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

results[0]: Reported in the paper results[1]: Directly test the provided model in your repositories results[2]: Experimental results reproduced with the same configuration provided in code, the test model used is the best result (avg-acc) on the high-quality data set selected in the vaildating phase.

Dataset	Avg on high quality	IJBB	IJBC	tinyface(Rank-1)	tinyface(Rank-5)
Reported in the paper [0]	97.19	95.67	96.89	68.21	71.54
Pre-trained model [1]	97.19	95.67	96.89	68.21	71.54
Reproducted [2]	97.18	95.39	96.65	67.78	71.22

As you can see in the table, our results are lower. How can I get the similar experimental results as you reported in the paper? Is the way we choose the final checkpoint same with you actually do? Should I save and test all the models during training? I would be grateful if you could reply.