Thanks for your excellent work!
We use the last command in the sample_training.sh (denote as IM320_Ensemble_L2-3_P001_D1024-384_PS-5_AN-5_S88) to train the model. To get the performance of pixel AUPRO, we use the function 'compute_pro' in RD4AD to calculate the PRO. The performance is 92.9 lower than the released results (94.9).
What is the reason for this?
Would you like to share the code about calculating pixel AUPRO and the pretrained model? Thanks a lot!
Thanks for your excellent work! We use the last command in the sample_training.sh (denote as IM320_Ensemble_L2-3_P001_D1024-384_PS-5_AN-5_S88) to train the model. To get the performance of pixel AUPRO, we use the function 'compute_pro' in RD4AD to calculate the PRO. The performance is 92.9 lower than the released results (94.9). What is the reason for this? Would you like to share the code about calculating pixel AUPRO and the pretrained model? Thanks a lot!