The performance of DET_t obtained by reproducing the model is lower

[LePetitSirius]

Thanks for your excellent work, after reproducing your code using 8*P40 the result is
OpenLane-V2 Score: 0.3952, DET_l: 0.2843010127544403, DET_t: 0.4807989299297333, TOP_ll: 0.1098, TOP_lt: 0.2345
There is a big difference between DET_t and the result you provided DET_t: 48.6. May I ask what causes this problem?
Looking forward to your reply! thank you

[sephyli]

Thank you for your feedback! The discrepancy you're observing, especially in the DET_t metric, is likely due to the inherent instability in the convergence of this type of topology model. The standard deviation of the metrics can be larger compared to simpler detection tasks. It's possible that a simple rerun of the training could lead to improved results.