How can i reproduce the performance in the paper?

[Anikily]

Thanks for your work and code. I trained the coot on the anet dataset by the default setting, and obtain the performance as follows: INFO vid | 0.574 | 0.839 | 0.911 | 0.980 | 1.0 | 7.6 | 2.393 INFO par | 0.563 | 0.837 | 0.906 | 0.980 | 1.0 | 7.8 | 2.380 INFO In total, training 55 epochs took 11289.156s (7587.835s train / 3701.320s val) and the recall performance in the paper is INFO vid | 0.607 | 0.879 | 0.934 | 0.987 | 1.0 | 5.1 | 2.472 INFO par | 0.613 | 0.867 | 0.929 | 0.987 | 1.0 | 5.1 | 2.467

How can I obtain this sota performance?

[simon-ging]

Hi, please first try to run the same experiment multiple times with different random seeds, maybe you just got really unlucky with the weight init. If it still doesn't work well, run the pretrained model that we provide and check it's performance to see if it matches the reported performance. Best