Converge trend of the model

[hwfan]

Hi authors,

Thanks for your open-source implementation, I read your instruction and tried to reproduce the final detection performance. However I realized the converge speed of the model is too low: it takes almost 2 days to reach 150 epoches on two nodes with 8 gpus on each node. Have you tried any way to accelerate the procedure? Will scaling up the learning rate at the start of the training be helpful?

[cjw2021]

Thanks for your interest in our work. We have tried 2x learning rate (the same setting as deformable DETR), but it didn't work. The heavy burden of predicting verb class may cause slow convergence.