Detr: ablation study of intermediate supervision

[gaopengcuhk]

Can you share the influence of intermediate supervision on the final performance for DETR?

[fmassa]

Hi,

The numbers I have are with an older version of the code, so I'm re-running this ablation with latest implementation and will update the issue once I get the results.

[fmassa]

Hi,

I've re-run the ablations using --no_aux_loss to compare against the baseline in the current state of the codebase.

Without using auxiliary classifiers, the mAP after 300 epochs is only 35.5, compared to 40.8 if you use auxiliary classifiers.

• Orange: baseline
• Blue: --no_aux_loss
  
  which, if we compare to Fig 4 in the paper, shows that without auxiliary losses, adding more than 1 decoder layer seems to bring marginal gains.

I believe I've answered your question, and as such I'm closing this issue, but let us know if you have further questions.

[PkuRainBow]

@fmassa Great explanation! I am wondering about the loss weights on the supervision over intermediate predictions. It would be great if could share with me how does the choice of the loss weights influence the final performance.

Besides, I found that the initial learning rate of DETR is much smaller compared to Faster-RCNN or FPN. It would be great if you could also share more results on the choices of the initial learning rate.