Feng Li

Hey, box prediction is supposed to base on the images before batch padding. Therefore, we only need to modify boxes in augmentation. Thank you.

Hey, you can refer to class Postprocess() to see how to do post process. For example, we post process with `img_h, img_w = target_sizes.unbind(1)`, in which `img_h, img_w` is the...

No, in Figure 3b it does not contain the denoising part. Yes, class label embedding is noised label, and we also input the noised box in the position query.

Hi, you can refer to our model zoo and check the config.json files. In addition, our provided command in ReadMe covers the parameters you need to set, and you can...

Hi, I think I use the same running command as you do. I am not sure why you get this result, can you give more information, like the training log?...

Hey, I cannot find the problem from these logs. I suggest you clone our clean code again and rerun it to see if you encounter the same problem. As no...

It seems you set `args.use_dn` to false so that `dn_args` does not go into the `forward` function. You need to set `args.use_dn` to true by adding --use_dn.

Hey, this means your modification to the code fails to make the model work and produces illegal box predictions. For example, here the mistake means the value of the bottom...

Hey, because deformable attention does not support half precision. You can refer to [deformable detr](https://github.com/fundamentalvision/Deformable-DETR) for more details.

Sorry I do not understand your question. We have released the inference code, for example ``` # for dn_deformable_detr: 49.5 AP python main.py -m dn_deformable_detr \ --output_dir logs/dab_deformable_detr/R50 \ --batch_size...