D-FINE not using any auxiliary losses when denoising is turned off

[m-matthias]

System Info

• transformers version: 5.7.0.dev0
• Platform: Linux-6.6.87.2-microsoft-standard-WSL2-x86_64-with-glibc2.39
• Python version: 3.12.3
• Huggingface_hub version: 1.7.1
• Safetensors version: 0.7.0
• Accelerate version: 1.13.0
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (accelerator?): 2.8.0+cu128 (CUDA)
• GPU type: NVIDIA RTX 500 Ada Generation Laptop GPU

Who can help?

@VladOS95-cyber
@yonigozlan

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Turn off denoising, by e.g.:

model = DFineForObjectDetection.from_pretrained("ustc-community/dfine-medium-obj365", num_denoising=0)

Expected behavior

Only the denoising losses should be turned off. However, DfineForObjectDetectionLoss only checks for denoising_meta_values to either use all or no auxiliary losses at all. See here: https://github.com/huggingface/transformers/blob/main/src/transformers/loss/loss_d_fine.py#L339

RT-DETR for example handles this in a different way and keeps the other losses even when denoising is off:
https://github.com/huggingface/transformers/blob/main/src/transformers/loss/loss_rt_detr.py#L454

I found that denoising can lead to degraded validation performance with complex datasets, and turning it off helps. But without the other losses, D-FINE does not train properly.

[Abineshabee]

Hi @m-matthias, thanks for the clear report and pointers to the relevant sections.

I’ve reviewed the issue and compared the current implementation in loss_d_fine.py with loss_rt_detr.py. It looks like the auxiliary losses are incorrectly gated by denoising_meta_values, which disables them entirely when num_denoising=0.

I’d like to work on this by:

• Decoupling auxiliary loss computation from the denoising condition
• Ensuring only denoising-specific losses are disabled when num_denoising=0
• Adding/adjusting tests to verify the expected behavior

Please let me know if that approach sounds good or if there are any additional considerations before I proceed.

Thanks!

[m-matthias]

Hi @Abineshabee yes that sounds great, thank you.