Performance Issue Fine-Tuning D-FINE S with custom dataset

[sergiosanchoasensio]

Hi everyone,

I’m fine-tuning the D-FINE S model initialized with Objects365 weights using a custom dataset, but I’m encountering significantly lower performance than expected.

Performance Comparison In just a single epoch:

• D-FINE S achieves mAP@50=0.384.
• YOLOv11 S (initialized with COCO weights) achieves mAP@50=0.842.

1. Dataset Preparation:

• Converted custom dataset annotations (x1,y1,x2,y2) to COCO format (x1, y1, w, h).
• Since the dataset includes truncated images, I added in the train.py: ImageFile.LOAD_TRUNCATED_IMAGES = True
• This dataset contains only one class, category_id=0, so I set num_classes: 1 and remap_mscoco_category: False.

2. Configuration Changes:

• Using dfine_hgnetv2_s_obj2custom.yml as the base configuration.
• Modified only the dataset path and adjusted the batch size to 8 for both training and validation.

3. Training Command:

CUDA_VISIBLE_DEVICES=0 torchrun --master_port=7777 --nproc_per_node=1 train.py \
  -c configs/dfine/custom/objects365/dfine_hgnetv2_s_obj2custom.yml \
  --seed=0 -t dfine_s_obj365.pth

What could be causing the poor performance? I’d appreciate any insights or suggestions for debugging this issue. Thanks in advance!

PS: I tried fine-tuning for 20 epochs, and the mAP@50 only improves slightly to 0.4. Using COCO weights instead of Objects365 weights does not lead to any improvement either.

[Peterande]

Can you try training from scratch without any pre-training weights?

[sergiosanchoasensio]

Can you try training from scratch without any pre-training weights?

I trained the model from scratch but obtained an mAP@50 of -1.000. I suspect this may be due to the annotation format not aligning with D-FINE's requirements. Could you kindly clarify the correct format? Should the annotations use COCO format x1, y1, w, h, YOLO format (normalized, compatible with MIT YOLOv9) x_center, y_center, w, h, or another standard?

Thank you in advance.

[prashant-dn]

Same here - I tried these 2 variants : dfine_hgnetv2_${model}_obj365.yml and dfine_hgnetv2_${model}_custom.yml and the mAP50 never goes above 0.35 while even YOLOv9c gets to 0.7

[AnishMathewOommen]

@sergiosanchoasensio Could you please tell me how to use the pretrained model checkpoint file. Please share codes if you are no issues in sharing

[S130111]

I also encountered the same problem. Does anyone solve it?

[ArgoHA]

@AnishMathewOommen I used this command for single GPU setup and custom dataset torchrun train.py -c configs/dfine/custom/objects365/dfine_hgnetv2_m_obj2custom.yml --use-amp --seed=0 -t dfine_m_coco.pth

[FrancescoSaverioZuppichini]

I am also experiencing terrible performanes with dfine, it just sucks (more likely I suck at training) but I am using this dataset with just one classes no matter what I do I end up with map at 0.15

[ArgoHA]

@FrancescoSaverioZuppichini try out this repo

[FrancescoSaverioZuppichini]

@FrancescoSaverioZuppichini try out this repo

Thanks a lot man, is on my todo list!

[FrancescoSaverioZuppichini]

@FrancescoSaverioZuppichini try out this repo

man what the fuck your code is amazing, it is training so much faster and it is so well done - are you an advance AI from the future?