Thanks for your work,excellent! some question about yolo-world finetune freeze, glip Pseudo label and prompt.

[Meicsu199345]

Question

Thanks for your work,excellent! background I have a dataset about risk_control 100classes 200k images and 400k bboxes。 class such as 1:vulgar tongue kiss 2:porn bed photos 3:human smoking scenes

i want use it not only detect person ,car, human smoking in bed, so i need Re-parameterization?

question Q1: can i get some suggestion about my method is correct?

• change 1 update torch >=2.0.0 WARNING ⚠️ Upgrade to torch>=2.0.0 for deterministic training.
• change 2 cache='disk'
• change 3 load_model yolov8s-worldv2.pt(maybe need freeze)
• change 4 freeze=10 (backbone),freeze only in backbone or all model is better for finetune? the paper show
• add 10% background image

Q2:how can i use yolo world to do prompt?

Q3:how can i get Pseudo label（bboxes）by glip use this github-rep?

Q4:how can i do A.1. Re-parameterization for RepVL-PAN use ultralytics?

thanks a lot, yolo world let my study easy~

[wondervictor]

Hi @Meicsu199345, sorry for the late reply!! Thanks for your interest in YOLO-World!

1. Using torch~=1.13 and requirements.txt to install is fine and we'll remove the support for torch~2.0.
2. We only freeze the CLIP backbone for fine-tuning, other parameters are trainable. If you have a large vocabulary, you can also make CLIP trainable.
3. For prompt tuning, you can refer to docs/prompt_yolo_world. You just need to precompute the text embeddings and use the SimpleYOLOWorldDetector now.
4. The pseudo labels are listed here: docs/data.
5. The repo is independent from Ultralytics and we only support re-parameterization in this repo.

[Meicsu199345]

Hi @Meicsu199345, sorry for the late reply!! Thanks for your interest in YOLO-World!

1. Using torch~=1.13 and requirements.txt to install is fine and we'll remove the support for torch~2.0.
2. We only freeze the CLIP backbone for fine-tuning, other parameters are trainable. If you have a large vocabulary, you can also make CLIP trainable.
3. For prompt tuning, you can refer to docs/prompt_yolo_world. You just need to precompute the text embeddings and use the SimpleYOLOWorldDetector now.
4. The pseudo labels are listed here: docs/data.
5. The repo is independent from Ultralytics and we only support re-parameterization in this repo.

Thank you very much for your answer～