Make sure download stats work on Hugging Face, improve metadata

[NielsRogge]

Hi @wangbinDL,

Thanks for this nice work, great to see the models being released on the hub! Your paper was featured on the daily papers: https://huggingface.co/papers/2410.12628. Would be great to link the models to the paper.

To make download stats work, I wrote a PR similar to THU-MIG/yolov10#168 which I did for the YOLOv10 repository.

It leverages the PyTorchModelHubMixin developed by the 🤗 team to make sure a custom PyTorch model like yours:

• automatically inherits from_pretrained and push_to_hub methods
• have better metadata
• make sure download stats work.

Usage is as follows:

from doclayout_yolo import YOLOv10
from huggingface_hub import hf_hub_download

# Load a pre-trained model
filepath = hf_hub_download(repo_id="juliozhao/DocLayout-YOLO-D4LA-from_scratch", filename="doclayout_yolo_d4la_imgsz1600_from_scratch.pt")
model = YOLOv10(filepath)

# One can optionally push this to the hub
model.push_to_hub("juliozhao/DocLayout-YOLO-D4LA-from_scratch")

# Can now be reloaded as follows (and will increment download count)
model = YOLOv10.from_pretrained("juliozhao/DocLayout-YOLO-D4LA-from_scratch")

Besides that, some more suggestions to improve the HF release:

• one could create a Gradio demo on 🤗 Spaces to showcase what the model can do.
• one can link the datasets to the models (by adding them to the model cards)
• one can link the models/datasets to the paper page (by including this link: https://huggingface.co/papers/2410.12628 in the model/dataset cards)
• the dataset viewer currently doesn't work for your datasets. See here on how to enable them: https://huggingface.co/docs/datasets/image_dataset#object-detection

Let me know whether you need any help!

Kind regards,

Niels from HF

[JulioZhao97]

Thank you very much for your suggestions! If there are updates I will let you know immediately.

[wangbinDL]

Hi @NielsRogge

Thanks for your valuable suggestions. We've now:

• Used PyTorchModelHubMixin for our model, enabling from_pretrained and push_to_hub.
• Ensured download stats are working.
• Created a Gradio demo: https://huggingface.co/spaces/opendatalab/DocLayout-YOLO
• Linked our datasets to the models and paper: https://huggingface.co/papers/2410.12628

Your assistance is greatly appreciated!

Best,
Bin

[NielsRogge]

Thanks, looking great!

Note that for download stats to work, you will need to run the code snippet above for each of the model repos (basically overwriting them). As the class now inherits from PyTorchModelHubMixin, a config.json along with safetensors weights will be pushed.

[JulioZhao97]

Thanks, looking great!

Note that for download stats to work, you will need to run the code snippet above for each of the model repos (basically overwriting them). As the class now inherits from PyTorchModelHubMixin, a config.json along with safetensors weights will be pushed.

Noted, thanks for your contribution again!

[NielsRogge]

Hi @JulioZhao97 are you planning to run the script above?

Cause I see downloads aren't working yet

[JulioZhao97]

@NielsRogge Hello! I tested the script above at the time you submit the PR and I find

model = YOLOv10.from_pretrained("juliozhao/DocLayout-YOLO-D4LA-from_scratch")

requires yolov10n.pt, and the model weights I upload to huggingface reposity have not been re-uploaded. But other in the script above is fine. I will see to it in these two days.

[JulioZhao97]

@NielsRogge Hello! I tried model.push_to_hub but no model weights found in the repository, could you please tell me where is wrong?

Follow is the target repository:

[NielsRogge]

Hi @JulioZhao97 I just tried, it works for me: https://huggingface.co/nielsr/DocLayout-YOLO-D4LA-from_scratch/tree/main

Colab notebook: https://colab.research.google.com/drive/1N8Rfdu875lu5VCKG4uRUcM-Gi0o6qHom?usp=sharing.

Perhaps you might need to update your huggingface_hub version