Improve README, load CLIPCount from HF

[NielsRogge]

Hi,

Thanks for this interesting work. I contributed YOLOS back in the day to the Transformers library, so it's nice to see you're leveraging it here. Since you do seem to use this model, and not the YOLO model from Ultralytics, I assume we can update the link in the README (and perhaps as well the mention in the paper, which mentions YOLO instead of YOLOS).

Next to that, I see you leverage CLIPCount, it would be nice to make this model available on 🤗 for easy inference, and better discoverability. Here's how it can be used:

from models import CLIPCount

model = CLIPCount()

# load weights
model.load_state_dict(...)

# push to hub
model.push_to_hub("your-hf-org-or-username/clip-count-base")

# reload
model = CLIPCount.from_pretrained("your-hf-org-or-username/clip-count-base")

This way, people also don't need to manually download the checkpoint, it loads automatically from the hub.

Moreover, it leverages safetensors for serialization rather than pickle, which is considered unsafe.

Would you be interested in this integration?

Cheers,

Niels
Open-source @ HF

Note

Please don't merge this PR before pushing and reloading :)

[ozzafar]

Thanks @NielsRogge for you inputs!
I take your suggestion to upload CLIPCount to hugging face :)
Regarding YOLO, at some point we moved to Yolov9 in our experiments, I'll modify the code accordingly - thanks for pointing this out!