Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Coordinates are all over the place for local testing, but works perfectly on your gradio webapp #10

Open
pentatonictritones opened this issue Jan 23, 2025 · 11 comments

Comments

@pentatonictritones
Copy link

Hi, excited to see this performance.

The gradio space has very accurate performance
https://huggingface.co/spaces/Aheader/gui_test_app

So i tried running it locally but the coordinate boxes once i scale it by the image width and height is just all over the place. I get very poor results locally.

Could you share more details about the gradio implementation?

I tried using both the prompt examples in your docs but it is awful compared to the webapp - which gives perfect results everytime

Thanks!!

@pentatonictritones
Copy link
Author

For reference, my user prompt is

"Click the edit button"

Image

Locally:

Image

On gradio:

Image

@AHEADer
Copy link
Collaborator

AHEADer commented Jan 23, 2025

Hi
We deploy a cloud server following the deployment guide in readme. The space app just make requests to the server, you can check the app's source code to check the prompt. The prompt should be "Output only the coordinate of one box in your response. "

@liy9426
Copy link

liy9426 commented Jan 23, 2025

Same here. Is there any difference deploying locally? I did the same instructions as in the guide but it outputs different results every time. Thanks

@pentatonictritones
Copy link
Author

Hi, thanks, i checked out the notion file but unfortunately at least with the 2b and 7b model, i get very poor results locally

I've tried running it locally with all different frameworks and the methods but the coordinates continue to have 0% accuracy while the gradio model is very accurate.

Also, i tried out many different prompts, but it definitely isn't the issue

It would be great if anyone has any success running this locally.

@JjjFangg
Copy link
Collaborator

Have you tried converting the output coordinates using the transformation described in Coordinate Mapping Section

@pentatonictritones
Copy link
Author

Have you tried converting the output coordinates using the transformation described in Coordinate Mapping Section

hi!

Image

Yes. I am certain it is not a coordinate conversion issue, because if i prompt it 5 times, i get 5 vastly different coordinates

@AHEADer
Copy link
Collaborator

AHEADer commented Jan 23, 2025

Hi, thanks, i checked out the notion file but unfortunately at least with the 2b and 7b model, i get very poor results locally

I've tried running it locally with all different frameworks and the methods but the coordinates continue to have 0% accuracy while the gradio model is very accurate.

Also, i tried out many different prompts, but it definitely isn't the issue

It would be great if anyone has any success running this locally.

Are you using Ollama? Or vLLM?

@pentatonictritones
Copy link
Author

pentatonictritones commented Jan 23, 2025

Hi, thanks, i checked out the notion file but unfortunately at least with the 2b and 7b model, i get very poor results locally
I've tried running it locally with all different frameworks and the methods but the coordinates continue to have 0% accuracy while the gradio model is very accurate.
Also, i tried out many different prompts, but it definitely isn't the issue
It would be great if anyone has any success running this locally.

Are you using Ollama? Or vLLM?

ollama as the doc suggested, as well as transformers

@liy9426
Copy link

liy9426 commented Jan 23, 2025

Image
Also noticed that it outputs some texts that are definitely not in the pic I provide. Wondering what's the reason for that

@JjjFangg
Copy link
Collaborator

It seems that the issue might be related to the precision of the model deployed with Ollama. We are currently working on optimizing it. In the meantime, you can prioritize testing with vLLM or the cloud deployment.

@pentatonictritones
Copy link
Author

It seems that the issue might be related to the precision of the model deployed with Ollama. We are currently working on optimizing it. In the meantime, you can prioritize testing with vLLM or the cloud deployment.

thanks

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

4 participants