Image Segmentation Tasks not returning Scores

🤗Hub
1

I am following this tutorial from Phil Schmid on how to deploy an image-segmentation model from the huggingface hub to a sagemaker endpoint.

The only difference I make to the code is that I deploy the model serverlessly.

The deployment works great, and the the masking works. However, I am not being returned the scores of each class. The values are all None for the scores.

e.g. {“score”: null, “label”: “wall”, “mask”: }

This applies to every model I try from the Hub, not just one. Would love clarity on this, and any potential fixes.

Thanks!

2

cc @philschmid

3

@ShuaHousetable have you tested deploying the model as real-time endpoint rather than serverless? does it work there?

4

Hi @philschmid , thanks for the response.

Yes, I tried it straight (without the serverless) like you suggested. Still receiving all “None” for the scores.

Running the notebook you created straight, using the DataScience3 Image in sagemaker.

5

Hey @philschmid, just wondering if there is any update on this. Thanks!

6

Hi,

It depends on which model you’re using. In case of semantic segmentation (like SegFormer), there’s no notion of scores per segment. This is because the model predicts a label per pixel. Hence the model only outputs scores per pixel, not per segment.

However, models like DETR and MaskFormer do output scores per segment.