Different SAM model not giving correct masks

[realtimshady1]

Actions before raising this issue

• I searched the existing issues and did not find anything similar.
• I read/searched the docs

Steps to Reproduce

1. Download sam_vit_b_01ec64.pth
2. Replace decoder.onnx in /root/cvat/cvat-ui/plugins/sam/assets/decoder.onnx
3. Update serverless/pytorch/facebookresearch/sam/nuclio/model_handler.py

self.sam_checkpoint = "/opt/nuclio/sam/sam_vit_b_01ec64.pth"
self.model_type = "vit_b" 

4. Update serverless/pytorch/facebookresearch/sam/nuclio/function.yaml

value: curl -O https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth

5. Build

docker compose -f docker-compose.yml -f docker-compose.override.yml -f docker-compose.https.yml -f components/serverless/docker-compose.serverless.yml -f docker-compose.dev.yml up -d --build

6. Deploy function deploy_cpu.sh serverless/pytorch/facebookresearch/sam/nuclio/
   

7. Use the SAM tool on an image and the results look like the following
   

Expected Behavior

I believe that the new decoder is not being used properly. Either I did not build correctly or further steps are needed to accommodate the SAM-b model

Possible Solution

No response

Context

No response

Environment

Version: ad1bcd5474582c24119a4a49909eb3f1e3979899 (HEAD -> develop, origin/develop, origin/HEAD)
Docker version: 20.10.22
Docker Swarm/ Kubernetes: No
OS: Debian GNU/Linux 11 (bullseye)

[bsekachev]

Make sure that the previous decoder.onnx not cached by the browser

[xx674967]

HI @bsekachev, I followed the steps above to try to change the SAM model, and I also modified the ONNX code according to the issue #6019 you sent. However, I'm encountering this problem when starting CVAT for annotation. How can I resolve this issue? Thank you.

[realtimshady1]

did you follow the advice to remove the cached model? Depending on your browser, you need to do a hard refresh to ensure that it is using the new model