Model Card for SAM2.1 Hiera Tiny (IB-Robot)
Segment Anything Model 2.1 (SAM2.1) with Hiera-Tiny backbone, packaged for the IB-Robot framework. Performs prompted segmentation: given an image and point/box prompts, generates instance masks.
Repository Structure
inference_manifest.json— deployment routing (schema v3)assets/— adapter config, torch weights (.pt)artifacts/ascend_310b/— encoder + decoder OM (batch=1)artifacts/ascend_310p/— encoder + decoder OM (batch=4)
Deployment Backends
| Target | Backend | Runtime | Batch | Hardware |
|---|---|---|---|---|
ascend_310b | ascend | ACL | 1 | Ascend 310B1 |
ascend_310p | ascend | ACL | 4 | Ascend 310P1 |
torch_cpu | torch | PyTorch | — | CPU |
torch_cuda | torch | PyTorch | — | NVIDIA GPU |
Each Ascend deployment runs a two-stage pipeline: encoder then decoder.
Encoder input: host.sam2.image float32 [N,3,1024,1024] NCHW
Decoder outputs: host.sam2.mask_logits [N,1,256,256], host.sam2.iou_predictions [N,1], host.sam2.low_res_masks [N,1,256,256]
Source Model
This bundle's torch weights originate from the upstream SAM2.1 checkpoint:
- HuggingFace: facebook/sam2.1-hiera-tiny
- Direct checkpoint:
https://dl.fbaipublicfiles.com/segment_anything_2/092824/sam2.1_hiera_tiny.pt
The Ascend OM artifacts were converted from these torch weights. Download script: scripts/download_perception_models.sh.
Citation
@inproceedings{sam2,
title = {SAM 2: Segment Anything in Images and Videos},
author = {Ravi, Nikhila and Gabeur, Valentin and Hu, Yuan-Ting and others},
booktitle = {arXiv preprint arXiv:2408.00714},
year = {2024}
}
@software{ib_robot,
title = {IB-Robot: Intelligence Boom Robot},
url = {https://gitcode.com/openeuler/IB_Robot},
license = {Apache-2.0}
}