ram_plus_swin_large_14m

Model Card for RAM++ Swin-Large-14M (IB-Robot)

Recognize Anything Model Plus (RAM++) with Swin-Large backbone, packaged for the IB-Robot framework. Recognizes semantic tags from images for robotic scene understanding.

Repository Structure

• inference_manifest.json — deployment routing (schema v3)
• assets/ — adapter config, torch weights (.pth), tag vocabulary (4585 classes), BERT tokenizer
• artifacts/ascend_310b/ — Ascend 310B1 OM (fp16)
• artifacts/ascend_310p/ — Ascend 310P1 OM

Deployment Backends

Input: observation.image float32 [1,3,384,384] NCHW Output: tag_logits float32 [1,4585] (4585-class multi-label classification)

Source Model

This bundle's torch weights (ram_plus_swin_large_14m.pth) are downloaded from the upstream RAM++ model:

• Model weights (HuggingFace): xinyu1205/recognize-anything-plus-model
• Source code (GitHub): github.com/xinyu1205/recognize-anything

The Ascend OM artifacts were converted from these torch weights. See scripts/download_perception_models.sh for the download procedure.

Citation

@inproceedings{ram_plus,
  title     = {Recognize Anything: A Strong Image Tagging Model},
  author    = {Zhang, Xinyu and Huang, Ziyu and Shen, Jun and Li, Zongtao and Zhao, Wen-Jun and Wang, Sheng and Sun, Yifan},
  booktitle = {arXiv preprint arXiv:2306.03514},
  year      = {2023}
}
@software{ib_robot,
  title   = {IB-Robot: Intelligence Boom Robot},
  url     = {https://gitcode.com/openeuler/IB_Robot},
  license = {Apache-2.0}
}