Transformers
Safetensors
surgical-video
spatio-temporal-grounding
medical-vision-language-model
eccv-2026
Instructions to use linzher/RefineRank with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use linzher/RefineRank with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("linzher/RefineRank", device_map="auto") - Notebooks
- Google Colab
- Kaggle
File size: 1,117 Bytes
12b6f85 4a24473 12b6f85 4a24473 12b6f85 4a24473 12b6f85 4a24473 12b6f85 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 | # Checkpoints
RefineRank 推理/训练所需的全部权重,平铺结构如下:
```
checkpoints/
├── vlm/ # 冻结的 MedVLM(Qwen2.5-VL 架构,HF 格式,全部文件直接放这里)
│ ├── config.json / model-0000X-of-00004.safetensors / tokenizer.json / ...
├── grounding_dino/ # 冻结的 GroundingDINO 权重
│ └── groundingdino_swinb_cogcoor.pth
└── refinenet/ # 唯一可训练模块 RefineNet(约 1.25M 参数)
├── proposal_adapter_full.pt # 随论文发布的最终提交权重
└── deployment_manifest.json
```
- 从 Hugging Face 一键恢复本目录:`hf download linzher/RefineRank --local-dir .`
- `vlm/` 与 `grounding_dino/` 在训练与推理中都保持冻结,仅用于特征与候选框提取。
- `python interface.py predict` 不带 `--checkpoint` 时优先使用本目录的平铺权重;
训练新产生的 `run_<时间戳>/` 目录(`python interface.py train` 的输出)也会被自动发现。
- 大权重文件不要提交到 git。
|