git clone https://gitcode.com/ascend/ModelZoo-PyTorch.git
cd ModelZoo-PyTorch/ACL_PyTorch/built-in/cv/GroundingDINO 

git clone https://github.com/open-mmlab/mmdetection
cd mmdetection
git reset --hard cfd5d3a9

# 下载BERT权重，并放置于mmdetection目录下
https://huggingface.co/google-bert/bert-base-uncased/tree/main

# 下载MM-GroundingDINO权重，并放置于weights目录下
https://download.openmmlab.com/mmdetection/v3.0/mm_grounding_dino/grounding_dino_swin-b_pretrain_obj365_goldg_v3det/grounding_dino_swin-b_pretrain_obj365_goldg_v3de-f83eef00.pth

# 下载animals.png，并放置于images目录下
https://github.com/microsoft/X-Decoder/tree/main/inference_demo/images

# 下载NLTK权重（可选）。MM-GroundingDINO在进行Phrase Grounding推理时可能会进行名词短语提取，虽然会在运行时下载特定的模型，但是考虑到有些用户运行环境无法联网，因此可以提前下载。
方式一：
下载模型到~/nltk_data路径下：https://www.nltk.org/nltk_data/

方式二：
import nltk
nltk.download('punkt', download_dir='~/nltk_data')
nltk.download('averaged_perceptron_tagger', download_dir='~/nltk_data')

 │   ├── image_demo_npu.py //本仓提供单图推理脚本
 │   ├── video_demo_npu.py //本仓提供视频推理脚本
 │   ├── register_im2col_to_torchair.py //本仓提供torchair算子注册文件
 │   ├── register_roll_to_torchair.py //本仓提供torchair算子注册文件
 │   ├── requirements.txt //本仓提供
 │   └── install_requirements.sh //本仓提供依赖一键安装脚本

 │   ├── mmdetection_diff.patch  //本仓提供
 │   ├── mmengine_diff.patch     //本仓提供
 │   └── mmcv_diff.patch         //本仓提供

 mmdetection
 ├── demo
 │   ├── image_demo.py
 │   ├── video_demo.py
 │   ├── image_demo_npu.py //本仓提供单图推理脚本
 │   ├── video_demo_npu.py //本仓提供视频推理脚本
 │   ├── register_im2col_to_torchair.py //本仓提供torchair算子注册文件
 │   ├── register_roll_to_torchair.py //本仓提供torchair算子注册文件
 │   ├── requirements.txt //本仓提供
 │   └── install_requirements.sh //本仓提供依赖一键安装脚本
 ├── diff_patch
 │   ├── mmdetection_diff.patch  //本仓提供
 │   ├── mmengine_diff.patch     //本仓提供
 │   └── mmcv_diff.patch         //本仓提供
 ├── bert-base-uncased //BERT权重
 ├── weights
 │   └── grounding_dino_swin-b_pretrain_obj365_goldg_v3de-f83eef00.pth //MM-GroundingDINO权重
 ├── images
 │   └── animals.png
 ├── mmdet
 ├── resources
 ├── README.md
 ├── tests
 ├── tools
 ├── config
 └── ...

conda create -n groundingdino python=3.8
conda activate groundingdino

#在mmdetection目录下执行依赖一键安装脚本，会在自动拉取依赖仓并应用diff patch文件
source demo/install_requirements.sh

# 指定使用NPU ID，默认为0
export ASCEND_RT_VISIBLE_DEVICES=0

# 执行图片推理命令
python demo/image_demo_npu.py images/animals.png configs/mm_grounding_dino/grounding_dino_swin-b_pretrain_obj365_goldg_v3det.py --weight weights/grounding_dino_swin-b_pretrain_obj365_goldg_v3de-f83eef00.pth --texts '$: coco' --device npu (--loop 10)

# 执行视频推理命令
python demo/video_demo_npu.py demo/demo.mp4 configs/mm_grounding_dino/grounding_dino_swin-b_pretrain_obj365_goldg_v3det.py weights/grounding_dino_swin-b_pretrain_obj365_goldg_v3de-f83eef00.pth (--batch_size 16)

[root@autodl-container-33ab4e930e-c62598b5 mmdetection]# python demo/image_demo_npu.py images/animals.png configs/mm_grounding_dino/grounding_dino_swin-b_pretrain_obj365_goldg_v3det.py --weight weights/grounding_dino_swin-b_pretrain_obj365_goldg_v3de-f83eef00.pth --texts '$: coco' --device npu
报错一：
/usr/local/lib64/python3.11/site-packages/torch/nn/modules/transformer.py:20: UserWarning: Failed to initialize NumPy: _ARRAY_API not found (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)
报错二：
ImportError: libGL.so.1: cannot open shared object file: No such file or directory
[ERROR] 2025-12-23-12:04:37 (PID:20889, Device:-1, RankID:-1) ERR99999 UNKNOWN application exception

# 先卸载现有NumPy
pip uninstall -y numpy
# 安装与PyTorch 3.11兼容的稳定版本（推荐1.24.3）
pip install numpy==1.24.3

# 安装libGL核心依赖
yum install -y mesa-libGL-devel mesa-libGLU-devel
# 补充OpenCV其他常见依赖（避免后续报错）
yum install -y libX11-devel libXext-devel libXrender-devel libXt-devel