王泽宇 2025-04-01 20:41 浙江
4月2日晚8点，更有核心作者直播分享SptialLM更多技术细节，来预约!
模型地址：
conda install -y nvidia/label/cuda-12.4.0::cuda-toolkit conda-forge::sparsehash[tool.poetry]name = "spatiallm"version = "0.0.1"description = "SpatialLM: Large Language Model for Spatial Understanding"authors = ["ManyCore Research Team"]license = "Llama3.2"readme = "README.md"[tool.poetry.dependencies]python = ">=3.10,<3.13"transformers = ">=4.41.2,<=4.46.1"safetensors = "^0.4.5"pandas = "^2.2.3"einops = "^0.8.1"numpy = "^1.26"scipy = "^1.15.2"scikit-learn = "^1.6.1"toml = "^0.10.2"tokenizers = ">=0.19.0,<0.20.4"huggingface_hub = ">=0.25.0"rerun-sdk = ">=0.21.0"shapely = "^2.0.7"bbox = "^0.9.4"terminaltables = "^3.1.10"open3d = "^0.19.0"nvidia-cudnn-cu12 = "*"nvidia-nccl-cu12 = "*"poethepoet = {extras = ["poetry-plugin"], version = "^0.33.1"}addict = "^2.4.0"[tool.poe.tasks]install-torchsparse = "pip install git+https://github.com/mit-han-lab/torchsparse.git"[build-system]requires = ["poetry-core", "setuptools", "wheel", "torch"]build-backend = "poetry.core.masonry.api"
pip install poetry && poetry config virtualenvs.create false --localpoetry install
TorchSparse是一种基于PyTorch构建的高效稀疏张量处理库，专门用于加速三维稀疏卷积神经网络的训练和推理。
poe install-torchsparse点云数据（Point Cloud）是一种由大量空间点组成的数据集合，每个点记录了其在三维空间中的位置坐标，通常还包含颜色或反射强度信息，可用于精确描述三维物体或场景的空间结构。 如果想要使用自己录制的视频，可以参照MASt3R-SLAM进行三维重建。
huggingface-cli download manycore-research/SpatialLM-Testset pcd/scene0000_00.ply --repo-type dataset --local-dir .
modelscope download manycore-research/SpatialLM-Llama-1B --local_dir ./manycore-research/SpatialLM-Llama-1Bpython inference.py --point_cloud ./pcd/scene0000_00.ply --output scene0000_00.txt --model_path ./manycore-research/SpatialLM-Llama-1B
python visualize.py --point_cloud ./pcd/scene0000_00.ply --layout scene0000_00.txt --save scene0000_00.rrdrerun scene0000_00.rrd考虑到部分同学配置环境可能会遇到一些问题，我们在AutoDL平台准备了SpatialLM的环境镜像，点击下方链接并直接创建Autodl示例即可。 https://www.codewithgpu.com/i/datawhalechina/self-llm/SpatialLM
