AI

Windows使用WSL2 Ubuntu运行HoloPart+SAMPart3D分割3D

Generative 3D Part Amodal Segmentation

Posted by Gumc on April 29, 2025

基础

要求

  • 电脑需要含有NVIDIA的显卡，且支持CUDA
  • 已安装NVIDIA显卡驱动
  • 注意显卡版本的计算能力要求，我的显卡NVIDIA GeForce GTX 1050 Ti只支持6.x算力，只能用CUDA11.x版本

Windows10使用wsl 2 安装Linux

安装

  • 在设置-应用-应用和功能-程序和功能-安装或关闭Windows功能 ✔Hyper-V和适用于Linux的Windows子系统

  • 1
2
3
4

      # 查看适用于Windows的Linux版本
  wsl.exe --list --online
  # 安装想要的版本
  wsl --install -d Ubuntu-24.04
Linux迁移到D盘
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

  # 查看已安装的发行版名称:
  wsl -l -v
  # 创建导出的目标文件夹
  D:\ProgramFiles\WSL
  # 执行导出命令
  wsl --export Ubuntu-24.04 D:\ProgramFiles\WSL\ubuntu-export.tar 
  # 注销 (卸载) 原来的 Ubuntu 发行版
  wsl --unregister Ubuntu-24.04
  # 创建目标安装目录
  D:\ProgramFiles\WSL\Ubuntu-24.04
  # 执行导入命令
  wsl --import Ubuntu-24.04 D:\ProgramFiles\WSL\Ubuntu-24.04 D:\ProgramFiles\WSL\ubuntu-export.tar --version 2
  # 验证导入
  wsl -l -v
  # 启动新导入的 Ubuntu
  wsl -d Ubuntu-24.04
添加新用户
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

  # 添加新用户
  adduser gumc
  # 显示
  # Adding user `gumc' ...
  # Adding new group `gumc' (1000) ...
  # Adding new user `gumc' (1000) with group `gumc' ...
  # Creating home directory `/home/gumc' ...
  # Copying files from `/etc/skel' ...
  # New password: <输入你的密码>
  # Retype new password: <再次输入密码>
  # passwd: password updated successfully
  # Changing the user information for gumc
  # Enter the new value, or press ENTER for the default
  #         Full Name []: <可以直接按 Enter 跳过>
  #         Room Number []: <可以直接按 Enter 跳过>
  #         Work Phone []: <可以直接按 Enter 跳过>
  #         Home Phone []: <可以直接按 Enter 跳过>
  #         Other []: <可以直接按 Enter 跳过>
  # Is the information correct? [Y/n] Y

  # 将新用户添加到 sudo 组 (使其拥有管理员权限):
  usermod -aG sudo gumc

  # 设置默认登录用户
  echo "default=gumc" >> /etc/wsl.conf

  # vim /etc/wsl.conf确认是以下内容
  [user]
  default=gumc
  
  # 关闭和重启操作
  exit
  wsl --shutdown
  wsl -d Ubuntu-24.04
创建Data/Pack文件夹
1
2
3
4

  # 注意:避免将软件包安装在Windows挂载目录,例如/mnt/d/ProgramFiles/WSL肯定会出问题，而是放在WSL的Linux文件系统，例如/home/gumc/
  sudo mkdir -p /data/pack
  # 修改文件夹权限
  sudo chown -R gumc:gumc /data

安装 Conda (Miniconda 推荐):

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58

  # 进入安装目录
  cd /data/pack
  # 下载 Miniconda 安装脚本：
  wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
  # 运行安装脚本
  bash Miniconda3-latest-Linux-x86_64.sh
  # 按 Enter 查看许可，然后输入 yes 同意。
  # 确认安装路径：
  /data/miniconda3
  # 当询问是否初始化 Miniconda3 (conda init) 时，输入 yes。
  # 初始化 Miniconda: 当询问是否运行 conda init 时 (例如 Do you wish the installer to initialize Miniconda3 by running conda init? [yes|no])，强烈建议输入 yes。这会自动修改你的 shell 配置文件 (.bashrc 或 .zshrc)，以便每次打开终端时都能自动配置好 Conda 环境。

  # 关闭后重新打开
  exit
  wsl --shutdown
  wsl -d Ubuntu-24.04

  # 检测Conda
  which conda

  # Conda基本操作
  # 显示所有环境
  conda env list
  # 删除环境,sampart3d是环境名
  conda remove -n sampart3d --all
  # 清除cache 索引缓存、锁文件、未使用的包、tarballs
  conda clean --all -y

  # 确保你安装在 Windows 上的 NVIDIA 驱动程序是最新的，并且支持 WSL2
  nvidia-smi
  # 启动 WSL2 终端并激活你的 Conda 环境(强制使用python 3.10)
  conda create -n sampart3d python=3.10 -y
  conda activate sampart3d
  # 其他操作：如需要退出当前环境
  # conda deactivate
  # # 导出环境
  # conda env export --from-history --no-builds > environment.yml
  # # 修改环境后如果要更新环境
  # conda env update --file environment.yml --prune

  # 添加 Channel (并影响顺序)
  vim /data/miniconda3/.condarc
  channels:
    - nvidia        # 优先级 1
    - conda-forge   # 优先级 2
    - defaults      # 优先级 4 (Conda 官方默认 channel)
    - rapidsai
    - pytorch 
  channel_priority: flexible #strict, flexible, disabled (后面会解释)
  # auto_update_conda: true (可选)
  # show_channel_urls: true (可选，在解决输出中显示 channel URL)
  # 此时，如果你查看 channels 列表，顺序可能是：
  # 0. rapidsai
  # 1. nvidia (最后添加的，优先级最高)
  # 2. pytorch
  # 3. conda-forge
  # 4. defaults (通常默认存在)
  # conda config --show-sources

WSL2 中安装 CUDA

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

  
  # 确保在想要的环境 conda activate sampart3d
  # 如果你之前在这个环境里尝试安装过 cudatoolkit 或 cudnn，先移除它们以确保干净的状态：
  # conda remove -n sampart3d cudatoolkit cuda --force -y
  # 查看cuda
  # conda list | grep -E "cuda|cuml|cudatoolkit|libcumlprims"
  # 安装cuda,明确指定所有 CUDA 组件版本
  conda install cudatoolkit=11.8.0 cuda-nvcc=11.8.89 cuda-cudart-dev=11.8.89 cuda-cccl=11.8.89 cuda-cudart=11.8.89 python=3.10  #不加-y方便确认要安装的版本是否正确
  #如不确定版本，可通过conda search cuda-cudart-dev=11.8.* 查找版本号
  # conda install cudatoolkit=11.8 python=3.10 
  # 固定cudatoolkit版本:创建或追加到 pinned 文件
  echo $CONDA_PREFIX
  echo "cudatoolkit ==11.8.*" >> $CONDA_PREFIX/conda-meta/pinned
  # 安装其他
  # conda install cuda=11.8 cuda-version=11.8 
  # conda install rapids=23.12 cudf=23.12 cuml=23.12  #不加-y方便确认要安装的版本是否正确
  # 在安装其他依赖包时，显式声明 CUDA 版本，防止 Conda 自动升级
  # conda install rapids=23.12 cudf=23.12 cuml=23.12 cudatoolkit=11.8.0 cuda-nvcc=11.8.89 cuda-cudart-dev=11.8.89 cuda-cccl=11.8.89 cuda-cudart=11.8.89 python=3.10 
  # 安装cuda相关
  # conda install rapids=23.12 cudf=23.12 cuml=23.12 python=3.10 cuda-version=11.8 -y
  # conda install -c rapidsai -c conda-forge -c nvidia  rapids=23.12 python=3.10 'cuda-version>=11.8,<=11.8' 'pytorch=*=*cuda*'
  
  # 验证 Conda 环境中的 CUDA
  nvcc --version
  # 检查 CUDA 开发头文件
  ls $CUDA_HOME/include/cuda/std/cudart*.h

  # 如安装失败，很可能是因为需要更新(不失败也建议进行)
  sudo apt update
  sudo apt install build-essential -y
  # 如提示gcc版本不对，则安装gcc11和g++11(不失败也建议进行)
  conda install -c conda-forge gcc=11 gxx=11 -y
  # 验证安装
  # gcc --version  # 应输出 11.x.x
  # g++ --version  # 应输出 11.x.x

安装SAMPart3D

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

  #下载SAMPart3D
  cd /data
  git clone https://github.com/Pointcept/SAMPart3D.git
  cd SAMPart3D

  # 确保在想要的环境
  conda activate sampart3d

  # 安装基本模块：torch 
  # pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118  
  # conda install pytorch torchvision torchaudio pytorch-cuda=11.8
  conda install pytorch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 torch-scatter=2.1.1 cudatoolkit=11.8.0 cuda-nvcc=11.8.89 cuda-cudart-dev=11.8.89 cuda-cccl=11.8.89 cuda-cudart=11.8.89 python=3.10 
  #[备注：cu118根据自己的cuda版本决定，可通过nvcc --version查看cuda版本],--index-url代表pip 将从这个地址查找和下载包，而不是使用默认的官方 PyPI 服务器 (pypi.org)
  # 验证
python - <<EOF
  import torch
  import torch_scatter
  print(f"PyTorch 版本: {torch.__version__}")        # 应输出 2.2.2 或 2.3.0
  print(f"torch-scatter 版本: {torch_scatter.__version__}")  # 应输出 2.1.2+ 或兼容版本
EOF
  # 如果torch-scatter报错则重新安装torch-scatter
  pip install torch-scatter -f https://data.pyg.org/whl/torch-2.2.0+cu118.html --force-reinstall --no-cache-dir

  # requirements.txt 中的包(使用aliyun会更快)
  pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt

  # 为 PTv3-object 安装模块
  cd libs/pointops
  pip install . # 不建议使用python setup.py install
  cd ../..
  # 如果显示   OSError: CUDA_HOME environment variable is not set. Please set it to your CUDA install root.
  # echo 'export CUDA_HOME=$CONDA_PREFIX' >> ~/.bashrc
  # source ~/.bashrc
  # echo $CUDA_HOME
  # 可能需要使用，但重启后又得重新设置
  export CUDA_HOME=$CONDA_PREFIX:$CONDA_PREFIX/include && echo "Set CUDA_HOME to: $CUDA_HOME"
  echo $CUDA_HOME #为了方便建议使用最后面的"Conda 环境激活时自动设置"
  # spconv (SparseUNet)
  # refer https://github.com/traveller59/spconv
  # pip install spconv-cu124  # choose version match your local cuda
  pip  install spconv-cu118

安装Flash-attention

1
2
3
4
5
6
7
8
9
10
11
12

# 确定是否启用C++11 ABI,如果是False则下载FALSE的
python - <<EOF
import torch
print("PyTorch 是否启用了 C++11 ABI:", torch._C._GLIBCXX_USE_CXX11_ABI)
EOF

# Linux or MacOs
# 下载地址https://github.com/Dao-AILab/flash-attention/releases/download
# 根据cuda、torch等的版本来下载安装
# 将 N 替换为你想要的最大并行作业数，例如 4
MAX_JOBS=4 pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu11torch2.2cxx11abiFALSE-cp310-cp310-linux_x86_64.whl --no-build-isolation
# MAX_JOBS=4 pip install flash-attn --no-build-isolation

安装加速模块

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

pip install ninja git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch
# # 如提示运行错误找不到libcuda.so
# # 运行 find 命令来查找 libcuda.so（可能带有版本号后缀 .1 或类似）：
# sudo find /usr/lib /lib -name "libcuda.so*"
# sudo mkdir -p /usr/local/cuda/lib64
# sudo ln -s /usr/lib/wsl/lib/libcuda.so.1 /usr/local/cuda/lib64/libcuda.so
# # # 验证
# ls -l /usr/local/cuda/lib64/libcuda.so
# # 更新链接器缓存，让系统知道这个新链接
# sudo ldconfig

# 可能需要使用，但重启后又得重新设置
# export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH && echo "Set LD_LIBRARY_PATH to: $LD_LIBRARY_PATH"
# export LD_LIBRARY_PATH=/data/miniconda3/envs/sampart3d/lib/:$LD_LIBRARY_PATH
# export LD_LIBRARY_PATH=/usr/lib/wsl/lib:/usr/local/cuda/lib64:$CONDA_PREFIX/lib:$LD_LIBRARY_PATH
# export LIBRARY_PATH=/usr/lib/wsl/lib:/usr/local/cuda/lib64:$CONDA_PREFIX/lib:$LIBRARY_PATH
# echo $LD_LIBRARY_PATH
# echo $LIBRARY_PATH

# export CPLUS_INCLUDE_PATH=$CONDA_PREFIX/include:$CPLUS_INCLUDE_PATH
# export C_INCLUDE_PATH=$CONDA_PREFIX/include:$CPLUS_INCLUDE_PATH
# echo $CPLUS_INCLUDE_PATH
# echo $C_INCLUDE_PATH
# export PATH="$CONDA_PREFIX/bin:$PATH"
# echo $PATH
# using GPU-based HDBSCAN clustering algorithm
# refer https://docs.rapids.ai/install
# pip install --extra-index-url=https://pypi.nvidia.com cudf-cu12==24.6.* cuml-cu12==24.6.*
# 进入https://docs.rapids.ai/install/#selector选择安装命令行
# conda install -c nvidia -c rapidsai -c conda-forge  \
#     cudf=25.04 cuml=25.04 python=3.12 "cuda-version>=12.0,<=12.8"
conda install rapids=23.12.* cudf=23.12.* cuml=23.12.* cudatoolkit=11.8.0 cuda-nvcc=11.8.89 cuda-cudart-dev=11.8.89 cuda-cccl=11.8.89 cuda-cudart=11.8.89 python=3.10 

# 如安装错误则卸载
# conda uninstall rapids
# conda list | grep -E "rapids|cudf|cuml" # Linux/macOS

SAMPart3D训练

下载预训练的 PTv3 对象

https://huggingface.co/yhyang-myron/SAMPart3D/tree/main (windows下载ptv3-object.pth ，放到d:/ProgramFiles/WSL/SAMPart3D/ckpt下)

数据预处理

使用 Blender 渲染 3D glb 网格的多视图 RGB 和深度

1
2
3

wget https://download.blender.org/release/Blender4.0/blender-4.0.0-linux-x64.tar.xz
tar -xf blender-4.0.0-linux-x64.tar.xz
# rm blender-4.0.0-linux-x64.tar.xz
  • 然后渲染 rgb 和深度
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71

# 在Windowsd的D:\ProgramFiles\WSL\下创建文件夹SAMPart3D文件夹,并创建ckpt,mesh_root,data_root文件夹
  SAMPart3D
  |-- ckpt
      |-- ptv3-object.pth
  |-- mesh_root
      |-- 000001.ZaanseSchans-Bike.glb
  |-- data_root
      |-- 000001.ZaanseSchans-Bike
          |-- meta.json
          |-- render_0000.webp
          |-- depth_0000.exr
          ...
# 连接windows下的文件夹到Linux下
ln -s "/mnt/d/ProgramFiles/WSL/SAMPart3D/ckpt" /data/SAMPart3D/ckpt
ln -s "/mnt/d/ProgramFiles/WSL/SAMPart3D/mesh_root" /data/SAMPart3D/mesh_root
ln -s "/mnt/d/ProgramFiles/WSL/SAMPart3D/data_root" /data/SAMPart3D/data_root

# 修改配置
vim /data/SAMPart3D/configs/sampart3d/sampart3d-trainmlp-render16views.py
# misc custom setting
batch_size = 16  # 增大批量大小，根据GPU显存调整
num_worker = 2   # 增加数据加载的工作进程数，建议设置为CPU核心数的2-4倍
# scheduler settings
epoch = 1000       # 减少总训练轮次
eval_epoch = 1000  # 减少评估轮次
# 数据采样数量
data = dict(
  train=dict(
      sample_num=10000,        # 减少每个点云的采样点数
      pixels_per_image=128,    # 减少每张图像的像素数
      batch_size=16,          # 增大批处理大小
  )
)
# 模型结构简化
model = dict(
  backbone = dict(
      enc_depths=(2, 2, 2, 4, 8),        # 减少编码器深度
      enc_channels=(32, 64, 128, 256, 384), # 可以适当减少通道数
      enc_patch_size=(512, 512, 512, 512, 512), # 减小patch大小
      # enable_flash=True,    # 启用Flash Attention加速
  )
)
# 优化器和调度器设置
# optimizer = dict(type="AdamW", lr=2e-4)  # 适当增大学习率
# scheduler = dict(
#     type="OneCycleLR",
#     max_lr=[2e-4],
#     pct_start=0.1,
#     div_factor=10.0,
#     final_div_factor=10.0,
# )

# 1050Ti不支持Flash-Attention，所以只能设置False
enable_flash=False

# ${PATH_TO_BLENDER} -b -P blender_render_16views.py ${MESH_PATH} ${TYPES} ${OUTPUT_PATH}
/data/pack/blender-4.0.0-linux-x64/blender -b -P /data/SAMPart3D/tools/blender_render_16views.py /data/SAMPart3D/mesh_root/ZaanseSchansBike.glb glb /data/SAMPart3D/data_root/ZaanseSchansBike
# 如提示blender-4.3.2-linux-x64/blend![1746457238438](image/2025-04-29-holopart/1746457238438.glb)er: error while loading shared libraries: libSM.so.6: cannot open shared object file: No such file or directory,则安装
sudo apt install libsm6
# 如提示 Unable to find 'libdecor-0.so' 
sudo apt install libdecor-0-0 libdecor-0-plugin-1-cairo
# 使用blender-4.3.2会出现BLENDER_EEVEE_NEXT,例如：
# bpy_struct: item.attr = val: enum "BLENDER_EEVEE" not found in ('BLENDER_EEVEE_NEXT', 'BLENDER_WORKBENCH', 'CYCLES')
# AttributeError: 'SceneEEVEE' object has no attribute 'use_ssr'
# AttributeError: 'Mesh' object has no attribute 'use_auto_smooth'
# 以上问题似乎比较难解决，建议用blender-4.0.0

# 注意
# glb的名字不能包含连字符 (-)。它们只能包含字母、数字和下划线 (_)，且不能以数字开头

训练

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84

# Download pretrained PTv3-object https://huggingface.co/yhyang-myron/SAMPart3D/tree/main
# # 下载ptv3-object.pth放到ckpt
# /mnt/d/ProgramFiles/WSL/SAMPart3D/ckpt

# 修改configs/sampart3d/sampart3d-trainmlp-render16views.py中的data_root、mesh_root和backbone_weight_path
vim /data/SAMPart3D/configs/sampart3d/sampart3d-trainmlp-render16views.py
# 修改为以下参数:
data_root = "/data/SAMPart3D/data_root"
mesh_root = "/data/SAMPart3D/mesh_root"
backbone_weight_path = "/data/SAMPart3D/ckpt/ptv3-object.pth"

# 链接cuda的路径
mkdir -p /usr/local/cuda/
# sudo ln -s $CONDA_PREFIX/lib /usr/local/cuda/lib # 链接 Conda 的 lib 到 cuda/lib
sudo ln -s $CONDA_PREFIX/lib /usr/local/cuda/lib64 # 链接 Conda 的 lib 到 cuda/lib64
sudo ln -s $CONDA_PREFIX/include /usr/local/cuda/include # 链接 Conda 的 include 到 cuda/include
sudo ln -s $CONDA_PREFIX/bin /usr/local/cuda/bin # 创建 bin 目录符号链接
sudo ln -s $CONDA_PREFIX/nvvm /usr/local/cuda/nvvm # NVIDIA 虚拟机的库，某些 CUDA 高级功能（如 JIT 编译）会依赖

# 训练
# export CUDA_VISIBLE_DEVICES=0
MAX_JOBS=2 sh /data/SAMPart3D/scripts/train.sh -g 1 -d sampart3d -c sampart3d-trainmlp-render16views -n ZaanseSchansBike -o ZaanseSchansBike

# 发生SyntaxError
pip uninstall networkx  # 先卸载确保干净
pip install --upgrade networkx

# 发生ModuleNotFoundError: No module named 'termcolor'
pip install termcolor

# 发生ModuleNotFoundError: No module named 'google'
conda install protobuf

# 发生ModuleNotFoundError: No module named 'narwhals'
conda install narwhals

# 如发生OSError: /data/miniconda3/envs/sampart3d/lib/python3.10/site-packages/torch_scatter/_version_cuda.so: undefined symbol: _ZN3c1017RegisterOperatorsD1Ev
python - <<EOF
import torch
print(f"PyTorch version: {torch.__version__}")
print(f"PyTorch CUDA version: {torch.version.cuda}") # CUDA version PyTorch was built with
print(f"Is CUDA available: {torch.cuda.is_available()}")
# 检查 PyTorch 使用的 C++ ABI
print(torch.__config__.show()) # 查找 -D_GLIBCXX_USE_CXX11_ABI=0 或 =1
EOF

# 发生OSError: Trainer: SAMPart3DDataset16Views: We couldn't connect to 'https://huggingface.co' to load the files, and couldn't find them in the cached files 
# 可以设置 hf mirror 来修复 huggingface.co 的网络连接问题。
export HF_ENDPOINT=https://hf-mirror.com 

# windows访问WSL，通过路径,可以使用vscode开发方便修改代码
\\wsl$\Ubuntu-24.04\data\SAMPart3D
# 如提示Using a slow image processor as use_fast is unset and a slow processor was saved with this model. use_fast=True will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with use_fast=False.
# 将from transformers import pipeline, SamModel修改为
from transformers import pipeline, SamModel, SamImageProcessor
# 将SAM_model = pipeline(xxx)修改为
def prepare_meta_data(self, data_path=None):
    # 显式创建图像处理器和模型
    processor = SamImageProcessor.from_pretrained("facebook/sam-vit-huge", use_fast=True)
    model = SamModel.from_pretrained("facebook/sam-vit-huge")
    
    # 创建pipeline时指定processor
    SAM_model = pipeline(
        "mask-generation",
        model=model,
        image_processor=processor,
        device=self.device
    )
    
    # ...其余代码保持不变...
  
# 判断CUDA头文件是否可被正确访问
python - <<EOF
import os
os.environ["CUDA_HOME"] = "/data/miniconda3/envs/sampart3d"  # 替换为你的路径
cassert_path = os.path.join(os.environ["CUDA_HOME"], "include", "cuda", "std", "cassert")

if os.path.exists(cassert_path):
    print(f"成功找到 CUDA 头文件！路径: {cassert_path}")
else:
    print(f"路径仍无效，请检查: {cassert_path}")
    print("手动验证文件是否存在: ", end="")
    os.system(f"ls -l {cassert_path}")
EOF

HoloPart

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45

conda activate sampart3d
# 
git clone https://github.com/VAST-AI-Research/HoloPart.git
cd HoloPart
# 安装依赖
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt

# 在Windowsd的D:\ProgramFiles\WSL\下创建文件夹HoloPart,并创建input_mesh文件夹
  HoloPart
  |-- input_mesh
# 连接windows下的文件夹到Linux下
ln -s "/mnt/d/ProgramFiles/WSL/HoloPart/input_mesh" /data/HoloPart/input_mesh

# 进入input_mesh
cd /data/HoloPart/input_mesh/

python - <<EOF
import trimesh
import numpy as np
import os

print(f"Python script CWD: {os.getcwd()}")
input_mesh_name = "AmsterdamPaleisopdeDam"

mesh_path = f"/data/SAMPart3D/mesh_root/{input_mesh_name}.glb"
mask_path = f"/data/SAMPart3D/exp/sampart3d/{input_mesh_name}/results/last/mesh_0.0.npy"
output_path = f"{input_mesh_name}_output.glb"

print(f"Attempting to load mesh from: {mesh_path}")
print(f"Attempting to load mask from: {mask_path}")

mesh = trimesh.load(mesh_path, force="mesh")
mask_npy = np.load(mask_path)
mesh_parts = []
for part_id in np.unique(mask_npy):
    mesh_part = mesh.submesh([mask_npy == part_id], append=True)
    mesh_parts.append(mesh_part)
mesh_parts = trimesh.Scene(mesh_parts).export(output_path)
EOF

# 将3D网格分解为完整的部分
python -m scripts.inference_holopart --mesh-input /data/HoloPart/input_mesh/AmsterdamPaleisopdeDam_output.glb

# 发生ModuleNotFoundError: No module named 'pymeshlab'
pip install pymeshlab

Conda 环境激活时自动设置

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90

# 找到你的 Conda 环境的激活脚本目录。对于名为 sampart3d 的环境，路径通常是：
/data/miniconda3/envs/sampart3d/etc/conda/activate.d/
# 如果这个目录不存在，请创建它：
mkdir -p /data/miniconda3/envs/sampart3d/etc/conda/activate.d/
# 在该目录下创建一个 .sh 脚本文件，例如 env_vars.sh:
vim /data/miniconda3/envs/sampart3d/etc/conda/activate.d/env_vars.sh
# 在文件中添加以下内容
#!/bin/bash
export OLD_CUDA_HOME="${CUDA_HOME}"
if [ -n "${CUDA_HOME}" ]; then
  export CUDA_HOME="${CONDA_PREFIX}:${CUDA_HOME}"
else
    export CUDA_HOME="${CONDA_PREFIX}"
fi
echo "Set CUDA_HOME to: $CUDA_HOME"

export OLD_LD_LIBRARY_PATH="${LD_LIBRARY_PATH}"
if [ -n "${LD_LIBRARY_PATH}" ]; then
  export LD_LIBRARY_PATH="/usr/lib/wsl/lib:/usr/local/cuda/lib64:${CONDA_PREFIX}/lib:${LD_LIBRARY_PATH}"
else
  export LD_LIBRARY_PATH="/usr/lib/wsl/lib:/usr/local/cuda/lib64:${CONDA_PREFIX}/lib"
fi
echo "Set LD_LIBRARY_PATH to: $LD_LIBRARY_PATH"

export OLD_LIBRARY_PATH="${LIBRARY_PATH}"
if [ -n "${LIBRARY_PATH}" ]; then
  export LIBRARY_PATH="/usr/lib/wsl/lib:/usr/local/cuda/lib64:${CONDA_PREFIX}/lib:${LIBRARY_PATH}"
else
  export LIBRARY_PATH="/usr/lib/wsl/lib:/usr/local/cuda/lib64:${CONDA_PREFIX}/lib"
fi
echo "Set LIBRARY_PATH to: $LIBRARY_PATH"

export OLD_CPLUS_INCLUDE_PATH="${CPLUS_INCLUDE_PATH}"
if [ -n "${CPLUS_INCLUDE_PATH}" ]; then
  export CPLUS_INCLUDE_PATH="${CONDA_PREFIX}/include:${CPLUS_INCLUDE_PATH}"
else
  export CPLUS_INCLUDE_PATH="${CONDA_PREFIX}/include"
fi
echo "Set CPLUS_INCLUDE_PATH to: $CPLUS_INCLUDE_PATH"

export OLD_C_INCLUDE_PATH="${C_INCLUDE_PATH}"
if [ -n "${C_INCLUDE_PATH}" ]; then
  export C_INCLUDE_PATH="${CONDA_PREFIX}/include:${C_INCLUDE_PATH}"
else
  export C_INCLUDE_PATH="${CONDA_PREFIX}/include"
fi

export HF_ENDPOINT=https://hf-mirror.com 
echo "Set C_INCLUDE_PATH to: $C_INCLUDE_PATH"

# export OLD_PATH="${PATH}"
# if [ -n "${PATH}" ]; then
#   export PATH="${CONDA_PREFIX}/bin:${PATH}"
# else
#   export PATH="${CONDA_PREFIX}/bin"
# fi
# echo "Set PATH to: $PATH"


# 还需要一个对应的反激活脚本来恢复原始值
# 创建反激活脚本目录：
mkdir -p /data/miniconda3/envs/sampart3d/etc/conda/deactivate.d/
# 创建反激活脚本文件 env_vars.sh:
vim /data/miniconda3/envs/sampart3d/etc/conda/deactivate.d/env_vars.sh
# 添加内容：
#!/bin/bash
export CUDA_HOME="${OLD_CUDA_HOME}"
unset OLD_CUDA_HOME
echo "set CUDA_HOME to: $CUDA_HOME"

export LD_LIBRARY_PATH="${OLD_LD_LIBRARY_PATH}"
unset OLD_LD_LIBRARY_PATH
echo "set LD_LIBRARY_PATH to: $LD_LIBRARY_PATH"

export LIBRARY_PATH="${OLD_LIBRARY_PATH}"
unset OLD_LIBRARY_PATH
echo "set LIBRARY_PATH to: $LIBRARY_PATH"

export CPLUS_INCLUDE_PATH="${OLD_CPLUS_INCLUDE_PATH}"
unset OLD_CPLUS_INCLUDE_PATH
echo "set CPLUS_INCLUDE_PATH to: $CPLUS_INCLUDE_PATH"

export C_INCLUDE_PATH="${OLD_C_INCLUDE_PATH}"
unset OLD_C_INCLUDE_PATH
echo "set C_INCLUDE_PATH to: $C_INCLUDE_PATH"

# export PATH="${OLD_PATH}"
# unset OLD_PATH
# echo "set PATH to: $PATH"
# 现在，每次你运行 conda activate sampart3d 时，CPLUS_INCLUDE_PATH等就会被设置。当你运行 conda deactivate 时，它会尝试恢复
CATALOG
    FEATURED TAGS
    Life Game Development Continuous Updates Poetry Read Movie Open Source Dreams Society and Science Market Server
    FRIENDS