手把手教ISBNET环境配置以及训练（遇到的问题） 全网最全

Ubuntu18.04 显卡：NVIDIA A40 驱动：525.147.05 cuda 11.3

环境配置

conda create -n isbnet python=3.7
conda activate isbnet

conda install pytorch==1.12.1 torchvision==0.13.1 cudatoolkit=11.3 -c pytorch
pip3 install spconv-cu113==2.1.25
pip3 install torch-scatter==2.0.9 -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
pip3 install -r requirements.txt

git clone https://github.com/Karbo123/segmentator.git

如果ubuntu18.04连接不了github，参考：http://t.csdnimg.cn/xEDV9

cd segmentator/csrc
mkdir build && cd build

cmake .. \
-DCMAKE_PREFIX_PATH=`python -c 'import torch;print(torch.utils.cmake_prefix_path)'` \
-DPYTHON_INCLUDE_DIR=$(python -c "from distutils.sysconfig import get_python_inc; print(get_python_inc())")  \
-DPYTHON_LIBRARY=$(python -c "import distutils.sysconfig as sysconfig; print(sysconfig.get_config_var('LIBDIR'))") \
-DCMAKE_INSTALL_PREFIX=`python -c 'from distutils.sysconfig import get_python_lib; print(get_python_lib())'` 

make && make install

这里问题比较多，先解决完问题再继续

问题：CMake Error at /usr/local/share/cmake-3.30/Modules/Internal/CMakeCUDAArchitecturesValidate.cmake:7 (message): CMAKE_CUDA_ARCHITECTURES must be non-empty if set.

原因： CMake 需要在启用 CUDA 时指定 CUDA 架构，
但 CMAKE_CUDA_ARCHITECTURES 变量没有被设置。
当 CMake 检测到 CUDA 时，如果没有指定架构，就会出现这个错误。

CMakeLists.txt在segmentator/csrc里面

解决办法：
1.使用 cuda 工具查看

cd /usr/local/cuda/samples/1_Utilities/deviceQuery
sudo make
./deviceQuery

2.修改 Cmakelist.txt set(CMAKE_CUDA_ARCHITECTURES 86)

cmake_minimum_required(VERSION 3.18 FATAL_ERROR)
set(CMAKE_CXX_STANDARD 14)
set(CMAKE_BUILD_TYPE Release)

project(libsegmentator)

# 设置 CUDA 架构，针对 NVIDIA A40 GPU
set(CMAKE_CUDA_ARCHITECTURES 86)

find_package(Torch REQUIRED)
find_package(PythonLibs REQUIRED)
find_library(TORCH_PYTHON_LIBRARY torch_python PATHS "${TORCH_INSTALL_PREFIX}/lib")

set(SOURCES_LIB segmentator.cpp)
add_library(libsegmentator SHARED ${SOURCES_LIB})
set_target_properties(libsegmentator PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(libsegmentator PROPERTIES PREFIX "")
target_link_libraries(libsegmentator ${TORCH_PYTHON_LIBRARY})
target_compile_definitions(libsegmentator PUBLIC TORCH_EXTENSION_NAME=libsegmentator)
target_compile_definitions(libsegmentator PUBLIC TORCH_API_INCLUDE_EXTENSION_H)
target_compile_definitions(libsegmentator PUBLIC ${TORCH_CXX_FLAGS})
target_include_directories(libsegmentator PUBLIC ${TORCH_INCLUDE_DIRS})
target_include_directories(libsegmentator PUBLIC ${PYTHON_INCLUDE_DIRS})

# 安装部分
set(BARE_PROJECT_NAME segmentator)
install(CODE "execute_process(                                        \
    COMMAND ${CMAKE_COMMAND} -E create_symlink                        \
    ${PROJECT_SOURCE_DIR}/../../${BARE_PROJECT_NAME}                  \
    ${CMAKE_INSTALL_PREFIX}/${BARE_PROJECT_NAME}                      \
    )"
)

问题：CMake Error at /usr/local/share/cmake-3.30/Modules/CMakeDetermineCompilerId.cmake:400 (file):file failed to open for writing (Permission denied):

原因：权限不够

解决办法：sudo cmake…

问题：`-- The CUDA compiler identification is unknown

CMake Error at /home/admin815/anaconda3/envs/isbnet/lib/python3.7/site-packages/torch/share/cmake/Caffe2/public/cuda.cmake:47 (enable_language):
The CMAKE_CUDA_COMPILER:

/usr/local/cuda/bin/nvcc-DCMAKE_CUDA_ARCHITECTURES=86

is not a full path to an existing compiler tool.

Tell CMake where to find the compiler by setting either the environment
variable “CUDACXX” or the CMake cache entry CMAKE_CUDA_COMPILER to the full
path to the compiler, or to the compiler name if it is in the PATH.`

原因：CMake 无法找到 CUDA 编译器 nvcc，原因是您提供的路径不正确或 nvcc 不在系统的 PATH 环境变量中。

解决办法：CMake 配置过程中直接指定 CMAKE_CUDA_COMPILER

cmake -DCMAKE_CUDA_COMPILER=/usr/local/cuda-11.3/bin/nvcc

有关cmake的报错已全部解决，接下来继续执行命令

sudo apt-get install libsparsehash-dev

cd ISBNet/isbnet/pointnet2
python3 setup.py bdist_wheel
cd ./dist
pip3 install <.whl>   <.whl>指的是dist文件夹的whl

cd ISBNet
python3 setup.py build_ext develop

问题：OSError: /home/admin815/anaconda3/envs/isbnet/lib/python3.7/site-packages/nvidia/cublas/lib/libcublas.so.11: symbol cublasLtGetStatusString version libcublasLt.so.11 not defined in file libcublasLt.so.11 with link time reference

原因：路径配置不对，无法找到文件libcublasLt.so.11

解决办法：修改配置环境变量的文件sh，使路径为libcublasLt.so.11所在路径，

sudo gedit ~/.bashrc

export LD_LIBRARY_PATH={libcublasLt.so.11所在路径}:$LD_LIBRARY_PATH
环境配置完成！

下面开始数据准备

本文使用s3dis数据集进行训练，所以自行准备数据集Stanford3dDataset_v1.2_Aligned_Version

数据集里有不少错误需要修改，不然后面会报错，请参考：http://t.csdnimg.cn/BPPZY

准备 superpoints
https://datasets.d2.mpi-inf.mpg.de/box2mask/segment_labels.tar.gz

格式如下：

注意！！！项目里面一定全部用绝对路径 ，后面很多问题说找不到文件，其实都是因为没有用绝对路径！！！！

cd ISBNet/dataset/s3dis
bash prepare_data.sh

接下来是训练

# Pretrain step
python3 tools/train.py configs/s3dis/isbnet_backbone_s3dis_area5.yaml --only_backbone  --exp_name default

# Train entire model
python3 tools/train.py configs/s3dis/isbnet_s3dis_area5.yaml --trainall  --exp_name default

问题： File “/home/admin815/PycharmProjects/ISBNet-master/isbnet/data/s3dis.py”, line 35, in get_filenames assert len(filenames) > 0, f"Empty {p}"AssertionError: Empty Area_1`

原因：isbnet/data/s3dis.py 代码有问题

filenames = glob(osp.join(self.data_root, "preprocess", p + "*" + self.suffix))

这里+"*" 原本是指取文件夹里面全部文件，但是不知道为什么他解析不出来当成了一个字符而已，导致preprocess文件夹里面的内容取不出来，于是重写代码

解决：
`# 指定文件夹路径
directory = ‘/home/admin815/PycharmProjects/ISBNet-master/dataset/s3dis/preprocess/’

        # 获取文件夹内的所有文件名
        filenames = os.listdir(directory)

        # 过滤掉非文件的项目（如子文件夹）
        filenames = [f for f in filenames if os.path.isfile(os.path.join(directory, f))]

        filenames = [os.path.join(directory, filenames) for filenames in filenames]
        # 对文件名进行排序
        filenames = sorted(filenames)`

问题：大概就是说superpoint…找不到，报错位置也是在isbnet/data/s3dis.py

在函数def load(self, filename):里
spp_filename = osp.join(self.data_root, “superpoints”, scan_id + “.pth”)

原因：scan_id = osp.basename(filename).replace(self.suffix, “”)没有执行成功

解决办法：修改configs/s3dis/isbnet_backbone_s3dis_area5.yaml
suffix:’_inst_nostuff.pth’ 修改为 ‘_inst_nostuff’

问题：RuntimeError: /io/build/temp.linux-x86_64-cpython-37/spconv/build/core_cc/src/csrc/sparse/all/SpconvOps/SpconvOps_get_indice_pairs.cc(65)not implemented for CPU ONLY build.

原因：这个错误表明你在构建或运行一个仅支持 GPU 的库或功能，但你的环境或配置是 CPU-only。spconv 是一个用于稀疏卷积的库，通常依赖于 GPU。

解决办法：pip install spconv-cu113

解决完之后终于跑起来了！！太不容易了

最后测试并可视化：

测试
python3 tools/test.py configs/s3dis/isbnet_s3dis_area5.yaml /home/admin815/PycharmProjects/ISBNet-master/work_dirs/s3dis/isbnet_backbone_s3dis_area5/default/latest.pth --out results/isbnet_s3dis_val

问题：visualization/vis_s3dis.py有个地方要改

xyz, rgb, semantic_label, instance_label = torch.load(
        f"{args.data_root}/{args.split}/{args.scene_name}_inst_nostuff.pth"
    )

改为

xyz, rgb, semantic_label, instance_label = torch.load(
        f"{args.data_root}/{args.split}/{args.scene_name}_inst_nostuff"
    )

可视化
python3 visualization/vis_s3dis.py --data_root /home/admin815/PycharmProjects/ISBNet-master/dataset/s3dis/ --scene_name Area_1_conferenceRoom_2 --prediction_path /home/admin815/PycharmProjects/ISBNet-master/results/isbnet_s3dis_val --task inst_pred

cd /home/admin815/PycharmProjects/ISBNet-master/visualization/pyviz3d
python -m http.server 6008

这里应该是我训练次数不够多，看不出实例分割（应该吧），训练30轮都要一天，不训练这么久了，赶紧试一下自己的数据集算了