python安装torch-cluster、torch-scatter、torch-sparse和torch-geometric | torch_geometric详细安装教程

1.检查CUDA版本

【方法1】用nvidia-smi

已装cuda，用nvidia-smi 或 nvcc -V 查看

【方法2】用torch

已装torch，用torch代码打印

import torch
print(torch.__version__)                # 查看pytorch安装的版本号
print(torch.cuda.is_available())        # 查看cuda是否可用。True为可用，即是gpu版本pytorch
print(torch.cuda.get_device_name(0))    # 返回GPU型号
print(torch.cuda.device_count())        # 返回可以用的cuda（GPU）数量，0代表一个
print(torch.version.cuda)               # 查看cuda的版本

2.安装torch（已装请跳过）

pip安装请看： https://download.pytorch.org/whl/torch_stable.html
找到自己系统环境+cuda版本+py版本 对应的torch轮子
我的是cu117/torch-1.13.1%2Bcu117-cp310-cp310-linux_x86_64.whl

• 【安装方法】 下载安装（Linux可以用wget安装 / windows可以点击下载到本地）
  pip install whl的完整路径

pip install https://download.pytorch.org/whl/torch_stable.htmlcu117/torch-1.13.1%2Bcu117-cp310-cp310-linux_x86_64.whl

3.安装torch-geometric

去到网站：https://pytorch-geometric.com/whl/index.html

• 【安装方法1】 ：pip install -f 指定网址(带torch和cuda版本的网址)
  逐个包指定网址，也可以全部的包，一起指定网址

  pip install torch-scatter torch-sparse torch-spline-conv torch-cluster torch-geometric -f https://pytorch-geometric.com/whl/torch-1.13.1%2Bcu117.html
  

• 【安装方法2】 ：下载后，再逐个pip（按顺序）
  • 先 torch_scatter

  pip install torch_scatter-2.1.0+pt113cu117-cp310-cp310-linux_x86_64.whl
  

  • 再 torch_sparse

  pip install torch_sparse-0.6.17+pt113cu117-cp310-cp310-linux_x86_64.whl
  

  • 再到 torch_cluster

  pip install torch_cluster-1.6.1+pt113cu117-cp310-cp310-linux_x86_64.whl
  

  • 然后到 torch_cluster

  pip install torch_spline_conv-1.2.1+pt113cu117-cp310-cp310-linux_x86_64.whl
  

  • 最后到 torch_geometric（需关闭机场）

  pip install torch_geometric==2.0.4 -i https://pypi.doubanio.com/simple
  或者
  pip install torch_geometric -i https://mirrors.aliyun.com/pypi/simple
  或者
  pip install torch_geometric -i https://pypi.tuna.tsinghua.edu.cn/simple
  

• 检验是否安装成功
  • import torch_geometric 一下未报错即可
    

4.常见错误

错误1. RuntimeError

RuntimeError: Detected that PyTorch and torch_scatter were compiled with different CUDA versions. PyTorch has CUDA version 10.2 and torch_scatter has CUDA version 11.3. Please reinstall the torch_scatter that matches your PyTorch install.

• 【解决方案】：检查torch版本 + CUDA版本 + py版本，和 torch-geometric相关包的版本，是否一致。

.

错误2. An issue occurred xx

UserWarning: An issue occurred while importing ‘torch-scatter’. xxxxxxxx torch_scatter/_scatter_cuda.so: undefined symbol: _ZNK3c107SymBool10guard_boolEPKcl

UserWarning: An issue occurred while importing ‘torch-cluster’. xxxxxxxx
torch_cluster/_grid_cuda.so: undefined symbol: _ZNK3c107SymBool10guard_boolEPKcl

UserWarning: An issue occurred while importing 'torch-spline-conv’.
torch_spline_conv/_basis_cuda.so: undefined symbol: _ZNK3c107SymBool10guard_boolEPKcl

UserWarning: An issue occurred while importing ‘torch-sparse’.
torch_sparse/_diag_cuda.so: undefined symbol: _ZN3c106detail19maybe_wrap_dim_slowIlEET_S2_S2_b

• 【解决方案】：卸载torch-geometric相关4个包，再按顺序 依次安装

  pip uninstall torch-cluster
  pip uninstall torch-scatter
  pip uninstall torch-sparse
  pip uninstall torch-geometric