解决报错ImportError: unique_cuda.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN6caffe28T

最近在调试有关pytorch-geometric包的代码的时候遇到了这个错误，具体报错如下所示：

ImportError: /home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_sparse/unique_cuda.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN6caffe28TypeMeta21_typeMetaDataInstanceIN3c108BFloat16EEEPKNS_6detail12TypeMetaDataEv
(SGE) amax@admin:~/XJY/Semantic-Graph/Semantic-Graph/SGE-GNN$ python train_cross-modal.py --data_dir /home/amax/XJY/Semantic-Graph/Semantic-Graph/SGE/data_process/pascal/ --model_dir /home/amax/XJY/Semantic-Graph/Semantic-Graph/SGE/data_process/pascal
Traceback (most recent call last):
  File "train_cross-modal.py", line 5, in 
    from GNN.model import *
  File "/home/amax/XJY/Semantic-Graph/Semantic-Graph/SGE-GNN/GNN/model.py", line 1, in 
    from .conv import *
  File "/home/amax/XJY/Semantic-Graph/Semantic-Graph/SGE-GNN/GNN/conv.py", line 5, in 
    from torch_geometric.nn import GCNConv, GATConv
  File "/home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_geometric/__init__.py", line 2, in 
    import torch_geometric.nn
  File "/home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_geometric/nn/__init__.py", line 2, in 
    from .data_parallel import DataParallel
  File "/home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_geometric/nn/data_parallel.py", line 5, in 
    from torch_geometric.data import Batch
  File "/home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_geometric/data/__init__.py", line 1, in 
    from .data import Data
  File "/home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_geometric/data/data.py", line 7, in 
    from torch_sparse import coalesce
  File "/home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_sparse/__init__.py", line 2, in 
    from .coalesce import coalesce
  File "/home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_sparse/coalesce.py", line 4, in 
    from .utils.unique import unique
  File "/home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_sparse/utils/unique.py", line 5, in 
    import torch_sparse.unique_cuda
ImportError: /home/amax/.conda/envs/SGE/lib/python3.7/site-packages/torch_sparse/unique_cuda.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN6caffe28TypeMeta21_typeMetaDataInstanceIN3c108BFloat16EEEPKNS_6detail12TypeMetaDataEv

可以看到这个错误最终是指向了pytorch-sparse这个包，我们知道pytorch-geometric这个包主要有三个依赖包分别是：

torch-cluster
torch-scatter
torch-sparse

解决方案

因为我之前安装这三个包都是直接使用pip命令安装，问题好像出在使用pip安装即是安装成功了也会出现编译错误？于是我跑到了github把三个包重新下载了一遍，再用如下命令进行编译：

python setup.py install

下面给出几个包的下载地址：

torch-cluster	https://github.com/rusty1s/pytorch_cluster/releases
torch-scatter	https://github.com/rusty1s/pytorch_scatter/releases
torch-sparse	https://github.com/rusty1s/pytorch_sparse/releases