search

DIMESTeam / DIMES

MIT License
39 stars 6 forks source link

Installation Issues #6

Closed xyfffff closed 11 months ago

xyfffff commented 11 months ago

Hi Ruizhong,

I am encountering persistent installation issues with several packages and am reaching out for assistance. I have attempted various combinations of versions for pytorch, scatter, sparse, cluster, and geometric, but keep running into errors.

The errors vary: sometimes I encounter an issue with LIB_PATH = os.getenv('LD_LIBRARY_PATH').split(':')[-1], other times I face the error torch_sparse/_convert_cuda.so: undefined symbol: _ZNSt15__exception_ptr13exception_ptr9_M_addrefEv, and at times I run into a segmentation fault as follows:

./tsp500_test_G.sh: line 24: 3291898 Segmentation fault (core dumped) python3 ./test_AS_G.py --seed 123456789 --device cuda:0 --n_nodes 500 --knn_k 50 --outer_opt AdamW --outer_opt_lr 0.001 --outer_opt_wd 1e-5 --inner_opt AdamW --inner_opt_lr 0.01 --inner_opt_wd 0.0001 --net_units 32 --net_act silu --emb_agg mean --emb_depth 12 --par_depth 3 --te_net 120 --te_range_l 0 --te_range_r 128 --te_batch_size 10 --te_tune_steps 0 --te_tune_sample_size 500 --te_sample_size 500 --save_name $1.

I am unclear about the root cause of these issues. It could be due to my NVIDIA CUDA version (12.1), my GCC version (11.4), or my operating system (Ubuntu 22.04.2 LTS).

I am unable to install libraries like scatter. I am wondering if these libraries are particularly sensitive to hardware configurations? Do you have any suggestions for resolving these issues or any alternative approaches that I might consider?

Thank you for your time and assistance.

Best regards, Yifan

q-rz commented 11 months ago

The segmentation fault seems due to the failed installation of our TSP sampling library. As far as I am aware of, there are two possible causes of your installation issue:

  1. Our TSP sampling library depends on a .so file in the torch_scatter library. You will have to properly install torch_scatter before installing our TSP sampling library.
  2. The installer torch_sampling/setup.py might fail to find the path of that .so file. If so, please search for the file manually and add its path to torch_sampling/setup.py. Note that there can be two similar .so files, one for CPU and the other for GPU. Please use the GPU version.

If you have any questions, please feel free to let me know.