search

Tiiiger / SGC

official implementation for the paper "Simplifying Graph Convolutional Networks"
MIT License
831 stars 145 forks source link

Bugfix: reference cuda arg in call to load_reddit_data #31

Closed kelseyball closed 3 years ago

kelseyball commented 3 years ago

The cuda argument is always True within the load_reddit_data function, leading to a torch error if cuda is not available. This change allows the user to run the reddit dataset experiments without cuda.

Test plan:

Running reddit evaluation without cuda produces this error:

python reddit.py --inductive --test
Traceback (most recent call last):
  File "reddit.py", line 38, in <module>
    adj, train_adj, features, labels, idx_train, idx_val, idx_test = load_reddit_data(args.normalization)
  File "/Users/ktb477/SGC/utils.py", line 127, in load_reddit_data
    adj = adj.cuda()
  File "/Users/ktb477/.pyenv/versions/SGC/lib/python3.8/site-packages/torch/cuda/__init__.py", line 164, in _lazy_init
    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

With the proposed change:

python reddit.py --inductive --test
Finished data loading.
Total Time: 22.9538s, Test F1: 0.9501
felixgwu commented 3 years ago

Hi @kelseyball, thanks a lot for the fix!