search

wujcan / GIF-torch

20 stars 4 forks source link

Gradient are all NAN when I run GAT/GIN. #1

Closed Billy1900 closed 1 year ago

Billy1900 commented 1 year ago

When I run python3 main.py --dataset_name cora --target_model GAT --is_train_target_model True --exp Unlearn --method GIF --is_use_node_feature True --num_runs 5 --unlearn_task edge --unlearn_ratio 0.05 --iteration 100 --scale 500, the h_estimate will become NAN after several unlearn iterations.

python3 main.py --dataset_name cora --target_model GIN --is_train_target_model True --exp Unlearn --method GIF --is_use_node_feature True --num_runs 5 --unlearn_task edge --unlearn_ratio 0.05 --iteration 100 --scale 500 meets the same issue. Do you know why? @wujcan

wujcan commented 1 year ago

We find a large value of 'scale' coefficient can avoid this issue, for example, 10^6 on GIN and 10^5 on GAT.