Edward-Sun
commented
4 years ago Hi Donnie,
wn18 has more entities than fb15k, so it takes more GPU memory. To reproduce the results on wn18, the recommended command is in best_config.sh: bash run.sh train RotatE wn18 0 0 512 1024 500 12.0 0.5 0.0001 80000 8 -de
It's equivalent to: CUDA_VISIBLE_DEVICES=$GPU_DEVICE python -u codes/run.py --do_train \ --cuda \ --do_valid \ --do_test \ --data_path data/wn18 \ --model RotatE \ -n 1024 -b 512 -d 500 \ -g 12.0 -a 0.5 -adv \ -lr 0.0001 --max_steps 80000 \ -save $SAVE --test_batch_size 8 \ -de
dear bro, very lucky to be able to read such a good paper, and open source, when I run the program, some errors occurred, the same server, when I use the data set FB15K, he is working, when I changed to wn18 , Then RuntimeError: CUDA out of memory
my code： CUDA_VISIBLE_DEVICES=1 python -u codes/run.py --do_train --cuda --do_valid --do_test --data_path data/wn18 --model RotatE -n 256 -b 256 -d 1000 -g 24.0 -a 1.0 -adv -lr 0.0001 --max_steps 80000 -save models/RotatE_wn18_0 --test_batch_size 16 -de -b{64,128,256,512}I have tried using these values，I also asked for help