Skip to content

Latest commit

 

History

History
29 lines (17 loc) · 808 Bytes

README.md

File metadata and controls

29 lines (17 loc) · 808 Bytes

Pretraining

pretrain Dink-Net on your own data.

complie kmcuda

git clone https://github.com/src-d/kmcuda

cd kmcuda/src

cmake -DCMAKE_BUILD_TYPE=Release . && make

copy the pretrained model parameters libKMCUDA.so to ./pretrain

run codes with comments

python pretrain.py --device cuda:0 --dataset cora --hid_units 512 --lr 1e-3 --epochs 200 --seed 47 --km_cuda

python pretrain.py --device cuda:0 --dataset citeseer --hid_units 1536 --lr 5e-4 --epochs 200 --km_cuda

python pretrain.py --device cuda:0 --dataset amazon_photo --hid_units 512 --lr 1e-3 --epochs 500 --seed 9 --km_cuda

python pretrain.py --device cuda:0 --dataset ogbn_arxiv --hid_units 1500 --encoder_layer 3 --lr 1e-4 --epochs 10 --seed 6 --km_cuda 

copy the pretrained model parameters DinkNet_xxx.pt to ./models