Make dgl-ke deterministic

[fburdet]

Dear all,

I'm trying to make dglk-ke deterministic (ie launch 2x the same dglke_train command, calculate the scores for the same edges and get exactly the same score). That would sound like more than a nice to have...

I tried some of the tips in https://github.com/awslabs/dgl-ke/issues/76 that is,

• add --num_thread 1 --num_proc 1 in the dglke_train command
• modify .local/lib/python3.8/site-packages/torch/backends/cuda/init.py that seems to load cuda and added torch.use_deterministic_algorithms(True)

but it doesn't work. Any easier way to make the command deterministic?

Thanks in advance for your help!

[classicsong]

The training is also random (edge sampling)

[fburdet]

Thanks for the answer, @classicsong !

Any way to also set the seed for the edge sampling?

[classicsong]

You can try set both numpy and torch seeds.