Controlling the random seed in train_test_split_edges for reproducibility

[sbonner0]

Is it possible to control the seed used for train_test_split_edges to consistently return the same dataset split? Conversely, would it also be possible to set the seed to get different dataset splits?

I guess setting the torch seed might fix this? However I am concerned setting it might affect other functions such as negative_sampling which would be better if it remained random.

[rusty1s]

You may want to call this function via:

torch.manual_seed(12345)
train_test_split_edges()
torch.manual_seed(time())

but I'm actually not sure if that's the best way how you want to do it in PyTorch.

[sbonner0]

Great thanks - maybe adding a seed argument for this function might be something that would be useful? Having some control over the dataset splits would be really great.

[rusty1s]

I'm not sure. PyTorch does not seem to utilize a seed argument anywhere in its codebase (please correct me if I'm wrong), and I generally try match the design principles of PyTorch quite closely. IMO, if you want controllable and deterministic experiments, it's best to set a global seed at the start of your script.

[sbonner0]

Perfectly valid thanks!

Am I right in assuming that the negative_sampling function would also be affected by setting the seed? This might not be desirable for training as it would be better to have a different set of negatives for each epoch.

[rusty1s]

You will still have a different set of negatives for each epoch, as long as you do not re-set the seed every time before calling negative_sampling.