Firstly, the original edges pre-split into train/val/test set and
negative_sampler.GlobalUniform in train/val Dataloaders is used while training the model.
Now since these sampler edges change, how can I be sure I am evaluating on the same trained positive and negative edges for best training accuracy from the saved best model? Does setting seeds just as in training setting enough to recreate the same train positive and negative edges from the dataloader again?
So, after we train the link prediction model, I can pass the entire original graph at the inference stage and get the node representations.
Now, how to evaluate model on test set edges? I have the positive test set edges and similarly positive train/val edges. Since, I used Dataloader and Sampler to generate negative edges in the training setup, what would be the way to proceed here for test setup?