Lab: Training Graph Embeddings
Lab 14: Training Graph Embeddings
Topics
Training knowledge graph embeddings with TorchKGE.
Classes and methods
The following TorchKGE class from the previous lab remains central:
- Model - contains the embeddings (entity and relation vectors) for the KG
In addition, we will also use:
- KG - contains the knowledge graph (KG)
More classes will be suggested below.
Tasks
Pre-trained models:
- Choose a KG and TransE model you want to work with. It should have a pre-trained model available. (Freebase FB15k is still a good choice, see the note below if you want to use Wikidata.)
- Load the pre-trained model (you do no need the KG yet). and evaluate it using the examples given here: https://torchkge.readthedocs.io/en/latest/tutorials/evaluation.html .
- Extra: You can also evaluate the model on relation prediction but, the way TransE is pre-trained, it is awful on this task.
Note: The Wikidata dataset returns two graphs. They are not train/test, but the dataset with and without additional attributes. Start with the one without attributes. You need to split it into train/validation/test yourself using KG.split_kg().
Train your own:
- Load the corresponding KG using a dataset loader.
- Run the Shortest training example, but use a much lower value for epoch (for example 200).
- Take note of the evaluation metrics and final loss, and re-run the example using different numbers of epochs. What happens when you increase the number?
- Also run the Simplest training example. Use the documentation to make sure you have an idea of what the different parts of the algorithm do.
Train with early stopping:
- Run the Training with Ignite example. Use the documentation to make sure you have an idea of what the different parts of the algorithm do. How do the results compare with your exploration of different epoch values?
If You Have More Time
- Try this out on the other models supported by TorchKGE, both other TransX models and a deep model (ConvKB).
- Try it out with different datasets, for example one you create youreself using SPARQL queries on an open KG.