This extends the git repo KnowledgeGraphEmbedding to make it interactive by creating a live embedding visualization that allows for adding new entities during training. For more information on how the actual embedding works click the link.
Some changes were made to the existing code of KnowledgeGraphEmbedding. These are i.a. fixing seeds and modifying some code parts such that it is possible to store embeddings during the training.
Most notably the module run.py was modified for the main purpose of this extended repo to visualize embeddings, change them via drag-and-drop and add new entities - all live during training.
Additionally, the run.py and model.py modules were changed to allow for a new optional feature: Pretraining.
Pretraining allows for the training of the semantics of new relations with the help of a mock dataset that should be saved as pretrain.txt under the corresponding data directory.
An example for such a file can be looked up in the data/Yamanishi_with_Similarity_Information directory.
While the regular training phase uses the data in train.txt, pretraining precedes regular training and only uses the data in pretrain.txt.
Pretraining lasts only a small number of epochs before regular training starts.
After Pretraining is completed the embeddings of the new relations will be frozen, so that the learned semantics don't get corrupted.
For detailed information on how to use the run.py for KGE training without visualization refer to KnowledgeGraphEmbedding.
You don't have to use the KnowledgeGraphEmbedding repo, since this repo (InteractiveKGE) provides the same features and even more.
Like training without visualization, training with visualization can be executed on any of the datasets in the data/ directory.
To use the Interactive Visualization set the --visualize argument. Example:
python -u codes/run.py
--visualize
--do_train --do_valid --do_test
--data_path data/countries_neighb_UsaSpaDen
--model TransE
--valid_steps 10
--save_checkpoint_steps 1
-n 2 -b 8 -d 2
-g 1.0 -a 1.0 -adv
-lr 0.1 --max_steps 100
-save models/countries_neighb_UsaSpaDen_TransE --test_batch_size 8
This countries_neighb_UsaSpaDen dataset is a simplified version of the countries_S1 dataset, that is filtered for the countries USA, Spain and Denmark and all of their neighbors.
You can add your own datasets and run the Interactive KGE on them. Therefore, add your datasets to the data/ directory,
but keep the structure and format same as the other datasets. Visualization caps the dimensionality -d at 2.
To use Pretraining set the --do_pretrain argument. Example:
CUDA_VISIBLE_DEVICES=0 python -u codes/run.py
--do_pretrain
--do_train --do_valid --do_test
--cuda
...
...
...
Pretraining cannot be used alongside the Interactive Visualization. The --visualize argument will override --do_pretrain.
Implemented Features:
- Tracking and Plotting the embeddings live while KGE training is running in the background
- Change embeddings during training via drag-and-drop
- Pretraining (only accessible if training without visualization)
In Progress Features:
- By clicking into the embedding space, you pause the training and create a new entity at the clicked coordinates. Choose the relations of the new entity in the dropdown menu