This repo is designed for implementing various embedding algorithms for knowledge graphs. It is under development (for my thesis).
((no GPU support (yet...)))
Before one does anything, it is required to produce mappings from raw triplets to ids. To produce this, one should run the following code for their given training triplets...
python triplets.py triplet_path
This will produce 2 json files, one called entity2id and the other relationship2id, which contain the aforementioned mappings! Those will be located in the same directory as the given triplet file.
To train an embedding:
python main.py saved_embedding_folder \
--train_data=train_path --val_data=valid_path \
--lr=0.01 --train_batch_size=128 --epochs=50 --patience=20 \
--algorithm=transe --emb_dim=100
python main.py --help (Will show only training arguments, without the specific class arguments)
To get all possible arguments for training, we have two types. First, we have the general training arguments (i.e --train_data). Secondly, having defined the --algorithm argument, one can fix all the arguments that define the class Model inside the algorithms folder. If one wants to train a TransE [1] model, just add --algorithm=transe, In general, one can define a personal type of embedding algorithm, by simply creating a Python script (for example experiment.py) and then add --algorithm=experiment .
To test an embedding algorithm with metrics like hits@10 or Mean rank.
python test.py save_path algorithm mean_rank
save_path is the path of the saved model. The test results will be appended in a file called test.txt located in the directory of the saved model.
Setting --filtering==True, one can get the results of the rankings with filtering.
To have the correct behaviour, one should define the arguments below:
- --test_data
- --train_data
- --valid_data
(Train and valid data are only used for the filtering of scores.)
To get all args:
python test.py --help
This code does NOT guarantee the accurate implementation of the implemented algorithms, like TransE, TransH [2] and RotatE [3]...
[3] Zhiqing Sun et al., "RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space"