| ---
|
| language: en
|
| tags:
|
| - knowledge-graph
|
| - link-prediction
|
| - ComplEx
|
| - CoDEx
|
| - LibKGE
|
| - wikidata
|
| datasets:
|
| - CoDEx-S
|
| metrics:
|
| - mrr
|
| - hits@1
|
| - hits@10
|
| ---
|
|
|
| # CoDEx-S ComplEx — Winner Model
|
|
|
| Knowledge graph link prediction on **CoDEx-S** using **ComplEx** embeddings,
|
| trained with the [LibKGE](https://github.com/uma-pi1/kge) framework.
|
| Reproduces and slightly improves results from the
|
| [CoDEx paper (EMNLP 2020)](https://arxiv.org/pdf/2009.07810.pdf).
|
|
|
| ## Results (Validation Set — Filtered with Test)
|
|
|
| | Metric | This Model | Paper |
|
| |---------|-----------|-------|
|
| | MRR | 0.474 | 0.465 |
|
| | Hits@1 | 0.377 | 0.372 |
|
| | Hits@3 | 0.522 | 0.504 |
|
| | Hits@10 | 0.664 | 0.646 |
|
|
|
| Training stopped early at epoch **345** via early stopping.
|
|
|
| ## Dataset — CoDEx-S
|
|
|
| | | Count |
|
| |-|-------|
|
| | Entities | 2,034 |
|
| | Relations | 42 |
|
| | Train triples | 32,888 |
|
| | Valid triples | 1,827 |
|
| | Test triples | 1,828 |
|
|
|
| ## Hyperparameters
|
|
|
| | Parameter | Value |
|
| |-----------|-------|
|
| | Embedding dim | 512 |
|
| | Optimizer | Adam |
|
| | Learning rate | 0.000339 |
|
| | Batch size | 1024 |
|
| | Max epochs | 400 |
|
| | Training type | 1vsAll |
|
| | Loss | KL divergence |
|
| | LR scheduler | ReduceLROnPlateau |
|
| | Entity dropout | 0.079 |
|
| | Relation dropout | 0.056 |
|
|
|
| ## Load in Your App
|
|
|
| ```python
|
| import sys
|
| sys.path.insert(0, r"C:/path/to/codex/kge")
|
|
|
| from huggingface_hub import hf_hub_download
|
| from kge.model import KgeModel
|
| from kge.util.io import load_checkpoint
|
| import torch
|
|
|
| # Download from Hugging Face
|
| path = hf_hub_download(
|
| repo_id="aaryaupadhya20/codex-s-complex-winner",
|
| filename="winner_model.pt"
|
| )
|
|
|
| # Load model
|
| checkpoint = load_checkpoint(path, device="cpu")
|
| winner_model = KgeModel.create_from(checkpoint)
|
| winner_model.eval()
|
|
|
| print("winner_model ready!")
|
|
|
| # Score a triple using entity/relation integer indices
|
| s = torch.tensor([0]) # head entity index
|
| p = torch.tensor([1]) # relation index
|
| o = torch.tensor([2]) # tail entity index
|
|
|
| score = winner_model.score_spo(s, p, o, direction="o")
|
| print("Score:", score.item())
|
| ```
|
|
|
| ## Citation
|
|
|
| ```bibtex
|
| @inproceedings{safavi-koutra-2020-codex,
|
| title = "CoDEx: A Comprehensive Knowledge Graph Completion Benchmark",
|
| author = "Safavi, Tara and Koutra, Danai",
|
| booktitle = "Proceedings of EMNLP 2020",
|
| year = "2020",
|
| url = "https://arxiv.org/pdf/2009.07810.pdf"
|
| }
|
| ```
|
|
|
