Skip to content

SETUP.md

Latest commit

 

History

History
86 lines (70 loc) · 3.23 KB

SETUP.md

File metadata and controls

86 lines (70 loc) · 3.23 KB

Setup

This setup documentation serves as an introduction for extending and maintaining the ChronoKGE framework.

Project structure

Folder structure

The chrono_kge module consists of the following folders:

  • main: The main modules.
    • handler: Handler for framework state.
    • manager: Managers for framework execution.
    • parser: Parsers for framework parameter.
  • experiment: The experiment modules.
  • knowledge: Modules related to data handling and preprocessing.
    • augmentation: Augmentation modules.
    • chrono: Time processing modules.
    • graph: Knowledge graph modules.
  • model: The model modules.
    • kge: Complete network models.
      • deepml: KGE models using deep learning.
      • geometric: KGE-models using distance scoring.
      • semantic: KGE-models using similarity scoring.
    • module: Model modules and extensions.
      • calculus: Basic tensor modules.
      • embedding: Embedding modules.
      • pooling: Bilinear pooling modules.
      • regularizer: Regularization modules.
      • scoring: Scoring modules.
  • statistics: The statistics modules.
  • trainer: The trainer modules.
  • tuner: The tuner modules.
  • utils: Additional helper functions
    • vars: Constants and enumerations.
    • web: Web utilities.
  • visualization: Visualization tools.

KGE Families

  • Deep Learning (network): model.kge.deepml
  • Geometric (distance): model.kge.geometric
  • Semantic matching (similarity): model.kge.semantic

Extending ChronoKGE

Adding a new model

In the following, we describe how to add a new embedding model.

  1. Add a new model into model/kge/<family>(Learn more) by either:
    1. creating a new model which inherits PyTorch's base nn.Module
    2. extending an existing model, e.g., inheriting from kge_model.KGE_Model
  2. Register your model within the model/__model__.py module
    1. Add an entry into the REGISTER dictionary, where:
      1. key: a unique string identifier used to call your model
      2. value: the class type of your model
  3. Done!

Adding a new dataset

In the following, we describe how to add a new benchmark dataset.

  1. Save the new dataset under one of the following folders:
    1. static: Datasets with static facts (s,p,o)
    2. temporal: Datasets with temporal facts (s,p,o,t)
    3. synthetic: Datasets which are synthetically created
  2. Register a new dataset in the knowledge/knowledge_base.py module
    1. Define a new KnowledgeBase in the KnowledgeBases object
      1. Specify the KB properties according to the dataset:
        1. name: a unique string identifier (same name as data-folder)
        2. genre: the genre of the KB (static, temporal, synthetic)
        3. start_date: (optional) The start date of the KB
        4. num_days: (optional) The number of days covered by the KB
        5. gran: (optional) The time granularity of the KB
    2. Add the new KB enum into the ALL_KB list in the KnowledgeBases object
    3. Register the new KB in the appropriate function attributes:
      1. has_timestamps: KB contains time in form of timestamps
      2. has_indices: KB contains entity indices in datasets
      3. has_labels: KB contains entity labels in datasets
  3. Done!