Last edited: week of Feb 12th 2024.
#TODO: rewrite for public consumption, see TODO subsections
Removes all output directories created in failed training runs, i.e. those which do not contain checkpoints or hyperparameters
Contains old code for reference, mostly methods for data module creation and batch collation.
- data/yago3-10/
- test/train/valid/valid_tiny.del: train, test, and validation splits as tab-delimited files. Each line in the file is a unique triple of {subject_idx, rel_idx, object_idx}.
- entity/relation_mentions.del: links ids with their canonical mentions. first column entity/relation id, second column corresponding entity/rel mention.
- describe contents / function of [and script creating] query_solutions.pckl for yago3-10 and wikidata
- describe data/wikidataXX-XX
- add credit for datasets and data preprocessing/train/val/test splits
- hydra.yaml: defines (default) hydra behaviour; default options for multirun execution in sweep
- what actually was in here?? @Daniel
- add default options in sweep if used.
The config folder where hydra goes to check all available config files and file groups.
- config.yaml: defines all (default) model/execution parameters which shouldn't be hardcoded; input to final config of run.
- default folders for config packages