In this phase of our assignment, we focused not only to explore but also to improve the DoDUO model variants of our selected paper (https://github.com/megagonlabs/doduo).
This repository consists mainly of 6 below notebooks:
-
DoDUO_Finetune_MSP_Inference.ipynb An implementation of finetuning the pretrained DoDUO (given) with Masked Column Prediction (MCP), along with its inference
-
T5_single_column_annotation.ipynb An implementation of single column annotation, same approach as mentioned in the paper, where the main architecture is based on the T5 model.
-
T5_single_colpair_columns_relationship.ipynb An implementation of single column relationship, same approach as mentioned in the paper, where the main architecture is based on the T5 model.
-
T5_single_column_MTL.ipynb An implementation of single column relationship & single column annotation, where we introduce a single unified T5 model training for both tasks though Multi Task Learning.
-
T5_tablewise_annotation.ipynb An implementation of multi-column (tablewise) annotation, same approach as mentioned in the paper, where the main architecture is based on the T5 model.
-
T5_tablewise_relationship.ipynb An implementation of multi-column (tablewise) relationship, same approach as mentioned in the paper, where the main architecture is based on the T5 model.
Results
- An overall graph with all the F1 validation score of the T5 model variants
Important Notes
- In all the respective notebooks, the official DoDUo repository is cloned and integrated via the "git clone" command.
- The requirements.txt file is generated from the Coolab-notebook envrironment of the notebooks.
- All the python scripts (*.py) are generated and downloaded from the respective notebooks.
- All the notebooks are executable, therefore it is easy to reproduce the results with no limitations at all.