Persona Aware Response Generation

Getting started : Preparing data

• Install all requirements listed in requirements.txt.
• Check arguments in data/save2json.py. Prepare data.

cd data
python save2json.py --subset train --encdec bart --pretrained_name facebook/bart-base --encrep first
python save2json.py --subset valid --encdec bart --pretrained_name facebook/bart-base --encrep first
cd ..

How to train?

• Check all arguments needed in main.py, then run it

python3 main.py --save_name models/trialRun --batch_size 128 --encdec bart --pretrained_name facebook/bart-base --encrep first

• It will save the model to models/trialRun.

How to generate?

• Check all arguments needed in evaluate.py, these aer generation parameters.

You can read on controlling sampling methods in How to Generate

• It will generate responses from the save and save them comprehensively in generation_results/ in a json, where you will get corresponding scores - per sentence and averaged.

To-do list

• Adj for other types of graphs
• T5
• Metrics and evaluate script.
• update dataloader fields to match model forward
• modify main.py
• Save only trainable parameters?
• Flexibility to switch Graph types : adj + args