dfilimon / discoflan

DiscoFlan

Overview: Refined and Unrefined models

This is Disrpt submission for our generative model. There are two models:

DiscoFlan (Unrefined)	DiscoFlan+Ref (Refined)
X	Our Submission
Generates classes outside of submission rules	Generates submission file ("task_preds.csv")
No analysis script in runner file	Runs task rel_eval.py to obtain official results.
Pass --refined False to runner.py	Pass --refined True to runner.py

a) Refined models will be considered for submission hence the evaluation script is also provided.

b) Unrefined models will not be considered for submission hence evaluation script is not provided.

Scripts and outputs

There are two way to run the model (synchronously or in the background):

1. submission_runner.sh trains the following models (refer to comments submission_runner.sh for other instructions)
2. submission_runner_nohup.sh does the same by spawning background processes.

Accuracy scores are automatically generated as output. Generated labels are stored in predictions.csv and the best model is saved in best.pt

Instructions

create conda environment

cd DiscoFlan
mkdir refined_runs
mkdir unrefined_runs
conda create --name flant5 python=3.10
conda activate flant5

Install dependencies:

pip install -r requirements.txt

Provide the correct arguments by modifying submission_runner.sh and THEN run submission_runner.sh (Set Arguments: $DATASET_NAME, $DATASETFOLDER)

bash submission_runner.sh

Misc Batchsize can be modified if necessary in runner.py. Cuda device id can be provided in submission_runner.sh.