Skip to content

SALT-NLP/IDBR

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

30 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Information-Disentanglement-Based-Regularization(IDBR)

This repo contains codes for the following paper:

Yufan Huang*, Yanzhe Zhang*, Jiaao Chen, Xuezhi Wang, Diyi Yang: Continual Learning for Text Classification with Information Disentanglement Based Regularization, NAACL 2021.

If you would like to refer to it, please cite the paper mentioned above.

Getting Started

These instructions will get you running the codes of IDBR.

Requirements

  • Python 3.8.5
  • Pytorch 1.4.0
  • transformers 3.5.1
  • tqdm, sklearn, numpy, pandas

Detailed env is included in ./package-list.txt.

Code Structure

|_src/
      |_read_data.py --> Codes for reading and processing datasets
      |_preprocess.py --> Preprocess datasets
      |_model.py --> Codes for baseline and IDBR model
      |_finetune.py --> Codes for finetune Baseline
      |_naivereplay.py --> Codes for naivereplay Baseline
      |_multitasklearning.py --> Codes for multitasklearning Baseline
      |_train.py --> Codes for IDBR
|_data/
      |_ag/
      |_amazon/
      |_dbpedia/
      |_yelp/
      |_yahoo/
      |_data/

All folders under ./data will be generated automatically in the "Downloading and Pre-processing the data" step.

Downloading and Pre-processing the data

We used the data provided by LAMOL. You can find the data from link to data. Please download it and put it into the data folder. Then uncompress and pre-process the data:

mkdir data
cd ./data
tar -xvzf LAMOL.tar.gz
cd ../src
python preprocess.py

Training models in Setting (Sampled)

Note that in the following exps, default epoch numbers should be set to 4.

We prune some of them to a smaller number due to certain tasks are easy to overfit.

Finetune

We use ./src/finetune.py to train the Finetune Baseline model:

# Example for length-3 task sequence
python finetune.py --tasks ag yelp yahoo --epochs 4 3 2   

# Example for length-5 task sequence
python finetune.py --tasks ag yelp amazon yahoo dbpedia --epochs 4 3 3 2 1   

Naive Replay

We use ./src/naivereplay.py to train the Naive Replay Baseline model:

# Example for length-3 task sequence
python naivereplay.py --tasks ag yelp yahoo --epochs 4 3 2   

# Example for length-5 task sequence
python naivereplay.py --tasks ag yelp amazon yahoo dbpedia --epochs 4 3 3 2 1

Regularization

We use ./src/train.py to train the Regularization Baseline model:

# Example for length-3 task sequence
python train.py --tasks ag yelp yahoo --epochs 4 3 2 --reg True --reggen 0.5 --regspe 0.5 --kmeans True --tskcoe 0.0

# Example for length-5 task sequence
python train.py --tasks ag yelp amazon yahoo dbpedia --epochs 4 3 3 2 1 --reg True --reggen 0.5 --regspe 0.5 --kmeans True --tskcoe 0.0

Information-Disentanglement-Based-Regularization

While we set reggen to 0.5, we select best regspe from {0.3, 0.4, 0.5}.

We use ./src/train.py to train the IDBR model:

# Example for length-3 task sequence
python train.py --tasks ag yelp yahoo --epochs 4 3 2 --disen True --reg True --reggen 0.5 --regspe 0.3 --kmeans True


# Example for length-5 task sequence
python train.py --tasks ag yelp amazon yahoo dbpedia --epochs 4 3 3 2 1 --disen True --reg True --reggen 0.5 --regspe 0.3 --kmeans True

Multitask Learning

We use ./src/multitasklearning.py to train the multitask-learning model:

# Multitask Learning
python multitasklearning.py --tasks ag yelp yahoo

Training models in Setting (Full)

We use ./src/train.py to train the IDBR model:

python train.py --tasks ag yelp amazon yahoo dbpedia --epochs 1 1 1 1 1 --disen True --reg True --reggen 0.5 --regspe 0.5 --kmeans True --n-labeled -1 --n-val 500

Questions

If you have any questions, please contact Yanzhe Zhang via z_yanzhe AT gatech.edu

About

Codes for the paper: "Continual Learning for Text Classification with Information Disentanglement Based Regularization"

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages