Variational Information Bottleneck for Effective Low-resource Finetuning

Python requirements

This code is tested on:

• Python 3.7.7
• transformers 2.8.0
• pytorch 1.2.0

Donwloading the datasets

You can download the datasets from the following paths, which are from previously published papers, and put them all in "data/datasets/"

• download the GLUE data from https://github.com/nyu-mll/GLUE-baselines, this include SNLI, MNLI, RTE, MRPC, STS-B
• for IMDB dataset, you can download it from https://github.com/allenai/sledgehammer
• for 12 used target NLI datasets, you can download them from https://github.com/azpoliak/robust-nli/tree/master/data
• we download the yelp datasets from https://github.com/jiacheng-xu/vmf\_vae\_nlp/tree/master/data

Parameters in the code

• num_samples Specifies the number of samples in case of running models on the subsampled datasets
• ib_dim Specifies the bottleneck size
• ib If this option is set, runs the VIBERT model
• deteministic If this option is set, runs the VIBERT model with beta=0
• beta Specifies the weight for the compression loss
• mixout defines the mixout propability
• weight_decay defines the weight for weight_decay regularization
• to run the model on the subsampled datasets, add --sample_train option and specify the number of samples with --num_samples N , where N is the number of samples.

Usage

We provide the following sample scripts. We using these scripts, please change bert\_path, path to the bert model.

1. To Train BERT base model:

sh sample_commands/bert.sh

2. To Train VIBERT model:

sh sample_commands/vibert.sh

3. To train Dropout model:

sh sample_commands/dropout.sh

4. To train Mixout model:

sh sample_commands/mixout.sh

5. To train WD model:

sh sample_commands/wd.sh

6. To train VIBERT(beta=0) used in ablation study:

sh sample_commands/bert_beta_zero.sh

Bibliography

If you find this repo useful, please cite our paper.

@inproceedings{
mahabadi2021variational,
title={Variational Information Bottleneck for Effective Low-Resource Fine-Tuning},
author={Rabeeh Karimi mahabadi and Yonatan Belinkov and James Henderson},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=kvhzKz-_DMF}
}

Final words

Hope this repo is useful for your research. For any questions, please create an issue or email [email protected] or [email protected], and we will get back to you as soon as possible.