Awesome

AcTune

This is the code repo for our paper `AcTune: Uncertainty-Based Active Self-Training for Active Fine-Tuning of Pretrained Language Models' (In Proceedings of NAACL 2022 Main Conference, Oral Presentation).

Requirements

python 3.7
transformers==4.2.0
pytorch==1.6.0
tqdm
scikit-learn
faiss-cpu==1.6.4

Datasets

Main Experiments

We use the following four datasets for the main experiments.

The processed data can be found at this link. The folder to put these datasets will be discribed in the following parts.

Weak Supervision Experiments

Most of the dataset are from the WRENCH benchmark. Please checkout their repo for dataset details.

Training

Please use the commands in commands folder for experiments. Take AG News dataset as an example, run_agnews_finetune.sh is used for running the experiment of standard active learning approaches, and run_agnews.sh is used for running active self-training experiments as unlabeled data is also used during fine-tuning.

Here, we suppose there is a folder for storing datasets as in ../datasets/, and a folder for logging the experiment results as in ../exp.

Hyperparameter Tuning

The key hyperparameter for our approach includes pool, pool_min, gamma, gamma_min, self_training_weight sample_per_group and n_centroids.

• pool stands for the number of samples selected for self-training on average for each round. For example, if
• pool_min stands for the initial number of samples selected for self-training. For example, if pool_min=3000 and pool=4000 and there are 10 rounds in total, it means that in the first round, it selects 3000 samples, and in the later rounds, the number of low-uncertainty samples used for self-training will increase linearly. Finally, the total number of unlabeled data used for self-training equals to 4000*10=40000.
• gamma stands for the final weight of momentum-based memory bank.
• gamma_min stands for the initial weight of momentum-based memory bank. The weight will gradually become closer to gamma in later rounds.
• n_centroids is the number of clusters used in region-aware sampling.
• sample_per_group is the number of samples selected in each high-uncertainty clusters.

Citation

Please cite the following paper if you are using our datasets/tool. Thanks!

@inproceedings{yu2022actune,
    title = "{A}c{T}une: Uncertainty-Based Active Self-Training for Active Fine-Tuning of Pretrained Language Models",
    author = "Yu, Yue and Kong, Lingkai and Zhang, Jieyu and Zhang, Rongzhi and Zhang, Chao",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.naacl-main.102",
    pages = "1422--1436",
}