Rethinking Semi-supervised Learning with Language Models

This repository contains the code for the paper titled Rethinking Semi-supervised Learning with Language Models, built upon huggingface and semilearn.

Quick Links

• Rethinking Semi-supervised Learning with Language Models
  • Quick Links
  • Overview
  • 1. Installation
  • 2. Preprocess the data
  • 3. Task Adaptive Pre-Training
  • 4. Fine-tuning from the roberta-base or the pre-trained checkpoints
  • 5. Self Training
  • Bugs or questions?
  • Citation
  • Authors
  • Security
  • License

Overview

You can reproduce the continued pre-training and self training experiments of our recent paper Rethinking Semi-supervised Learning with Language Models.

1. Installation

conda create --name storpt python=3.8
conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
pip install -r requirements.txt

2. Preprocess the data

The dataset can be downloaded via the following repository: semilearn. To preprocess data for language modeling, follow these steps:

• Set TASK_NAME as one of the following: aclImdb, ag_news, amazon_review, yahoo_answers, or amazon_review.
• Run the provided code snippets for the desired format.

For task adaptive pre-training format:

for TASK_NAME in aclImdb ag_news amazon_review yahoo_answers; do
    python tools/convert_to_pretrain_format.py --task_name ${TASK_NAME};
done

For the format of the text classification:

for TASK_NAME in aclImdb ag_news amazon_review yahoo_answers; do
    python tools/convert_to_finetune_format.py --task_name ${TASK_NAME};
done

For text classification with a partially labeled dataset (Use the index of labeled data to ensure fair comparison with semi-supervised approaches):

TASK_NAME=aclImdb
LABEL_SIZE=20
python tools/convert_to_finetune_semi_format.py --num_labels ${LABEL_SIZE} --task_name ${TASK_NAME}

To change the size of the dataset:

TASK_NAME=amazon_review
TRAIN_SIZE=23000
VAL_SIZE=2000
TEST_SIZE=25000
python tools/convert_dataset_size.py --task_name ${TASK_NAME} --train_size ${TRAIN_SIZE} --valid_size ${VAL_SIZE} --test_size ${TEST_SIZE}
python tools/convert_dataset_size.py --task_name ${TASK_NAME} --train_size ${TRAIN_SIZE} --keep_original_dev_test

3. Task Adaptive Pre-Training

The code for masked language modeling can be found in run_mlm.py. To perform task adaptive pre-training, execute the command below:

for TASK_NAME in aclImdb ag_news amazon_review yahoo_answers; do \
    python run_mlm.py \
        --model_name_or_path roberta-base \
        --train_file data/${TASK_NAME}/train.txt \
        --validation_file data/${TASK_NAME}/dev.txt \
        --line_by_line \
        --per_device_train_batch_size 8 \
        --per_device_eval_batch_size 8 \
        --gradient_accumulation_steps 4 \
        --learning_rate 1e-04 \
        --optim adamw_torch \
        --weight_decay 0.01 \
        --adam_beta1 0.9 \
        --adam_beta2 0.98 \
        --adam_epsilon 1e-06 \
        --do_train \
        --do_eval \
        --save_steps 500 \
        --evaluation_strategy steps \
        --eval_steps 500 \
        --num_train_epochs 100 \
        --warmup_ratio 0.06 \
        --mlm_probability 0.15 \
        --fp16 \
        --output_dir saved_tapt/${TASK_NAME} \
        --load_best_model_at_end; \
done

Note: The above command assumes training on 8x16GB V100 GPUs. Each GPU uses a batch size of 8 sequences and accumulates gradients for a total batch size of 256 sequences. If you have a GPU with mixed precision capabilities (architecture Pascal or more recent), you can use mixed precision training with PyTorch 1.6.0 or latest by adding the flag --fp16 to the scripts mentioned above! The learning rate and batch size are closely related and should be adjusted together. More details can be found here.

4. Fine-tuning from the roberta-base or the pre-trained checkpoints

The code for masked language modeling can be found in run_glue.py. To perform task adaptive fine-tuning, execute the command below:

TASK_NAME=
LABEL_SIZE=
CHECKPOINT_DIR=roberta-base
for seed in 1 2 3 4 5; do \
    CUDA_VISIBLE_DEVICES=0 python run_glue.py \
        --train_file data/${TASK_NAME}/labeled_idx/train_ft_label${LABEL_SIZE}_seed${seed}.json \
        --validation_file data/${TASK_NAME}/dev_ft.json \
        --test_file data/${TASK_NAME}/test_ft.json \
        --model_name_or_path ${CHECKPOINT_DIR} \
        --seed ${seed} \
        --do_train \
        --do_eval \
        --do_predict \
        --per_device_train_batch_size 16 \
        --per_device_eval_batch_size 16 \
        --max_seq_length 256 \
        --num_train_epochs 50 \
        --save_strategy epoch \
        --evaluation_strategy epoch \
        --learning_rate 2e-05 \
        --warmup_ratio 0.0 \
        --fp16 \
        --metric_for_best_model eval_f1 \
        --load_best_model_at_end \
        --save_total_limit 1 \
        --output_dir saved_finetuned/${TASK_NAME}_label${LABEL_SIZE}_seed${seed}; \
done

5. Self Training

Run self training

ALGORITHM=
TASK_NAME=
LABEL_SIZE=
for seed in 1 2 3 4 5; do \
    CUDA_VISIBLE_DEVICES=0 python train.py \
        --seed ${seed} \
        --c config_roberta/${ALGORITHM}/${ALGORITHM}_${TASK_NAME}_${LABEL_SIZE}_0.yaml \
        --save_dir ./saved_models \
        --save_name ${ALGORITHM}_${TASK_NAME}_${LABEL_SIZE}_${seed} \
        --load_path ./saved_models/${ALGORITHM}_${TASK_NAME}_${LABEL_SIZE}_${seed}/latest_model.pth; \
done

Train models under the self-taught learning (STL) settings.

ALGORITHM=
TASK_NAME=
UNLABEL_DATASET=
LABEL_SIZE=
for seed in 1 2 3 4 5; do \
    CUDA_VISIBLE_DEVICES=0 python train.py \
        --seed ${seed} \
        --num_labels ${LABEL_SIZE} \
        --dataset ${TASK_NAME} \
        --custom_unlabeled_data_file data/${UNLABEL_DATASET}/train.json \
        --c config_roberta/${ALGORITHM}/${ALGORITHM}_aclImdb_100_0.yaml \
        --save_dir saved_domainshift_stl \
        --save_name ${ALGORITHM}_${TASK_NAME}_${UNLABEL_DATASET}_${LABEL_SIZE}_${seed} \
        --load_path ./saved_domainshift_stl/${ALGORITHM}_${TASK_NAME}_${UNLABEL_DATASET}_${LABEL_SIZE}_${seed}/latest_model.pth; \
done

Train models under Unsupervised Domain Adaptation (UDA) settings.

ALGORITHM=
SOURCE_DATASET=
TARGET_DATASET=
LABEL_SIZE=
for seed in 1 2 3 4 5; do \
    CUDA_VISIBLE_DEVICES=0 python train.py \
        --seed ${seed} \
        --num_labels ${LABEL_SIZE} \
        --dataset ${SOURCE_DATASET} \
        --custom_unlabeled_data_file data/${TARGET_DATASET}/train.json \
        --custom_dev_data_file data/${TARGET_DATASET}/dev.json \
        --custom_test_data_file data/${TARGET_DATASET}/test.json \
        --c config_roberta/${ALGORITHM}/${ALGORITHM}_${SOURCE_DATASET}_${LABEL_SIZE}_0.yaml \
        --save_dir saved_domainshift_uda \
        --save_name ${ALGORITHM}_${SOURCE_DATASET}_${TARGET_DATASET}_${LABEL_SIZE}_${seed} \
        --load_path ./saved_domainshift_uda/${ALGORITHM}_${SOURCE_DATASET}_${TARGET_DATASET}_${LABEL_SIZE}_${seed}/latest_model.pth; \
done

Bugs or questions?

If you have any inquiries pertaining to the code or the paper, please do not hesitate to contact Zhengxiang Shi. In case you encounter any issues while utilising the code or wish to report a bug, you may open an issue. We kindly request that you provide specific details regarding the problem so that we can offer prompt and efficient assistance.

Citation

@inproceedings{shi2023rethinking,
  title={Rethinking Semi-supervised Learning with Language Models},
  author={Shi, Zhengxiang and Tonolini, Francesco and Aletras, Nikolaos and Yilmaz, Emine and Kazai, Gabriella and Jiao, Yunlong},
  year={2023},
  address = {Toronto, Canada},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2023},
  publisher = {Association for Computational Linguistics},
}

Authors

• Zhengxiang Shi: Main contributor

Security

See CONTRIBUTING for more information.

License

This project is licensed under the Apache-2.0 License.