MPNet

MPNet: Masked and Permuted Pre-training for Language Understanding, by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu, is a novel pre-training method for language understanding tasks. It solves the problems of MLM (masked language modeling) in BERT and PLM (permuted language modeling) in XLNet and achieves better accuracy.

News: We have updated the pre-trained models now.

Supported Features

A unified view and implementation of several pre-training models including BERT, XLNet, MPNet, etc.
Code for pre-training and fine-tuning for a variety of language understanding (GLUE, SQuAD, RACE, etc) tasks.

Installation

We implement MPNet and this pre-training toolkit based on the codebase of fairseq. The installation is as follow:

pip install --editable pretraining/
pip install pytorch_transformers==1.0.0 transformers scipy sklearn

Pre-training MPNet

Our model is pre-trained with bert dictionary, you first need to pip install transformers to use bert tokenizer. We provide a script encode.py and a dictionary file dict.txt to tokenize your corpus. You can modify encode.py if you want to use other tokenizers (like roberta).

1) Preprocess data

We choose WikiText-103 as a demo. The running script is as follow:

wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-raw-v1.zip
unzip wikitext-103-raw-v1.zip

for SPLIT in train valid test; do \
    python MPNet/encode.py \
        --inputs wikitext-103-raw/wiki.${SPLIT}.raw \
        --outputs wikitext-103-raw/wiki.${SPLIT}.bpe \
        --keep-empty \
        --workers 60; \
done

Then, we need to binarize data. The command of binarizing data is following:

fairseq-preprocess \
    --only-source \
    --srcdict MPNet/dict.txt \
    --trainpref wikitext-103-raw/wiki.train.bpe \
    --validpref wikitext-103-raw/wiki.valid.bpe \
    --testpref wikitext-103-raw/wiki.test.bpe \
    --destdir data-bin/wikitext-103 \
    --workers 60

2) Pre-train MPNet

The below command is to train a MPNet model:

TOTAL_UPDATES=125000    # Total number of training steps
WARMUP_UPDATES=10000    # Warmup the learning rate over this many updates
PEAK_LR=0.0005          # Peak learning rate, adjust as needed
TOKENS_PER_SAMPLE=512   # Max sequence length
MAX_POSITIONS=512       # Num. positional embeddings (usually same as above)
MAX_SENTENCES=16        # Number of sequences per batch (batch size)
UPDATE_FREQ=16          # Increase the batch size 16x

DATA_DIR=data-bin/wikitext-103

fairseq-train --fp16 $DATA_DIR \
    --task masked_permutation_lm --criterion masked_permutation_cross_entropy \
    --arch mpnet_base --sample-break-mode complete --tokens-per-sample $TOKENS_PER_SAMPLE \
    --optimizer adam --adam-betas '(0.9,0.98)' --adam-eps 1e-6 --clip-norm 0.0 \
    --lr-scheduler polynomial_decay --lr $PEAK_LR --warmup-updates $WARMUP_UPDATES --total-num-update $TOTAL_UPDATES \
    --dropout 0.1 --attention-dropout 0.1 --weight-decay 0.01 \
    --max-sentences $MAX_SENTENCES --update-freq $UPDATE_FREQ \
    --max-update $TOTAL_UPDATES --log-format simple --log-interval 1 --input-mode 'mpnet'

Notes: You can replace arch with mpnet_rel_base and add command --mask-whole-words --bpe bert to use relative position embedding and whole word mask.

Notes: You can specify --input-mode as mlm or plm to train masked language model or permutation language model.

Pre-trained models

We have updated the final pre-trained MPNet model for fine-tuning.

You can load the pre-trained MPNet model like this:

from fairseq.models.masked_permutation_net import MPNet
mpnet = MPNet.from_pretrained('checkpoints', 'checkpoint_best.pt', 'path/to/data', bpe='bert')
assert isinstance(mpnet.model, torch.nn.Module)

Fine-tuning MPNet on down-streaming tasks

Acknowledgements

Our code is based on fairseq-0.8.0. Thanks for their contribution to the open-source commuity.

Reference

If you find this toolkit useful in your work, you can cite the corresponding papers listed below:

@article{song2020mpnet,
    title={MPNet: Masked and Permuted Pre-training for Language Understanding},
    author={Song, Kaitao and Tan, Xu and Qin, Tao and Lu, Jianfeng and Liu, Tie-Yan},
    journal={arXiv preprint arXiv:2004.09297},
    year={2020}
}

Related Works

MASS: Masked Sequence to Sequence Pre-training for Language Generation, by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu. GitHub: https://github.com/microsoft/MASS
LightPAFF: A Two-Stage Distillation Framework for Pre-training and Fine-tuning, by Kaitao Song, Hao Sun, Xu Tan, Tao Qin, Jianfeng Lu, Hongzhi Liu, Tie-Yan Liu

microsoft/MPNet