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