779f22fc47
This pr is auto merged as it contains a mandatory file and is opened for more than 10 days. |
||
---|---|---|
MASS-summarization | ||
MASS-supNMT | ||
MASS-unsupNMT | ||
figs | ||
.gitignore | ||
CONTRIBUTING.md | ||
LICENSE.TXT | ||
NOTICE.md | ||
README.md | ||
SECURITY.md |
README.md
MASS
MASS: Masked Sequence to Sequence Pre-training for Language Generation, by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu, is a novel pre-training method for sequence to sequence based language generation tasks. It randomly masks a sentence fragment in the encoder, and then predicts it in the decoder.
MASS can be applied on cross-lingual tasks such as neural machine translation (NMT), and monolingual tasks such as text summarization. The current codebase supports unsupervised NMT (implemented based on XLM), supervised NMT, text summarization and conversational response generation, which are all based on Fairseq. We will release our implementation for other sequence to sequence generation tasks in the future.
What is New!
We release MPNet, a new pre-trained method for language understanding. GitHub: https://github.com/microsoft/MPNet
Unsupervised NMT
Unsupervised Neural Machine Translation just uses monolingual data to train the models. During MASS pre-training, the source and target languages are pre-trained in one model, with the corresponding langauge embeddings to differentiate the langauges. During MASS fine-tuning, back-translation is used to train the unsupervised models. Code is under MASS-unsupNMT. We provide pre-trained and fine-tuned models:
Languages | Pre-trained Model | Fine-tuned Model | BPE codes | Vocabulary |
---|---|---|---|---|
EN - FR | MODEL | MODEL | BPE codes | Vocabulary |
EN - DE | MODEL | MODEL | BPE codes | Vocabulary |
En - RO | MODEL | MODEL | BPE_codes | Vocabulary |
We are also preparing larger models on more language pairs, and will release them in the future.
Dependencies
Currently we implement MASS for unsupervised NMT based on the codebase of XLM. The depencies are as follows:
- Python 3
- NumPy
- PyTorch (version 0.4 and 1.0)
- fastBPE (for BPE codes)
- Moses (for tokenization)
- Apex (for fp16 training)
Data Ready
We use the same BPE codes and vocabulary with XLM. Here we take English-French as an example.
cd MASS
wget https://dl.fbaipublicfiles.com/XLM/codes_enfr
wget https://dl.fbaipublicfiles.com/XLM/vocab_enfr
./get-data-nmt.sh --src en --tgt fr --reload_codes codes_enfr --reload_vocab vocab_enfr
Pre-training:
python train.py \
--exp_name unsupMT_enfr \
--data_path ./data/processed/en-fr/ \
--lgs 'en-fr' \
--mass_steps 'en,fr' \
--encoder_only false \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--tokens_per_batch 3000 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 200000 \
--max_epoch 100 \
--eval_bleu true \
--word_mass 0.5 \
--min_len 5 \
During the pre-training prcess, even without any back-translation, you can observe the model can achieve some intial BLEU scores:
epoch -> 4
valid_fr-en_mt_bleu -> 10.55
valid_en-fr_mt_bleu -> 7.81
test_fr-en_mt_bleu -> 11.72
test_en-fr_mt_bleu -> 8.80
Distributed Training
To use multiple GPUs e.g. 3 GPUs on same node
export NGPU=3; CUDA_VISIBLE_DEVICES=0,1,2 python -m torch.distributed.launch --nproc_per_node=$NGPU train.py [...args]
To use multiple GPUS across many nodes, use Slurm to request multi-node job and launch the above command. The code automatically detects the SLURM_* environment vars to distribute the training.
Fine-tuning
After pre-training, we use back-translation to fine-tune the pre-trained model on unsupervised machine translation:
MODEL=mass_enfr_1024.pth
python train.py \
--exp_name unsupMT_enfr \
--data_path ./data/processed/en-fr/ \
--lgs 'en-fr' \
--bt_steps 'en-fr-en,fr-en-fr' \
--encoder_only false \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--tokens_per_batch 2000 \
--batch_size 32 \
--bptt 256 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 200000 \
--max_epoch 30 \
--eval_bleu true \
--reload_model "$MODEL,$MODEL" \
We also provide a demo to use MASS pre-trained model on the WMT16 en-ro bilingual dataset. We provide pre-trained and fine-tuned models:
Model | Ro-En BLEU (with BT) |
---|---|
Baseline | 34.0 |
XLM | 38.5 |
MASS | 39.1 |
Download dataset by the below command:
wget https://dl.fbaipublicfiles.com/XLM/codes_enro
wget https://dl.fbaipublicfiles.com/XLM/vocab_enro
./get-data-bilingual-enro-nmt.sh --src en --tgt ro --reload_codes codes_enro --reload_vocab vocab_enro
After download the mass pre-trained model from the above link. And use the following command to fine tune:
MODEL=mass_enro_1024.pth
python train.py \
--exp_name unsupMT_enro \
--data_path ./data/processed/en-ro \
--lgs 'en-ro' \
--bt_steps 'en-ro-en,ro-en-ro' \
--encoder_only false \
--mt_steps 'en-ro,ro-en' \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--tokens_per_batch 2000 \
--batch_size 32 \
--bptt 256 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 200000 \
--max_epoch 50 \
--eval_bleu true \
--reload_model "$MODEL,$MODEL"
Supervised NMT
We also implement MASS on fairseq, in order to support the pre-training and fine-tuning for large scale supervised tasks, such as neural machine translation, text summarization. Unsupervised pre-training usually works better in zero-resource or low-resource downstream tasks. However, in large scale supervised NMT, there are plenty of bilingual data, which brings challenges for conventional unsupervised pre-training. Therefore, we design new pre-training loss to support large scale supervised NMT. The code is under MASS-supNMT.
We extend the MASS to supervised setting where the supervised sentence pair (X, Y) is leveraged for pre-training. The sentence X is masked and feed into the encoder, and the decoder predicts the whole sentence Y. Some discret tokens in the decoder input are also masked, to encourage the decoder to extract more informaiton from the encoder side.
During pre-training, we combine the orignal MASS pre-training loss and the new supervised pre-training loss together. During fine-tuning, we directly use supervised sentence pairs to fine-tune the pre-trained model. Except for NMT, this pre-trainig paradigm can be also applied on other superviseed sequence to sequence tasks.
We release the pre-trained model and example codes of how to pre-train and fine-tune on WMT Chinese<->English (Zh<->En) translation.:
Languages | Pre-trained Model | BPE codes | English-Dict | Chinese-Dict |
---|---|---|---|---|
Zh - En | MODEL | CODE | VOCAB | VOCAB |
Prerequisites
After download the repository, you need to install fairseq
by pip
:
pip install fairseq==0.7.1
Data Ready
We first prepare the monolingual and bilingual sentences for Chinese and English respectively. The data directory looks like:
- data/
├─ mono/
| ├─ train.en
| ├─ train.zh
| ├─ valid.en
| ├─ valid.zh
| ├─ dict.en.txt
| └─ dict.zh.txt
└─ para/
├─ train.en
├─ train.zh
├─ valid.en
├─ valid.zh
├─ dict.en.txt
└─ dict.zh.txt
The files under mono
are monolingual data, while under para
are bilingual data. dict.en(zh).txt
in different directory should be identical. The dictionary for different language can be different. Running the following command can generate the binarized data:
# Ensure the output directory exists
data_dir=data/
mono_data_dir=$data_dir/mono/
para_data_dir=$data_dir/para/
save_dir=$data_dir/processed/
# set this relative path of MASS in your server
user_dir=mass
mkdir -p $data_dir $save_dir $mono_data_dir $para_data_dir
# Generate Monolingual Data
for lg in en zh
do
fairseq-preprocess \
--task cross_lingual_lm \
--srcdict $mono_data_dir/dict.$lg.txt \
--only-source \
--trainpref $mono_data_dir/train --validpref $mono_data_dir/valid \
--destdir $save_dir \
--workers 20 \
--source-lang $lg
# Since we only have a source language, the output file has a None for the
# target language. Remove this
for stage in train valid
do
mv $save_dir/$stage.$lg-None.$lg.bin $save_dir/$stage.$lg.bin
mv $save_dir/$stage.$lg-None.$lg.idx $save_dir/$stage.$lg.idx
done
done
# Generate Bilingual Data
fairseq-preprocess \
--user-dir $mass_dir \
--task xmasked_seq2seq \
--source-lang en --target-lang zh \
--trainpref $para_data_dir/train --validpref $para_data_dir/valid \
--destdir $save_dir \
--srcdict $para_data_dir/dict.en.txt \
--tgtdict $para_data_dir/dict.zh.txt
Pre-training
We provide a simple demo code to demonstrate how to deploy mass pre-training.
save_dir=checkpoints/mass/pre-training/
user_dir=mass
data_dir=data/processed/
mkdir -p $save_dir
fairseq-train $data_dir \
--user-dir $user_dir \
--save-dir $save_dir \
--task xmasked_seq2seq \
--source-langs en,zh \
--target-langs en,zh \
--langs en,zh \
--arch xtransformer \
--mass_steps en-en,zh-zh \
--memt_steps en-zh,zh-en \
--optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
--lr-scheduler inverse_sqrt --lr 0.00005 --min-lr 1e-09 \
--criterion label_smoothed_cross_entropy \
--max-tokens 4096 \
--dropout 0.1 --relu-dropout 0.1 --attention-dropout 0.1 \
--max-update 100000 \
--share-decoder-input-output-embed \
--valid-lang-pairs en-zh \
We also provide a pre-training script which is used for our released model.
Fine-tuning
After pre-training stage, we fine-tune the model on bilingual sentence pairs:
data_dir=data/processed
save_dir=checkpoints/mass/fine_tune/
user_dir=mass
model=checkpoint/mass/pre-training/checkpoint_last.pt # The path of pre-trained model
mkdir -p $save_dir
fairseq-train $data_dir \
--user-dir $user_dir \
--task xmasked_seq2seq \
--source-langs zh --target-langs en \
--langs en,zh \
--arch xtransformer \
--mt_steps zh-en \
--save-dir $save_dir \
--optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
--lr-scheduler inverse_sqrt --lr-shrink 0.5 --lr 0.00005 --min-lr 1e-09 \
--criterion label_smoothed_cross_entropy \
--max-tokens 4096 \
--max-update 100000 --max-epoch 50 \
--dropout 0.1 --relu-dropout 0.1 --attention-dropout 0.1 \
--share-decoder-input-output-embed \
--valid-lang-pairs zh-en \
--reload_checkpoint $model
We also provide a fine-tuning script which is used for our pre-trained model.
Inference
After the fine-tuning stage, you can generate translation results by using the below script:
model=checkpoints/mass/fine_tune/checkpoint_best.pt
data_dir=data/processed
user_dir=mass
fairseq-generate $data_dir \
--user-dir $user_dir \
-s zh -t en \
--langs en,zh \
--source-langs zh --target-langs en \
--mt_steps zh-en \
--gen-subset valid \
--task xmasked_seq2seq \
--path $model \
--beam 5 --remove-bpe
Text Summarization
MASS for text summarization is also implemented on fairseq. The code is under MASS-summarization.
Dependency
pip install torch==1.0.0
pip install fairseq==0.8.0
MODEL
MASS uses default Transformer structure. We denote L, H, A as the number of layers, the hidden size and the number of attention heads.
Model | Encoder | Decoder | Download |
---|---|---|---|
MASS-base-uncased | 6L-768H-12A | 6L-768H-12A | MODEL |
MASS-middle-uncased | 6L-1024H-16A | 6L-1024H-16A | MODEL |
Results on Abstractive Summarization (12/03/2019)
Dataset | RG-1 | RG-2 | RG-L |
---|---|---|---|
CNN/Daily Mail | 43.05 | 20.02 | 40.08 |
Gigaword | 38.93 | 20.20 | 36.20 |
XSum | 39.75 | 17.24 | 31.95 |
Evaluated by files2rouge.
Pipeline for Pre-Training
Download data
Our model is trained on Wikipekia + BookCorpus. Here we use wikitext-103 to demonstrate how to process data.
wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-raw-v1.zip
unzip wikitext-103-raw-v1.zip
Tokenize corpus
We use wordpiece vocabuary (from bert) to tokenize the original text data directly. We provide a script to deal with data. You need to pip install pytorch_transformers
first to generate tokenized data.
mkdir -p mono
for SPLIT in train valid test; do
python encode.py \
--inputs wikitext-103-raw/wiki.${SPLIT}.raw \
--outputs mono/${SPLIT}.txt \
--workers 60; \
done
Binarized data
wget -c https://modelrelease.blob.core.windows.net/mass/mass-base-uncased.tar.gz
tar -zxvf mass-base-uncased.tar.gz
# Move dict.txt from tar file to the data directory
fairseq-preprocess \
--user-dir mass --only-source \
--trainpref mono/train.txt --validpref mono/valid.txt --testpref mono/test.txt \
--destdir processed --srcdict dict.txt --workers 60
Pre-training
TOKENS_PER_SAMPLE=512
WARMUP_UPDATES=10000
PEAK_LR=0.0005
TOTAL_UPDATES=125000
MAX_SENTENCES=8
UPDATE_FREQ=16
fairseq-train processed \
--user-dir mass --task masked_s2s --arch transformer_mass_base \
--sample-break-mode none \
--tokens-per-sample $TOKENS_PER_SAMPLE \
--criterion masked_lm \
--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 \
--ddp-backend=no_c10d \
Pipeline for Fine-tuning (CNN / Daily Mail)
Data
Download, tokenize and truncate data from this link, and use the above tokenization to generate wordpiece-level data. Rename the shuffix article
and title
as src
and tgt
. Assume the tokenized data is under cnndm/para
fairseq-preprocess \
--user-dir mass --task masked_s2s \
--source-lang src --target-lang tgt \
--trainpref cnndm/para/train --validpref cnndm/para/valid --testpref cnndm/para/test \
--destdir cnndm/processed --srcdict dict.txt --tgtdict dict.txt \
--workers 20
dict.txt
is included in mass-base-uncased.tar.gz
. A copy of binarized data can be obtained from here.
Running
fairseq-train cnndm/processed/ \
--user-dir mass --task translation_mass --arch transformer_mass_base \
--optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
--lr 0.0005 --min-lr 1e-09 \
--lr-scheduler inverse_sqrt --warmup-init-lr 1e-07 --warmup-updates 4000 \
--weight-decay 0.0 \
--criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
--update-freq 8 --max-tokens 4096 \
--ddp-backend=no_c10d --max-epoch 25 \
--max-source-positions 512 --max-target-positions 512 \
--skip-invalid-size-inputs-valid-test \
--load-from-pretrained-model mass-base-uncased.pt \
lr=0.0005
is not the optimal choice for any task. It is tuned on the dev set (among 1e-4, 2e-4, 5e-4).
Inference
MODEL=checkpoints/checkpoint_best.pt
fairseq-generate $DATADIR --path $MODEL \
--user-dir mass --task translation_mass \
--batch-size 64 --beam 5 --min-len 50 --no-repeat-ngram-size 3 \
--lenpen 1.0 \
min-len
is sensitive for different tasks, lenpen
needs to be tuned on the dev set.
Reference
If you find MASS useful in your work, you can cite the paper as below:
@inproceedings{song2019mass,
title={MASS: Masked Sequence to Sequence Pre-training for Language Generation},
author={Song, Kaitao and Tan, Xu and Qin, Tao and Lu, Jianfeng and Liu, Tie-Yan},
booktitle={International Conference on Machine Learning},
pages={5926--5936},
year={2019}
}
Related Works
- MPNet: Masked and Permuted Pre-training for Language Understanding, by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu. GitHub: https://github.com/microsoft/MPNet