OR-NMT: Bridging the Gap between Training and Inference for Neural Machine Translation

Wen Zhang, Yang Feng, Fandong Meng, Di You and Qun Liu. Bridging the Gap between Training and Inference for Neural Machine Translation. In Proceedings of ACL, 2019. [paper][code]

Codes in the two directories are the OR-NMT systems based on the RNNsearch and Transformer models correspondingly

• OR-RNNsearch: based on the RNNsearch system which we implemented from scratch
• OR-Transformer: based on the Transformer system fairseq implemented by Facebook

Runtime Environment

This system has been tested in the following environment.

• OS: Ubuntu 16.04.1 LTS 64 bits
• Python version >=3.6
• Pytorch version >=1.2

For OR-Transformer:

First, go into the OR-Transformer directory.

Then, the training script is the same with fairseq, except for the following arguments:

• add --use-word-level-oracles for training Transformer by word-level oracle.
• add --use-sentence-level-oracles for training Transformer by sentence-level oracle.

By default, the probability is decayed based on the update index.

• add --use-epoch-numbers-decay for decaying based on the epoch index.
• the hyperparameter --decay-k is used to control the speed of the inverse sigmoid decay, which is in Eq.(15) in the paper.
  • set 8~15 for the decaying based on epoch index
  • set 3000~8000 for the decaying based on update index
  • The larger the value, the slower the decay, vice versa.

NOTE: For a new data set, the hyperparameter --decay-k needs to be manually adjusted according to the maximum number of training updates (default) or epochs (--use-epoch-numbers-decay) to ensure that the probability of sampling golden words does not decay so quickly.

For Eq.(11~13) in the paper, is actually the same as . The operation is not needed in the code implementation.

Gumbel noise:

• add --use-greed-gumbel-noise to sample word-level oracle with Gumbel noise
• add --use-bleu-gumbel-noise to sample sentence-level oracle with Gumbel noise
• --gumbel-noise is used as the hyper-parameter in the calculation of Gumbel noise
• --oracle-search-beam-size is used to set the beam size in length-constrained decoding

As for the --arch and --criterion arguments, oracle_ should be used as the prefix for OR-NMT training, such as:

• --arch transformer_vaswani_wmt_en_de_big -> --arch oracle_transformer_vaswani_wmt_en_de_big
• --criterion label_smoothed_cross_entropy -> --criterion oracle_label_smoothed_cross_entropy

Example of the script for word-level training and decaying the probability based on epoch index:

export CUDA_VISIBLE_DEVICES=0,1,2,3
batch_size=4096
accum=2
data_dir=directory_of_data_bin
model_dir=./ckpt
python train.py $data_dir \
    --arch oracle_transformer_vaswani_wmt_en_de_big --share-all-embeddings \
    --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --lr-scheduler inverse_sqrt \
    --warmup-init-lr 1e-07 --warmup-updates 4000 --lr 0.0005 --min-lr 1e-09 \
    --weight-decay 0.0 --criterion oracle_label_smoothed_cross_entropy --label-smoothing 0.1 \
    --max-tokens $batch_size --update-freq $accum --no-progress-bar --log-format json --max-update 200000 \
    --log-interval 10 --save-interval-updates 10000 --keep-interval-updates 10 --save-interval 10000 \
    --seed 1111 --skip-invalid-size-inputs-valid-test \
    --distributed-port 28888 --distributed-world-size 4 --ddp-backend=no_c10d \
    --source-lang en --target-lang de --save-dir $model_dir \
    --use-word-level-oracles --use-epoch-numbers-decay --decay-k 10 \
    --use-greed-gumbel-noise --gumbel-noise 0.5 | tee -a $model_dir/training.log

Model settings on NIST Chinese->English (Zh->En) and WMT'14 English->German (En->De).

#Toks. means batchsize on single GPU.

#Freq. means the times of gradient accumulation.

Max represents the maximum number of training epochs (30) or updates (80k).

Results and Settings on NIST Chinese->English translation task

We calculate the case-insensitive 4-gram tokenized BLEU by script multibleu.perl

Models	Dev. (MT02)	MT03	MT04	MT05	MT06	MT08	Average
Transformer-big	48.50	47.29	47.79	48.28	47.50	38.50	45.87	30
+Word-level Oracle (==10)	49.18	48.70	48.67	48.69	48.49	39.58	46.83	30
+Word-level Oracle (==15)	49.05	48.57	48.73	48.68	48.59	39.68	46.85	30
+Word-level Oracle (==20)	49.30	48.46	48.57	48.87	48.57	39.46	46.79	30
+Word-level Oracle (==25)	48.88	48.32	48.66	48.74	48.32	39.38	46.68	30
+Word-level Oracle (==30)	48.47	48.37	48.50	48.63	48.07	39.54	46.62	46.74	30

We also evaluate by the case-insensitive 4-gram detokenized BLEU with SacreBLEU, which is calculated the script score.py provided by fairseq: BLEU+case.mixed+lang.en-{de,fr}+numrefs.4+smooth.exp+tok.13a+version.1.4.4

Models	Dev. (MT02)	MT03	MT04	MT05	MT06	MT08	Average
Transformer-big	48.46	47.41	47.88	48.25	47.52	38.60	45.93	30
+Word-level Oracle (==10)	49.20	48.80	48.77	48.64	48.49	39.79	46.90	30
+Word-level Oracle (==15)	49.07	48.64	48.81	48.63	48.65	39.88	46.92	30
+Word-level Oracle (==20)	49.32	48.54	48.73	48.82	48.51	39.50	46.82	30
+Word-level Oracle (==25)	48.90	48.18	48.70	48.59	47.73	39.14	46.47	30
+Word-level Oracle (==30)	48.53	48.59	48.74	48.58	48.07	39.71	46.74	30

The setting of the NIST Chinese->English:

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
data_bin_dir=directory_of_data_bin
model_dir=./ckpt
python train.py $data_bin_dir \
    --arch oracle_transformer_vaswani_wmt_en_de_big --share-all-embeddings \
    --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --lr-scheduler inverse_sqrt \
    --warmup-init-lr 1e-07 --warmup-updates 4000 --lr 0.0007 --min-lr 1e-09 \
    --weight-decay 0.0 --criterion oracle_label_smoothed_cross_entropy --label-smoothing 0.1 \
    --max-tokens 4096 --update-freq 3 --no-progress-bar --log-format json --max-epoch 30 \
    --log-interval 10 --save-interval 2 --keep-last-epochs 10 \
    --seed 1111 --use-epoch-numbers-decay \
    --use-word-level-oracles --decay-k 15 --use-greed-gumbel-noise --gumbel-noise 0.5 \
    --distributed-port 32222 --distributed-world-size 8 --ddp-backend=no_c10d \
    --source-lang zh --target-lang en --save-dir $model_dir | tee -a $model_dir/training.log

As Eq.(15) in the paper, the probability of sampling golden words decays with the number of epochs as follows:

Results and Settings on WMT'14 English->German translation task

We calculate the case-sensitive 4-gram tokenized BLEU by script multibleu.perl

Models	newstest2014	#update
Transformer-base	27.54	80000
+Word-level Oracle (==50, ==0.8)	28.01	80000
+Sentence-level Oracle (==5800, ==0.5, beam_size==4)	28.45	40000

We also evaluate by the case-sensitive 4-gram detokenized BLEU with SacreBLEU, which is calculated the script score.py provided by fairseq: BLEU+case.mixed+lang.en-{de,fr}+numrefs.1+smooth.exp+tok.13a+version.1.4.4

Models	newstest2014	#update
Transformer-base	26.45	80000
+Word-level Oracle (==50, ==0.8)	26.86	80000
+Sentence-level Oracle (==5800, ==0.5, beam_size==4)	27.24	40000

Setting of the word-level oracle for the WMT'14 English->German dataset:

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
data_bin_dir=directory_of_data_bin
model_dir=./ckpt
python train.py $data_bin_dir \
    --arch oracle_transformer_wmt_en_de --share-all-embeddings \
    --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --lr-scheduler inverse_sqrt \
    --warmup-init-lr 1e-07 --warmup-updates 4000 --lr 0.0007 --min-lr 1e-09 \
    --weight-decay 0.0 --criterion oracle_label_smoothed_cross_entropy --label-smoothing 0.1 \
    --max-tokens 12288 --update-freq 1 --no-progress-bar --log-format json --max-update 80000 \
    --log-interval 10 --save-interval-updates 4000 --keep-interval-updates 10 --save-interval 10000 \
    --seed 1111 --use-epoch-numbers-decay \
    --use-word-level-oracles --decay-k 50 --use-greed-gumbel-noise --gumbel-noise 0.8 \
    --distributed-port 31111 --distributed-world-size 8 --ddp-backend=no_c10d \
    --source-lang en --target-lang de --save-dir $model_dir | tee -a $model_dir/training.log

As Eq.(15) in the paper, the probability of sampling golden words decays with the number of epochs as follows:

In order to save training time, we use the sentence-level oracle method to finetune the best base model.

Setting of the sentence-level oracle for the WMT'14 English->German dataset:

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
data_bin_dir=directory_of_data_bin
model_dir=./ckpt
python train.py $data_bin_dir \
    --arch oracle_transformer_wmt_en_de --share-all-embeddings \
    --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --lr-scheduler inverse_sqrt \                                                                                                                
    --warmup-init-lr 1e-07 --warmup-updates 4000 --lr 0.0007 --min-lr 1e-09 \
    --weight-decay 0.0 --criterion oracle_label_smoothed_cross_entropy --label-smoothing 0.1 \
    --max-tokens 12288 --update-freq 1 --no-progress-bar --log-format json --max-update 40000 \
    --log-interval 10 --save-interval-updates 2000 --keep-interval-updates 10 --save-interval 10000 \
    --seed 1111 --reset-optimizer --reset-meters \
    --use-sentence-level-oracles --decay-k 5800 --use-bleu-gumbel-noise --gumbel-noise 0.5 --oracle-search-beam-size 4 \ 
    --distributed-port 31111 --distributed-world-size 8 --ddp-backend=no_c10d \
    --source-lang en --target-lang de --save-dir $model_dir | tee -a $model_dir/training.log

As Eq.(15) in the paper, the probability of sampling golden words decays with the number of udpates as follows:

NOTE

• The speed of word-level training is almost the same as original transformer.
• Sentence-level training is slower than word-level training.
• --use-epoch-numbers-decay and --decay-k need to be adapted on different training data.
• The prob field in the training log means the decay probability of sampling golden words.

Test training speed and GPU memory usage on iwslt de2en training set

Citation

please cite as:

@inproceedings{zhang2019bridging,
    title = "Bridging the Gap between Training and Inference for Neural Machine Translation",
    author = "Zhang, Wen and Feng, Yang and Meng, Fandong and You, Di and Liu, Qun",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1426",
    doi = "10.18653/v1/P19-1426",
    pages = "4334--4343",
}

Fairseq(-py) is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling and other text generation tasks.

What's New:

• March 2020: Byte-level BPE code released
• February 2020: mBART model and code released
• February 2020: Added tutorial for back-translation
• December 2019: fairseq 0.9.0 released
• November 2019: VizSeq released (a visual analysis toolkit for evaluating fairseq models)
• November 2019: CamemBERT model and code released
• November 2019: BART model and code released
• November 2019: XLM-R models and code released
• September 2019: Nonautoregressive translation code released
• August 2019: WMT'19 models released
• July 2019: fairseq relicensed under MIT license
• July 2019: RoBERTa models and code released
• June 2019: wav2vec models and code released

Features:

Fairseq provides reference implementations of various sequence-to-sequence models, including:

• Convolutional Neural Networks (CNN)
  • Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017)
  • Convolutional Sequence to Sequence Learning (Gehring et al., 2017)
  • Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018)
  • Hierarchical Neural Story Generation (Fan et al., 2018)
  • wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)
• LightConv and DynamicConv models
  • Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)
• Long Short-Term Memory (LSTM) networks
  • Effective Approaches to Attention-based Neural Machine Translation (Luong et al., 2015)
• Transformer (self-attention) networks
  • Attention Is All You Need (Vaswani et al., 2017)
  • Scaling Neural Machine Translation (Ott et al., 2018)
  • Understanding Back-Translation at Scale (Edunov et al., 2018)
  • Adaptive Input Representations for Neural Language Modeling (Baevski and Auli, 2018)
  • Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)
  • RoBERTa: A Robustly Optimized BERT Pretraining Approach (Liu et al., 2019)
  • Facebook FAIR's WMT19 News Translation Task Submission (Ng et al., 2019)
  • Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)
  • Multilingual Denoising Pre-training for Neural Machine Translation (Liu et at., 2020)
  • Neural Machine Translation with Byte-Level Subwords (Wang et al., 2020)
• Non-autoregressive Transformers
  • Non-Autoregressive Neural Machine Translation (Gu et al., 2017)
  • Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement (Lee et al. 2018)
  • Insertion Transformer: Flexible Sequence Generation via Insertion Operations (Stern et al. 2019)
  • Mask-Predict: Parallel Decoding of Conditional Masked Language Models (Ghazvininejad et al., 2019)
  • Levenshtein Transformer (Gu et al., 2019)

Additionally:

• multi-GPU (distributed) training on one machine or across multiple machines
• fast generation on both CPU and GPU with multiple search algorithms implemented:
  • beam search
  • Diverse Beam Search (Vijayakumar et al., 2016)
  • sampling (unconstrained, top-k and top-p/nucleus)
• large mini-batch training even on a single GPU via delayed updates
• mixed precision training (trains faster with less GPU memory on NVIDIA tensor cores)
• extensible: easily register new models, criterions, tasks, optimizers and learning rate schedulers

We also provide pre-trained models for translation and language modeling with a convenient torch.hub interface:

en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de.single_model')
en2de.translate('Hello world', beam=5)
# 'Hallo Welt'

See the PyTorch Hub tutorials for translation and RoBERTa for more examples.

Requirements and Installation

• PyTorch version >= 1.4.0
• Python version >= 3.6
• For training new models, you'll also need an NVIDIA GPU and NCCL
• For faster training install NVIDIA's apex library with the --cuda_ext and --deprecated_fused_adam options

To install fairseq:

pip install fairseq

On MacOS:

CFLAGS="-stdlib=libc++" pip install fairseq

If you use Docker make sure to increase the shared memory size either with --ipc=host or --shm-size as command line options to nvidia-docker run.

Installing from source

To install fairseq from source and develop locally:

git clone https://github.com/pytorch/fairseq
cd fairseq
pip install --editable .

Getting Started

The full documentation contains instructions for getting started, training new models and extending fairseq with new model types and tasks.

Pre-trained models and examples

We provide pre-trained models and pre-processed, binarized test sets for several tasks listed below, as well as example training and evaluation commands.

• Translation: convolutional and transformer models are available
• Language Modeling: convolutional and transformer models are available
• wav2vec: wav2vec large model is available

We also have more detailed READMEs to reproduce results from specific papers:

• Neural Machine Translation with Byte-Level Subwords (Wang et al., 2020)
• Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)
• Levenshtein Transformer (Gu et al., 2019)
• Facebook FAIR's WMT19 News Translation Task Submission (Ng et al., 2019)
• RoBERTa: A Robustly Optimized BERT Pretraining Approach (Liu et al., 2019)
• wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)
• Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)
• Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)
• Understanding Back-Translation at Scale (Edunov et al., 2018)
• Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018)
• Hierarchical Neural Story Generation (Fan et al., 2018)
• Scaling Neural Machine Translation (Ott et al., 2018)
• Convolutional Sequence to Sequence Learning (Gehring et al., 2017)
• Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017)

Join the fairseq community

• Facebook page: https://www.facebook.com/groups/fairseq.users
• Google group: https://groups.google.com/forum/#!forum/fairseq-users

License

fairseq(-py) is MIT-licensed. The license applies to the pre-trained models as well.

Citation

Please cite as:

@inproceedings{ott2019fairseq,
  title = {fairseq: A Fast, Extensible Toolkit for Sequence Modeling},
  author = {Myle Ott and Sergey Edunov and Alexei Baevski and Angela Fan and Sam Gross and Nathan Ng and David Grangier and Michael Auli},
  booktitle = {Proceedings of NAACL-HLT 2019: Demonstrations},
  year = {2019},
}