DINOISER: Diffused Conditional Sequence Learning by Manipulating Noises

This repository contains the official implementation of paper DINOISER: Diffused Conditional Sequence Learning by Manipulating Noises.

---

Dependencies

The code is implemented with fairseq. To setup the dependencies, run

pip3 install -r requirements.txt

Data Preparation

The following commands utilize the script from fairseq's examples to prepare IWSLT14 En<->De data.

mkdir iwslt14
cd iwslt14

wget -O - https://raw.githubusercontent.com/facebookresearch/fairseq/main/examples/translation/prepare-iwslt14.sh | bash

fairseq-preprocess -s en -t de \
    --trainpref iwslt14.tokenized.de-en/train \
    --validpref iwslt14.tokenized.de-en/valid \
    --testpref iwslt14.tokenized.de-en/test \
    --destdir iwslt14.en-de.real.bin
    --workers 32

fairseq-preprocess -s de -t en \
    --trainpref iwslt14.tokenized.de-en/train \
    --validpref iwslt14.tokenized.de-en/valid \
    --testpref iwslt14.tokenized.de-en/test \
    --destdir iwslt14.de-en.real.bin
    --workers 32

cd ..

For WMT14 En<->De and WMT16 En<->Ro, we obtain the preprocessed data from the links that Fully-NAT provides.

wget https://dl.fbaipublicfiles.com/nat/fully_nat/datasets/wmt14.en-de.zip
wget https://dl.fbaipublicfiles.com/nat/fully_nat/datasets/wmt16.ro-en.zip
unzip wmt14.en-de -d wmt14
unzip wmt16.ro-en -d wmt16

Training and Evaluation

Use the follwoing scripts to train the models for machin translation.

cd scripts
CUDA_VISIBLE_DEVICES=0 bash train_mt.sh -d iwslt14 -b ../iwslt14/iwslt14.de-en.real.bin -o ../outputs/iwslt14.de-en
CUDA_VISIBLE_DEVICES=0 bash train_mt.sh -d iwslt14 -b ../iwslt14/iwslt14.en-de.real.bin -o ../outputs/iwslt14.en-de
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash train_mt.sh -d wmt14 -b ../wmt14/wmt14.de-en.real.bin -o ../outputs/wmt14.de-en
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash train_mt.sh -d wmt14 -b ../wmt14/wmt14.en-de.real.bin -o ../outputs/wmt14.en-de
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash train_mt.sh -d wmt16 -b ../wmt16/wmt16.ro-en.real.bin -o ../outputs/wmt16.ro-en
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash train_mt.sh -d wmt16 -b ../wmt16/wmt16.en-ro.real.bin -o ../outputs/wmt16.en-ro

• -d: Identifier for dataset. We use it to setup dataset specific arguments (e.g., model architecture and batch size) in the script.
• -b: Directory of the binarize data.
• -o: Directory for saving checkpoints and logs.
• You can append -e "--bf16" to accelerate training if your devices support bfloat16.

Then, use the following scipts to make inferences and evaluate the models.

bash eval_mt.sh -b ../iwslt14/iwslt14.de-en.real.bin -l 10 -m 5 -c ../outputs/iwslt14.de-en/checkpoints_length/checkpoint_best.pt -o ../outputs/iwslt14.de-en -t en
bash eval_mt.sh -b ../iwslt14/iwslt14.en-de.real.bin -l 10 -m 5 -c ../outputs/iwslt14.en-de/checkpoints_length/checkpoint_best.pt -o ../outputs/iwslt14.en-de -t de
bash eval_mt.sh -b ../wmt14/wmt14.de-en.real.bin -l 10 -m 5 -c ../outputs/wmt14.de-en/checkpoints_length/checkpoint_best.pt -o ../outputs/wmt14.de-en -t en
bash eval_mt.sh -b ../wmt14/wmt14.en-de.real.bin -l 10 -m 5 -c ../outputs/wmt14.en-de/checkpoints_length/checkpoint_best.pt -o ../outputs/wmt14.en-de -t de
bash eval_mt.sh -b ../wmt16/wmt16.de-en.real.bin -l 10 -m 5 -c ../outputs/wmt16.ro-en/checkpoints_length/checkpoint_best.pt -o ../outputs/wmt16.ro-en -t en
bash eval_mt.sh -b ../wmt16/wmt16.de-en.real.bin -l 10 -m 5 -c ../outputs/wmt15.de-en/checkpoints_length/checkpoint_best.pt -o ../outputs/wmt16.en-ro -t ro

Arguments

• -b: Directory of the binarize data.
• -l: Length beam.
• -m: Number of MBR candidates for each length beam.
• -c: The checkpoint to be evaluated.
• -o: Directory to place the generation result.
• -t: The language of the target. This affects the tokenizer for computing sacrebleu.

Please check the scripts for more details.

Checkpoints

Dataset	Checkpoint
IWSLT14 De->En	Download
IWSLT14 En->De	Download
WMT14 De->En	Download
WMT14 En->De	Download
WMT16 Ro->En	Download
WMT16 En->Ro	Download

Citation

@article{ye2023dinoiser,
  title={DINOISER: Diffused Conditional Sequence Learning by Manipulating Noises},
  author={Ye, Jiasheng and Zheng, Zaixiang and Bao, Yu and Qian, Lihua and Wang, Mingxuan},
  journal={arXiv preprint arXiv:2302.10025},
  year={2023}
}