Giter Site home page Giter Site logo

icml21_oaxe's Introduction

Order-Agnostic Cross Entropy (OAXE) for Non-Autoregressive Machine Translation

Note

We based this code heavily on the original code of mask-predict and disco.

Download model

Description Dataset Model
CMLM-OAXE [WMT14 English-German] download (.tar.bz2)
CMLM-OAXE [WMT14 German-English] download (.tar.bz2)
CMLM-OAXE [WMT16 English-Romanian] download (.tar.bz2)
CMLM-OAXE [WMT16 Romanian-English] download (.tar.bz2)
CMLM-OAXE [WMT17 English-Chinese] download (.tar.bz2)
CMLM-OAXE [WMT17 Chinese-English] download (.tar.bz2)

Preprocess

TBD

Train

Here we offer the XE finetuning strategy, you have to copy a pretrained CMLM model into your model dir. We adopt the pretrain models from Mask-Predict. It is better to keep update-freq x GPU NUM >= 8.

data_dir = data dir
model_dir = model dir
python3 -u train.py ${data_dir} --arch bert_transformer_seq2seq \
    --criterion oaxe --lr 5e-6  \
    --optimizer adam --adam-betas '(0.9, 0.999)' --adam-eps 1e-6 --task translation_self \
    --max-tokens 16384 --weight-decay 0.01 --dropout 0.2 --encoder-layers 6 --encoder-embed-dim 512 \
    --decoder-layers 6 --decoder-embed-dim 512 --max-source-positions 10000 \
    --max-target-positions 10000 --max-epoch 245 --seed 1 \
    --save-dir ${model_dir} \
    --keep-last-epochs 10 --share-all-embeddings \
    --keep-interval-updates 1 --fp16 \
    --update-freq 1 --ddp-backend=no_c10d \
    --warmup-init-lr 1e-7 --min-lr 1e-9 --lr-scheduler inverse_sqrt --warmup-updates 500 \
    --label-smoothing 0.1 --reset-optimizer --skip 0.15 \

Evaluation

TBD

License

Following MASK-PREDICT, our code is also CC-BY-NC 4.0. The license applies to the pre-trained models as well.

Citation

Please cite as:

@inproceedings{Du2021OAXE,
  title = {Order-Agnostic Cross Entropy for Non-Autoregressive Machine Translation},
  author = {Cunxiao Du and Zhaopeng Tu and Jing Jiang},
  booktitle = {Proc. of ICML},
  year = {2021},
}

icml21_oaxe's People

Contributors

nonvolatilememory avatar

Stargazers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar

Watchers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar

icml21_oaxe's Issues

Code diff to original fairseq

Can you share the code-diff to the original fairseq implementation so it's clear which parts were added/adapted? This would make things a lot easier to integrate with other papers. Thanks!

instruction on training?

can I ask whether it is possible to provide any instruction for training or any other change in fairseq training instructions?

about code of ngram-OAXE

I notice that the code of your new paper "ngram-OAXE : Phrase-Based Order-Agnostic Cross Entropy for Non-Autoregressive Machine Translation" is invalid. Could you fix it?

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.