Skip to content

Latest commit

 

History

History
146 lines (126 loc) · 10.4 KB

File metadata and controls

146 lines (126 loc) · 10.4 KB

Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)

This page contains pointers to pre-trained models as well as instructions on how to train new models for our paper

Citation:

@inproceedings{wu2018pay,
  title = {Pay Less Attention with Lightweight and Dynamic Convolutions},
  author = {Felix Wu and Angela Fan and Alexei Baevski and Yann Dauphin and Michael Auli},
  booktitle = {International Conference on Learning Representations},
  year = {2019},
  url = {https://arxiv.org/abs/1901.10430},
}

Translation

Pre-trained models

For some datasets we release models without GLUs which are faster at inference.

Description Dataset Model Test set(s)
LightConv (without GLUs) IWSLT14 German-English download (.tar.bz2) IWSLT14 test:
download (.tar.bz2)
DynamicConv (without GLUs) IWSLT14 German-English download (.tar.bz2) IWSLT14 test:
download (.tar.bz2)
LightConv (without GLUs) WMT16 English-German download (.tar.bz2) newstest2014 (shared vocab):
download (.tar.bz2)
DynamicConv (without GLUs) WMT16 English-German download (.tar.bz2) newstest2014 (shared vocab):
download (.tar.bz2)
LightConv WMT16 English-German download (.tar.bz2) newstest2014 (shared vocab):
download (.tar.bz2)
DynamicConv WMT16 English-German download (.tar.bz2) newstest2014 (shared vocab):
download (.tar.bz2)
LightConv WMT14 English-French download (.tar.bz2) newstest2014:
download (.tar.bz2)
DynamicConv WMT14 English-French download (.tar.bz2) newstest2014:
download (.tar.bz2)
LightConv WMT17 Chinese-English download (.tar.bz2) newstest2017:
download (.tar.bz2)
DynamicConv WMT17 Chinese-English download (.tar.bz2) newstest2017:
download (.tar.bz2)
LightConv (CUDA module) WMT17 English-German download (.tar.gz) newstest2014 (shared vocab):
download (.tar.bz2)
DynamicConv (CUDA module) WMT17 English-German download (.tar.gz) newstest2014:
download (.tar.bz2)

Memory-Efficient CUDA Kernels

Since the PyTorch implementations of Light/Dynamic conv are quite memory intensive, we have developed CUDA kernels that implement the light and dynamic convolution operator in a memory-efficient and performant manner. For large sequence lengths, these kernels save about 50% memory compared to the PyTorch equivalent.

To install the kernels, use the commands below. Once installed, they will automatically be used in place of the PyTorch implementations whenever a light or dynamic convolution is used.

# to install lightconv
cd fairseq/modules/lightconv_layer
python cuda_function_gen.py
python setup.py install

# to install dynamicconv
cd fairseq/modules/dynamicconv_layer
python cuda_function_gen.py
python setup.py install

Preprocessing the training datasets

Please follow the instructions in examples/translation/README.md to preprocess the data.

Training and evaluation options:

To use the model without GLU, please set --encoder-glu 0 --decoder-glu 0. For LightConv, please use --encoder-conv-type lightweight --decoder-conv-type lightweight, otherwise the default is DynamicConv. For best BLEU results, lenpen may need to be manually tuned.

To use the CUDA kernels, first install the PyTorch modules using the commands below

# to install lightconv
python fairseq/modules/lightconv_layer/cuda_function_gen.py
python fairseq/modules/lightconv_layer/setup.py install

# to install dynamicconv
python fairseq/modules/dynamicconv_layer/cuda_function_gen.py
python fairseq/modules/dynamicconv_layer/setup.py install

Once the CUDA modules are installed, they will automatically be used instead of the PyTorch modules.

IWSLT14 De-En

Training and evaluating DynamicConv (without GLU) on a GPU:

# Training
SAVE="save/dynamic_conv_iwslt"
mkdir -p $SAVE 
CUDA_VISIBLE_DEVICES=0 $(which fairseq-train) data-bin/iwslt14.tokenized.de-en \
    --clip-norm 0 --optimizer adam --lr 0.0005 \
    --source-lang de --target-lang en --max-tokens 4000 --no-progress-bar \
    --log-interval 100 --min-lr '1e-09' --weight-decay 0.0001 \
    --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
    --lr-scheduler inverse_sqrt \
    --ddp-backend=no_c10d \
    --max-update 50000 --warmup-updates 4000 --warmup-init-lr '1e-07' \
    --adam-betas '(0.9, 0.98)' --keep-last-epochs 10 \
    -a lightconv_iwslt_de_en --save-dir $SAVE \
    --dropout 0.3 --attention-dropout 0.1 --weight-dropout 0.1 \
    --encoder-glu 0 --decoder-glu 0
python scripts/average_checkpoints.py --inputs $SAVE \
    --num-epoch-checkpoints 10 --output "${SAVE}/checkpoint_last10_avg.pt"

# Evaluation
CUDA_VISIBLE_DEVICES=0 fairseq-generate data-bin/iwslt14.tokenized.de-en --path "${SAVE}/checkpoint_last10_avg.pt" --batch-size 128 --beam 4 --remove-bpe --lenpen 1 --gen-subset test --quiet 

WMT16 En-De

Training and evaluating DynamicConv (with GLU) on WMT16 En-De using cosine scheduler on one machine with 8 V100 GPUs:

# Training
SAVE="save/dynamic_conv_wmt16en2de"
mkdir -p $SAVE
python -m torch.distributed.launch --nproc_per_node 8 $(which fairseq-train) \
    data-bin/wmt16_en_de_bpe32k --fp16  --log-interval 100 --no-progress-bar \
    --max-update 30000 --share-all-embeddings --optimizer adam \
    --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --weight-decay 0.0 \
    --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
    --min-lr 1e-09 --update-freq 16 --attention-dropout 0.1 --keep-last-epochs 10 \
    --ddp-backend=no_c10d --max-tokens 3584 \
    --lr-scheduler cosine --warmup-init-lr 1e-7 --warmup-updates 10000 \
    --lr-shrink 1 --max-lr 0.001 --lr 1e-7 --min-lr 1e-9 --warmup-init-lr 1e-07 \
    --t-mult 1 --lr-period-updates 20000 \
    --arch lightconv_wmt_en_de_big --save-dir $SAVE \
    --dropout 0.3 --attention-dropout 0.1 --weight-dropout 0.1 \
    --encoder-glu 1 --decoder-glu 1

# Evaluation
CUDA_VISIBLE_DEVICES=0 fairseq-generate data-bin/wmt16.en-de.joined-dict.newstest2014 --path "${SAVE}/checkpoint_best.pt" --batch-size 128 --beam 5 --remove-bpe --lenpen 0.5 --gen-subset test > wmt16_gen.txt
bash scripts/compound_split_bleu.sh wmt16_gen.txt

WMT14 En-Fr

Training DynamicConv (with GLU) on WMT14 En-Fr using cosine scheduler on one machine with 8 V100 GPUs:

# Training
SAVE="save/dynamic_conv_wmt14en2fr"
mkdir -p $SAVE
python -m torch.distributed.launch --nproc_per_node 8 $(which fairseq-train) \
    data-bin/wmt14_en_fr --fp16  --log-interval 100 --no-progress-bar \
    --max-update 30000 --share-all-embeddings --optimizer adam \
    --adam-betas '(0.9, 0.98)' --clip-norm 0.0 --weight-decay 0.0 \
    --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
    --min-lr 1e-09 --update-freq 16 --attention-dropout 0.1 --keep-last-epochs 10 \
    --ddp-backend=no_c10d --max-tokens 3584 \
    --lr-scheduler cosine --warmup-init-lr 1e-7 --warmup-updates 10000 \
    --lr-shrink 1 --max-lr 0.001 --lr 1e-7 --min-lr 1e-9 --warmup-init-lr 1e-07 \
    --t-mult 1 --lr-period-updates 70000 \
    --arch lightconv_wmt_en_fr_big --save-dir $SAVE \
    --dropout 0.1 --attention-dropout 0.1 --weight-dropout 0.1 \
    --encoder-glu 1 --decoder-glu 1

# Evaluation
CUDA_VISIBLE_DEVICES=0 fairseq-generate data-bin/wmt14.en-fr.joined-dict.newstest2014 --path "${SAVE}/checkpoint_best.pt" --batch-size 128 --beam 5 --remove-bpe --lenpen 0.9 --gen-subset test