Attention-based encoder-decoder model for neural machine translation
This package is based on the dl4mt-tutorial by Kyunghyun Cho et al. ( https://github.com/nyu-dl/dl4mt-tutorial ). It was used to produce top-scoring systems at the WMT 16 shared translation task.
The changes to Nematus include:
-
new architecture variants for better performance:
- arbitrary input features (factored neural machine translation) http://www.statmt.org/wmt16/pdf/W16-2209.pdf
- deep models (Zhou et al., 2016; Wu et al., 2016; Miceli Barone et al., 2017) https://arxiv.org/abs/1606.04199 https://arxiv.org/abs/1609.08144 https://arxiv.org/abs/1707.07631
- dropout on all layers (Gal, 2015) http://arxiv.org/abs/1512.05287
- tied embeddings (Press and Wolf, 2016) https://arxiv.org/abs/1608.05859
- layer normalisation (Ba et al, 2016) https://arxiv.org/abs/1607.06450
- weight normalisation (Salimans and Kingma, 2016) https://arxiv.org/abs/1602.07868
-
improvements to scoring and decoding:
- ensemble decoding (and new translation API to support it)
- n-best output for decoder
- scripts for scoring (given parallel corpus) and rescoring (of n-best output)
-
usability improvements:
- command line interface for training
- more output options (attention weights; word-level probabilities) and visualization scripts
- execute arbitrary validation scripts (for BLEU early stopping)
- vocabulary files and model parameters are stored in JSON format (backward-compatible loading)
- server mode
see changelog for more info
For general support requests, there is a Google Groups mailing list at https://groups.google.com/d/forum/nematus-support . You can also send an e-mail to nematus-support@googlegroups.com .
Nematus requires the following packages:
- Python 2.7
- numpy
- Theano >= 0.7 (and its dependencies).
we recommend executing the following command in a Python virtual environment:
pip install numpy numexpr cython tables theano bottle bottle-log tornado
the following packages are optional, but highly recommended
- CUDA >= 7 (only GPU training is sufficiently fast)
- cuDNN >= 4 (speeds up training substantially)
you can run Nematus locally. To install it, execute python setup.py install
You can also create docker image by running following command, where you change suffix
to either cpu
or gpu
:
docker build -t nematus-docker -f Dockerfile.suffix .
To run a CPU docker instance with the current working directory shared with the Docker container, execute:
docker run -v `pwd`:/playground -it nematus-docker
For GPU you need to have nvidia-docker installed and run:
nvidia-docker run -v `pwd`:/playground -it nematus-docker
Training speed depends heavily on having appropriate hardware (ideally a recent NVIDIA GPU), and having installed the appropriate software packages.
To test your setup, we provide some speed benchmarks with `test/test_train.sh', on an Intel Xeon CPU E5-2620 v3, with a Nvidia GeForce GTX 1080 GPU and CUDA 8.0:
CPU, theano 0.8.2:
THEANO_FLAGS=mode=FAST_RUN,floatX=float32,device=cpu ./test_train.sh
2.37 sentences/s
GPU, no CuDNN, theano 0.8.2:
THEANO_FLAGS=mode=FAST_RUN,floatX=float32,device=gpu ./test_train.sh
71.62 sentences/s
GPU, CuDNN 5.1, theano 0.8.2:
THEANO_FLAGS=mode=FAST_RUN,floatX=float32,device=gpu ./test_train.sh
139.73 sentences/s
GPU, CuDNN 5.1, theano 0.9.0dev5.dev-d5520e:
THEANO_FLAGS=mode=FAST_RUN,floatX=float32,device=gpu ./test_train.sh
173.15 sentences/s
GPU, CuDNN 5.1, theano 0.9.0dev5.dev-d5520e, new GPU backend:
THEANO_FLAGS=mode=FAST_RUN,floatX=float32,device=cuda ./test_train.sh
209.21 sentences/s
GPU, float16, CuDNN 5.1, theano 0.9.0-RELEASE, new GPU backend:
222.28 sentences/s
See SPEED.md for more benchmark results on different hardware and hyperparameter configurations.
execute nematus/nmt.py to train a model.
parameter | description |
---|---|
--datasets PATH PATH | parallel training corpus (source and target) |
--dictionaries PATH [PATH ...] | network vocabularies (one per source factor, plus target vocabulary) |
--model PATH | model file name (default: model.npz) |
--deep_fusion_lm PATH | deep fusion language model file name (default: None) |
--saveFreq INT | save frequency (default: 30000) |
--reload | load existing model (if '--model' points to existing model) |
--overwrite | write all models to same file |
--multilingual | Use a multilingual data iterator to iterate over multiple sets of parallel sentences in equal batches (requires language_source_target to be set, default: False) |
--language_source_target | Target languages and file names for multilingual NMT training (For example [('ru','multi.en', 'multi.ru'),('et','multi.en', 'multi.et')] , default: None) |
parameter | description |
---|---|
--dim_word INT | embedding layer size (default: 512) |
--dim INT | hidden layer size (default: 1000) |
--n_words_src INT | source vocabulary size (default: None) |
--n_words INT | target vocabulary size (default: None) |
--factors INT | number of input factors (default: 1) |
--dim_per_factor INT [INT ...] | list of word vector dimensionalities (one per factor): '--dim_per_factor 250 200 50' for total dimensionality of 500 (default: None) |
--use_dropout | use dropout layer (default: False) |
--dropout_embedding FLOAT | dropout for input embeddings (0: no dropout) (default: 0.2) |
--dropout_hidden FLOAT | dropout for hidden layer (0: no dropout) (default: 0.2) |
--dropout_source FLOAT | dropout source words (0: no dropout) (default: 0) |
--dropout_target FLOAT | dropout target words (0: no dropout) (default: 0) |
--layer_normalisation | use layer normalisation (default: False) |
--weight_normalisation | use weight normalisation (default: False) |
--tie_decoder_embeddings | tie the input embeddings of the decoder with the softmax output embeddings |
--tie_encoder_decoder_embeddings | tie the input embeddings of the encoder and the decoder (first factor only). Source and target vocabulary size must the same |
--encoder | encoder cell type (default: gru) |
--enc_depth INT | number of encoder layers (default: 1) |
--enc_depth_bidirectional | number of bidirectional encoder layers; if enc_depth is greater, remaining layers are unidirectional; by default, all layers are bidirectional. |
--decoder | type of recurrent layer for first decoder layer (default: gru_cond |
--decoder_deep | type of recurrent layer for decoder layers after the first (default: gru) |
--dec_depth INT | number of decoder layers (default: 1) |
--dec_deep_context | pass context vector (from first layer) to deep decoder layers |
--enc_recurrence_transition_depth | number of GRU transition operations applied in an encoder layer (default: 1) |
--dec_base_recurrence_transition_depth | number of GRU transition operations applied in first decoder layer (default: 2) |
--dec_high_recurrence_transition_depth | number of GRU transition operations applied in decoder layers after the first (default: 1) |
--concatenate_lm_decoder | concatenate LM state and decoder state (deep fusion) |
parameter | description |
---|---|
--maxlen INT | maximum sequence length (default: 100) |
--optimizer {adam,adadelta,rmsprop,sgd} | optimizer (default: adam) |
--batch_size INT | minibatch size (default: 80) |
--max_epochs INT | maximum number of epochs (default: 5000) |
--finish_after INT | maximum number of updates (minibatches) (default: 10000000) |
--decay_c FLOAT | L2 regularization penalty (default: 0) |
--map_decay_c FLOAT | MAP-L2 regularization penalty towards original weights (default: 0) |
--prior_model STR | Prior model for MAP-L2 regularization. Unless using "--reload", this will also be used for initialization. |
--clip_c FLOAT | gradient clipping threshold (default: 1) |
--lrate FLOAT | learning rate (default: 0.0001) |
--no_shuffle | disable shuffling of training data (for each epoch) |
--no_sort_by_length | do not sort sentences in maxibatch by length |
--maxibatch_size INT | size of maxibatch (number of minibatches that are sorted by length) (default: 20) |
--objective {CE,MRT,RAML} | training objective. CE: cross-entropy minimization (default); MRT: Minimum Risk Training (https://www.aclweb.org/anthology/P/P16/P16-1159.pdf); RAML: Reward Augmented Maximum Likelihood (https://arxiv.org/pdf/1609.00150.pdf) |
parameter | description |
---|---|
--valid_datasets PATH PATH | parallel validation corpus (source and target) |
--valid_batch_size INT | validation minibatch size (default: 80) |
--validFreq INT | validation frequency (default: 10000) |
--patience INT | early stopping patience (default: 10) |
--anneal_restarts INT | when patience runs out, restart training INT times with annealed learning rate (default: 0) |
--anneal_decay FLOAT | learning rate decay on each restart (default: 0.5) |
--external_validation_script PATH | location of validation script (to run your favorite metric for validation) (default: None) |
parameter | description |
---|---|
--dispFreq INT | display loss after INT updates (default: 1000) |
--sampleFreq INT | display some samples after INT updates (default: 10000) |
parameter | description |
---|---|
--mrt_alpha FLOAT | MRT alpha (default: 0.005) |
--mrt_samples INT | samples per source sentence (default: 100) |
--mrt_samples_meanloss INT | draw n independent samples to calculate mean loss (which is subtracted from loss) (default: 10) |
--mrt_loss STR | loss used in MRT (default: SENTENCEBLEU n=4) |
--mrt_reference | add reference to MRT samples. |
--mrt_ml_mix | mix in ML objective in MRT training with this scaling factor (default: 0) |
parameter | description |
---|---|
--raml_tau FLOAT | RAML tau (default: 0.85) |
--raml_samples INT | samples per source sentence (default: 1) |
--raml_reward {hamming_distance,edit_distance,bleu} | reward for RAML sampling |
more instructions to train a model, including a sample configuration and preprocessing scripts, are provided in https://github.com/rsennrich/wmt16-scripts
parameter | description |
---|---|
-k K | Beam size (default: 5)) |
-p P | Number of processes (default: 5)) |
-n | Normalize scores by sentence length |
-v | verbose mode. |
--models MODELS [MODELS ...], -m MODELS [MODELS ...] | model to use. Provide multiple models (with same vocabulary) for ensemble decoding |
--input PATH, -i PATH | Input file (default: standard input) |
--output PATH, -o PATH | Output file (default: standard output) |
--output_alignment PATH, -a PATH | Output file for alignment weights (default: standard output) |
--json_alignment | Output alignment in json format |
--n-best | Write n-best list (of size k) |
--suppress-unk | Suppress hypotheses containing UNK. |
--print-word-probabilities, -wp | Print probabilities of each word |
--search_graph, -sg | Output file for search graph visualisation. File format is determined by file name, e.g., PDF for search_graph.pdf |
--device-list, -dl | User specified device list for multi-processing decoding. For example: --device-list gpu0 gpu1 gpu2 |
parameter | description |
---|---|
-b B | Minibatch size (default: 80)) |
-n | Normalize scores by sentence length |
-v | verbose mode. |
--models MODELS [MODELS ...], -m MODELS [MODELS ...] | model to use. Provide multiple models (with same vocabulary) for ensemble decoding |
--source PATH, -s PATH | Source text file |
--target PATH, -t PATH | Target text file |
--output PATH, -o PATH | Output file (default: standard output) |
--walign, -w | Whether to store the alignment weights or not. If specified, weights will be saved in .alignment.json |
The n-best list is assumed to have the same format as Moses:
sentence-ID (starting from 0) ||| translation ||| scores
new scores will be appended to the end. rescore.py
has the same arguments as score.py
, with the exception of this additional parameter:
parameter | description |
---|---|
--input PATH, -i PATH | Input n-best list file (default: standard input) |
sample models, and instructions on using them for translation, are provided in the test
directory, and at http://statmt.org/rsennrich/wmt16_systems/
-
The pipe symbol "|" serves as a factor separator and should not otherwise appear in the text. See the documentation on factors for more information. Using the Moses tokenizer also escapes pipe symbols.
-
Support for float16 may not be fully functional or efficient using depending on the Theano version and GPU model. If you use float16 for training, consider using a lower learning rate for increased numerical stability.
if you use Nematus, please cite the following paper:
Rico Sennrich, Orhan Firat, Kyunghyun Cho, Alexandra Birch, Barry Haddow, Julian Hitschler, Marcin Junczys-Dowmunt, Samuel Läubli, Antonio Valerio Miceli Barone, Jozef Mokry and Maria Nadejde (2017): Nematus: a Toolkit for Neural Machine Translation. In Proceedings of the Software Demonstrations of the 15th Conference of the European Chapter of the Association for Computational Linguistics, Valencia, Spain, pp. 65-68.
@InProceedings{sennrich-EtAl:2017:EACLDemo,
author = {Sennrich, Rico and Firat, Orhan and Cho, Kyunghyun and Birch, Alexandra and Haddow, Barry and Hitschler, Julian and Junczys-Dowmunt, Marcin and L\"{a}ubli, Samuel and Miceli Barone, Antonio Valerio and Mokry, Jozef and Nadejde, Maria},
title = {Nematus: a Toolkit for Neural Machine Translation},
booktitle = {Proceedings of the Software Demonstrations of the 15th Conference of the European Chapter of the Association for Computational Linguistics},
month = {April},
year = {2017},
address = {Valencia, Spain},
publisher = {Association for Computational Linguistics},
pages = {65--68},
url = {http://aclweb.org/anthology/E17-3017}
}
the code is based on the following model:
Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio (2015): Neural Machine Translation by Jointly Learning to Align and Translate, Proceedings of the International Conference on Learning Representations (ICLR).
please refer to the Nematus paper for a description of implementation differences
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreements 645452 (QT21), 644333 (TraMOOC), 644402 (HimL) and 688139 (SUMMA).