Effective Approaches to Attention-based Neural Machine Translationsm

Code to train Neural Machine Translation systems as described in our EMNLP paper Effective Approaches to Attention-based Neural Machine Translation.

Features:

• Multi-layer bilingual encoder-decoder models: GPU-enabled.
• Attention mechanisms: global and local models.
• Beam-search decoder: can decode multiple models.
• Other features: dropout, train monolingual language models.
• End-to-end pipeline: scripts to preprocess, compute evaluation scores.

Citations:

If you make use of this code in your research, please cite our paper

@InProceedings{luong-pham-manning:2015:EMNLP,
  author    = {Luong, Minh-Thang  and  Pham, Hieu  and  Manning, Christopher D.},
  title     = {Effective Approaches to Attention-based Neural Machine Translation},
  booktitle = {Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing},
  year      = {2015},
}

• Thang Luong [email protected], 2014, 2015
• With contributions from: Hieu Pham [email protected] -- beam-search decoder.

Files

README.md       - this file
code/           - main Matlab code
  trainLSTM.m: train models
  testLSTM.m: decode models
data/           - toy data
scripts/        - utility scripts

The code directory further divides into sub-directories:

  basic/: define basic functions like sigmoid, prime. It also has an efficient way to aggreate embeddings.
  layers/: we define various layers like attention, LSTM, etc.
  misc/: things that we haven't categorized yet.
  preprocess/: deal with data.
  print/: print results, logs for debugging purposes.

Examples

• Prepare the data

./scripts/run_prepare_data.sh ./data/train.10k.en ./data/valid.100.en ./data/test.100.en 1000 ./output/id.1000
./scripts/run_prepare_data.sh ./data/train.10k.de ./data/valid.100.de ./data/test.100.de 1000 ./output/id.1000

Here, we convert train/valid/test files in text format into integer format that can be handled efficiently in Matlab. The script syntax is:

run_prepare_data.sh <trainFile> <validFile> <testFile> <vocabSize> <outDir>

• Train a bilingual, encoder-decoder model

In Matlab, go into the code/ directory and run:

trainLSTM('../output/id.1000/train.10k', '../output/id.1000/valid.100', '../output/id.1000/test.100', 'de', 'en', '../output/id.1000/train.10k.de.vocab.1000', '../output/id.1000/train.10k.en.vocab.1000', '../output/basic', 'isResume', 0)

This trains a very basic model with all the default settings. We set 'isResume' to 0 so that it will train a new model each time you run the command instead of loading existing models. See trainLSTM.m for more.

(To run directly from your terminal, checkout scripts/train.sh)

The syntax is:

trainLSTM(trainPrefix,validPrefix,testPrefix,srcLang,tgtLang,srcVocabFile,tgtVocabFile,outDir,varargin)
% Arguments:
%   trainPrefix, validPrefix, testPrefix: expect files trainPrefix.srcLang,
%     trainPrefix.tgtLang. Similarly for validPrefix and testPrefix.
%     These data files contain sequences of integers one per line.
%   srcLang, tgtLang: languages, e.g. en, de.
%   srcVocabFile, tgtVocabFile: one word per line.
%   outDir: output directory.
%   varargin: other optional arguments.

The trainer code outputs logs such as:

1, 20, 6.38K, 1, 6.52, gN=11.62

which means: at epoch 1, mini-batches 20, training speed is 6.38K words/s, learning rate is 1, train cost is 6.52, and grad norm is 11.62.

Once in a while, the code will evaluate on the valid and test sets:

# eval 34.40, 2, 144, 6.19K, 1.00, train=4.76, valid=3.63, test=3.54, W_src{1}=0.050 W_tgt{1}=0.081 W_emb_src=0.050 W_emb_tgt=0.056 W_soft=0.076, time=0.57s

which tells us additional information such as the current test perplexity, 34.40, the valid / test costs, and the average abs values of the model paramters.

• Decode

testLSTM('../output/basic/modelRecent.mat', 2, 10, 1, '../output/basic/translations.txt')

Decode with beamSize 2, collect maximum 10 translations, batchSize 1.

Note that the testLSTM implicitly decodes the test file specified during training. To specifiy a different test file, use 'testPrefix', see testLSTM.m for more.

testLSTM('../output/basic/modelRecent.mat', 2, 10, 1, '../output/basic/translations.txt', 'testPrefix', '../output/id.1000/valid.100')

(To run directly from your terminal, checkout scripts/test.sh)

Syntax:

testLSTM(modelFiles, beamSize, stackSize, batchSize, outputFile,varargin)
% Arguments:
%   modelFiles: single or multiple models to decode. Multiple models are
%     separated by commas.
%   beamSize: number of hypotheses kept at each time step.
%   stackSize: number of translations retrieved.
%   batchSize: number of sentences decoded simultaneously. We only ensure
%     accuracy of batchSize = 1 for now.
%   outputFile: output translation file.
%   varargin: other optional arguments.

• Grad check

trainLSTM('', '', '', '', '', '', '', '../output/gradcheck', 'isGradCheck', 1)

Advanced Examples

• Profiling

trainLSTM('../output/id.1000/train.10k', '../output/id.1000/valid.100', '../output/id.1000/test.100', 'de', 'en', '../output/id.1000/train.10k.de.vocab.1000', '../output/id.1000/train.10k.en.vocab.1000', '../output/basic', 'isProfile', 1, 'isResume', 0)

• Train a monolingual language model

trainLSTM('../output/id.1000/train.10k', '../output/id.1000/valid.100','../output/id.1000/test.100', '', 'en', '','../output/id.1000/train.10k.en.vocab.1000', '../output/mono', 'isBi', 0, 'isResume', 0)

• Train multi-layer model with dropout

trainLSTM('../output/id.1000/train.10k', '../output/id.1000/valid.100', '../output/id.1000/test.100', 'de', 'en', '../output/id.1000/train.10k.de.vocab.1000', '../output/id.1000/train.10k.en.vocab.1000', '../output/advanced', 'numLayers', 2, 'lstmSize', 100, 'dropout', 0.8, 'isResume', 0)

We train a 2 stacking LSTM layers with 100 LSTM cells and 100-dim embeddings. Here, 0.8 means the dropout "keep" probability.

• More control of hyperparameters

trainLSTM('../output/id.1000/train.10k', '../output/id.1000/valid.100', '../output/id.1000/test.100', 'de', 'en', '../output/id.1000/train.10k.de.vocab.1000', '../output/id.1000/train.10k.en.vocab.1000', '../output/hypers', 'learningRate', 1.0, 'maxGradNorm', 5, 'initRange', 0.1, 'batchSize', 128, 'numEpoches', 10, 'finetuneEpoch', 5, 'isResume', 0)

Apart from "obvious" hyperparameters, we want to scale our gradients whenever its norm averaged by batch size (128) is greater than 5. After training for 5 epochs, we start halving our learning rate each epoch. To have further control of learning rate schedule, see 'epochFraction' and 'finetuneRate' options in trainLSTM.m

• Train attention model:

trainLSTM('../output/id.1000/train.10k', '../output/id.1000/valid.100', '../output/id.1000/test.100', 'de', 'en', '../output/id.1000/train.10k.de.vocab.1000', '../output/id.1000/train.10k.en.vocab.1000', '../output/attn', 'attnFunc', 1, 'attnOpt', 1, 'isReverse', 1, 'feedInput', 1, 'isResume', 0)

Here, we also use source reversing 'isReverse' and the input feeding approach 'feedInput' as described in the paper. Other attention architectures can be specified as follows:

  % attnFunc=0: no attention.
  %          1: global attention
  %          2: local attention + monotonic alignments
  %          4: local attention  + regression for absolute pos (multiplied distWeights)
  % attnOpt: decide how we generate the alignment weights:
  %          0: location-based
  %          1: content-based, dot product
  %          2: content-based, general dot product
  %          3: content-based, concat Montreal style

'isResume' is set to 0 to avoid loading existing models (done by default), so that you can try different attention architectures.

• More grad checks:

./scripts/run_grad_checks.sh > output/grad_checks.txt 2>&1

Then compare with the provided grad check outputs data/grad_checks.txt. They should look similar.

Note: many different configurations will be run with the run_grad_checks.sh script. For many configuration, we set the 'initRange' to a large value 10, so you will notice the total gradient differences are large. This is to debug subtle mistakes; and if the total diff < 10, you can mostly be assured. We do note that with attnFunc=4, attnOpt=1, the diff is quite large; this is something to be checked though the model seems to work in practice.