Sequence-to-Sequence Learning with Attentional Neural Networks
UPDATE: Check-out the beta release of OpenNMT a fully supported feature-complete rewrite of seq2seq-attn. Seq2seq-attn will remain supported, but new features and optimizations will focus on the new codebase.
Torch implementation of a standard sequence-to-sequence model with (optional) attention where the encoder-decoder are LSTMs. Encoder can be a bidirectional LSTM. Additionally has the option to use characters (instead of input word embeddings) by running a convolutional neural network followed by a highway network over character embeddings to use as inputs.
The attention model is from Effective Approaches to Attention-based Neural Machine Translation, Luong et al. EMNLP 2015. We use the global-general-attention model with the input-feeding approach from the paper. Input-feeding is optional and can be turned off.
The character model is from Character-Aware Neural Language Models, Kim et al. AAAI 2016.
There are a lot of additional options on top of the baseline model, mainly thanks to the fantastic folks at SYSTRAN. Specifically, there are functionalities which implement:
- Effective Approaches to Attention-based Neural Machine Translation. Luong et al., EMNLP 2015.
- Character-based Neural Machine Translation. Costa-Jussa and Fonollosa, ACL 2016.
- Compression of Neural Machine Translation Models via Pruning. See et al., COLING 2016.
- Sequence-Level Knowledge Distillation. Kim and Rush., EMNLP 2016.
- Deep Recurrent Models with Fast Forward Connections for Neural Machine Translation. Zhou et al, TACL 2016.
- Guided Alignment Training for Topic-Aware Neural Machine Translation. Chen et al., arXiv:1607.01628.
- Linguistic Input Features Improve Neural Machine Translation. Senrich et al., arXiv:1606.02892
See below for more details on how to use them.
This project is maintained by Yoon Kim. Feel free to post any questions/issues on the issues page.
You will need the following packages:
GPU usage will additionally require:
If running the character model, you should also install:
We are going to be working with some example data in
First run the data-processing code
python preprocess.py --srcfile data/src-train.txt --targetfile data/targ-train.txt --srcvalfile data/src-val.txt --targetvalfile data/targ-val.txt --outputfile data/demo
This will take the source/target train/valid files (
src-val.txt, targ-val.txt) and make some hdf5 files to be consumed by Lua.
demo.src.dict: Dictionary of source vocab to index mappings.
demo.targ.dict: Dictionary of target vocab to index mappings.
demo-train.hdf5: hdf5 containing the train data.
demo-val.hdf5: hdf5 file containing the validation data.
*.dict files will be needed when predicting on new data.
Now run the model
th train.lua -data_file data/demo-train.hdf5 -val_data_file data/demo-val.hdf5 -savefile demo-model
This will run the default model, which consists of a 2-layer LSTM with 500 hidden units
on both the encoder/decoder.
You can also add
-gpuid 1 to use (say) GPU 1 in the cluster.
Now you have a model which you can use to predict on new data. To do this we are going to be running beam search
th evaluate.lua -model demo-model_final.t7 -src_file data/src-val.txt -output_file pred.txt -src_dict data/demo.src.dict -targ_dict data/demo.targ.dict
This will output predictions into
pred.txt. The predictions are going to be quite terrible,
as the demo dataset is small. Try running on some larger datasets! For example you can download
millions of parallel sentences for translation
Preprocessing options (
srcvocabsize, targetvocabsize: Size of source/target vocabularies. This is constructed by taking the top X most frequent words. Rest are replaced with special UNK tokens.
srcfile, targetfile: Path to source/target training data, where each line represents a single source/target sequence.
srcvalfile, targetvalfile: Path to source/target validation data.
batchsize: Size of each mini-batch.
seqlength: Maximum sequence length (sequences longer than this are dropped).
outputfile: Prefix of the output file names.
maxwordlength: For the character models, words are truncated (if longer than maxwordlength) or zero-padded (if shorter) to
chars: If 1, construct the character-level dataset as well. This might take up a lot of space depending on your data size, so you may want to break up the training data into different shards.
srcvocabfile, targetvocabfile: If working with a preset vocab, then including these paths will ignore the
unkfilter: Ignore sentences with too many UNK tokens. Can be an absolute count limit (if > 1) or a proportional limit (0 < unkfilter < 1).
shuffle: Shuffle sentences.
alignvalfile: If provided with filenames that contain 'Pharaoh' format alignment on the train and validation data, source-to-target alignments are stored in the dataset.
Training options (
data_file, val_data_file: Path to the training/validation
*.hdf5files created from running
savefile: Savefile name (model will be saved as
save_everyepoch where X is the X-th epoch and PPL is the validation perplexity at the epoch.
num_shards: If the training data has been broken up into different shards, then this is the number of shards.
train_from: If training from a checkpoint then this is the path to the pre-trained model.
num_layers: Number of layers in the LSTM encoder/decoder (i.e. number of stacks).
rnn_size: Size of LSTM hidden states.
word_vec_size: Word embedding size.
attn: If = 1, use attention over the source sequence during decoding. If = 0, then it uses the last hidden state of the encoder as the context at each time step.
brnn: If = 1, use a bidirectional LSTM on the encoder side. Input embeddings (or CharCNN if using characters) are shared between the forward/backward LSTM, and hidden states of the corresponding forward/backward LSTMs are added to obtain the hidden representation for that time step.
use_chars_enc: If = 1, use characters on the encoder side (as inputs).
use_chars_dec: If = 1, use characters on the decoder side (as inputs).
reverse_src: If = 1, reverse the source sequence. The original sequence-to-sequence paper found that this was crucial to achieving good performance, but with attention models this does not seem necessary. Recommend leaving it to 0.
init_dec: Initialize the hidden/cell state of the decoder at time 0 to be the last hidden/cell state of the encoder. If 0, the initial states of the decoder are set to zero vectors.
input_feed: If = 1, feed the context vector at each time step as additional input (via concatenation with the word embeddings) to the decoder.
multi_attn: If > 0, then use a multi-attention on this layer of the decoder. For example, if
num_layers = 3and
multi_attn = 2, then the model will do an attention over the source sequence on the second layer (and use that as input to the third layer) and the penultimate layer. We've found that this did not really improve performance on translation, but may be helpful for other tasks where multiple attentional passes over the source sequence are required (e.g. for more complex reasoning tasks).
res_net: Use residual connections between LSTM stacks whereby the input to the l-th LSTM layer of the hidden state of the l-1-th LSTM layer summed with hidden state of the l-2th LSTM layer. We didn't find this to really help in our experiments.
Below options only apply if using the character model.
char_vec_size: If using characters, size of the character embeddings.
kernel_width: Size (i.e. width) of the convolutional filter.
num_kernels: Number of convolutional filters (feature maps). So the representation from characters will have this many dimensions.
num_highway_layers: Number of highway layers in the character composition model.
To build a model with guided alignment (implemented similarly to Guided Alignment Training for Topic-Aware Neural Machine Translation (Chen et al. 2016)):
guided_alignment: If 1, use external alignments to guide the attention weights
guided_alignment_weight: weight for guided alignment criterion
guided_alignment_decay: decay rate per epoch for alignment weight
epochs: Number of training epochs.
start_epoch: If loading from a checkpoint, the epoch from which to start.
param_init: Parameters of the model are initialized over a uniform distribution with support
optim: Optimization method, possible choices are 'sgd', 'adagrad', 'adadelta', 'adam'. For seq2seq I've found vanilla SGD to work well but feel free to experiment.
learning_rate: Starting learning rate. For 'adagrad', 'adadelta', and 'adam', this is the global learning rate. Recommended settings vary based on
optim: sgd (
learning_rate = 1), adagrad (
learning_rate = 0.1), adadelta (
learning_rate = 1), adam (
learning_rate = 0.1).
layer_lrs: Comma-separated learning rates for encoder, decoder, and generator when using 'adagrad', 'adadelta', or 'adam' for 'optim' option. Layer-specific learning rates cannot currently be used with sgd.
max_grad_norm: If the norm of the gradient vector exceeds this, renormalize to have its norm equal to
dropout: Dropout probability. Dropout is applied between vertical LSTM stacks.
lr_decay: Decay learning rate by this much if (i) perplexity does not decrease on the validation set or (ii) epoch has gone past the
start_decay_at: Start decay after this epoch.
curriculum: For this many epochs, order the minibatches based on source sequence length. (Sometimes setting this to 1 will increase convergence speed).
feature_embeddings_dim_exponent: If the additional feature takes
Nvalues, then the embbeding dimension will be set to
pre_word_vecs_enc: If using pretrained word embeddings (on the encoder side), this is the path to the *.hdf5 file with the embeddings. The hdf5 should have a single field
word_vecs, which references an array with dimensions vocab size by embedding size. Each row should be a word embedding and follow the same indexing scheme as the *.dict files from running
preprocess.py. In order to be consistent with
beam.lua, the first 4 indices should always be
pre_word_vecs_dec: Path to *.hdf5 for pretrained word embeddings on the decoder side. See above for formatting of the *.hdf5 file.
fix_word_vecs_enc: If = 1, fix word embeddings on the encoder side.
fix_word_vecs_dec: If = 1, fix word embeddings on the decoder side.
max_batch_l: Batch size used to create the data in
preprocess.py. If this is left blank (recommended), then the batch size will be inferred from the validation set.
start_symbol: Use special start-of-sentence and end-of-sentence tokens on the source side. We've found this to make minimal difference.
gpuid: Which GPU to use (-1 = use cpu).
gpuid2: If this is >=0, then the model will use two GPUs whereby the encoder is on the first GPU and the decoder is on the second GPU. This will allow you to train bigger models.
cudnn: Whether to use cudnn or not for convolutions (for the character model).
cudnnhas much faster convolutions so this is highly recommended if using the character model.
save_every: Save every this many epochs.
print_every: Print various stats after this many batches.
seed: Change the random seed for random numbers in torch - use that option to train alternate models for ensemble
prealloc: when set to 1 (default), enable memory preallocation and sharing between clones - this reduces by a lot the used memory - there should not be any situation where you don't need it. Also - since memory is preallocated, there is not (major) memory increase during the training. When set to 0, it rolls back to original memory optimization.
Decoding options (
model: Path to model .t7 file.
src_file: Source sequence to decode (one line per sequence).
targ_file: True target sequence (optional).
output_file: Path to output the predictions (each line will be the decoded sequence).
src_dict: Path to source vocabulary (
targ_dict: Path to target vocabulary (
feature_dict_prefix: Prefix of the path to the features vocabularies (
char_dict: Path to character vocabulary (
beam: Beam size (recommend keeping this at 5).
max_sent_l: Maximum sentence length. If any of the sequences in
srcfileare longer than this it will error out.
simple: If = 1, output prediction is simply the first time the top of the beam ends with an end-of-sentence token. If = 0, the model considers all hypotheses that have been generated so far that ends with end-of-sentence token and takes the highest scoring of all of them.
replace_unk: Replace the generated UNK tokens with the source token that had the highest attention weight. If
srctarg_dictis provided, it will lookup the identified source token and give the corresponding target token. If it is not provided (or the identified source token does not exist in the table) then it will copy the source token.
srctarg_dict: Path to source-target dictionary to replace UNK tokens. Each line should be a source token and its corresponding target token, separated by
|||. For example
This dictionary can be obtained by, for example, running an alignment model as a preprocessing step. We recommend fast_align.
score_gold: If = 1, score the true target output as well.
n_best: If > 1, then it will also output an n_best list of decoded sentences in the following format.
1 ||| sentence_1 ||| sentence_1_score 2 ||| sentence_2 ||| sentence_2_score
gpuid: ID of the GPU to use (-1 = use CPU).
gpuid2: ID if the second GPU (if specified).
cudnn: If the model was trained with
cudnn, then this should be set to 1 (otherwise the model will fail to load).
rescore: when set to scorer name, use scorer to find hypothesis with highest score - available 'bleu', 'gleu'
rescore_param: parameter to rescorer - for bleu/gleu ngram length
Using additional input features
Linguistic Input Features Improve Neural Machine Translation (Senrich et al. 2016) shows that translation performance can be increased by using additional input features.
Similarly to this work, you can annotate each word in the source text by using the
It supports an arbitrary number of features with arbitrary labels. However, all input words must have the same number of annotations. See for example
data/src-train-case.txt which annotates each word with the case information.
To evaluate the model, the option
-feature_dict_prefix is required on
evaluate.lua which points to the prefix of the features dictionnaries generated during the preprocessing.
Pruning a model
Compression of Neural Machine Translation Models via Pruning (See et al. 2016) shows that a model can be aggressively pruned while keeping the same performace.
To prune a model - you can use
prune.lua which implement class-bind, and class-uniform pruning technique from the paper.
model: the model to prune
savefile: name of the pruned model
gpuid: Which gpu to use. -1 = use CPU. Depends if the model is serialized for GPU or CPU
ratio: pruning rate
prune: pruning technique
uniform, by default
note that the pruning cut connection with lowest weight in the linear models by using a boolean mask. The size of the file is a little larger since it stores the actual full matrix and the binary mask.
Models can be retrained - typically you can recover full capacity of a model pruned at 60% or even 80% by few epochs of additional trainings.
Switching between GPU/CPU models
By default, the model will always save the final model as a CPU model, but it will save the
intermediate models as a CPU/GPU model depending on how you specified
If you want to run beam search on the CPU with an intermediate model trained on the GPU,
you can use
convert_to_cpu.lua to convert the model to CPU and run beam search.
GPU memory requirements/Training speed
Training large sequence-to-sequence models can be memory-intensive. Memory requirements will dependent on batch size, maximum sequence length, vocabulary size, and (obviously) model size. Here are some benchmark numbers on a GeForce GTX Titan X. (assuming batch size of 64, maximum sequence length of 50 on both the source/target sequence, vocabulary size of 50000, and word embedding size equal to rnn size):
prealloc = 0)
- 1-layer, 100 hidden units: 0.7G, 21.5K tokens/sec
- 1-layer, 250 hidden units: 1.4G, 14.1K tokens/sec
- 1-layer, 500 hidden units: 2.6G, 9.4K tokens/sec
- 2-layers, 500 hidden units: 3.2G, 7.4K tokens/sec
- 4-layers, 1000 hidden units: 9.4G, 2.5K tokens/sec
Thanks to some fantastic work from folks at SYSTRAN, turning
will lead to much more memory efficient training
prealloc = 1)
- 1-layer, 100 hidden units: 0.5G, 22.4K tokens/sec
- 1-layer, 250 hidden units: 1.1G, 14.5K tokens/sec
- 1-layer, 500 hidden units: 2.1G, 10.0K tokens/sec
- 2-layers, 500 hidden units: 2.3G, 8.2K tokens/sec
- 4-layers, 1000 hidden units: 6.4G, 3.3K tokens/sec
Tokens/sec refers to total (i.e. source + target) tokens processed per second.
If using different batch sizes/sequence length, you should (linearly) scale
the above numbers accordingly. You can make use of memory on multiple GPUs by using
-gpuid2 option in
train.lua. This will put the encoder on the GPU specified by
-gpuid, and the decoder on the GPU specified by
For translation, evaluation via BLEU can be done by taking the output from
beam.lua and using the
multi-bleu.perl script from Moses. For example
perl multi-bleu.perl gold.txt < pred.txt
Evaluation of States and Attention
attention_extraction.lua can be used to extract the attention and the LSTM states. It uses the following (required) options:
model: Path to model .t7 file.
src_file: Source sequence to decode (one line per sequence).
targ_file: True target sequence.
src_dict: Path to source vocabulary (
targ_dict: Path to target vocabulary (
Output of the script are two files,
decoder.hdf5. The encoder contains the states for every layer of the encoder LSTM and the offsets for the start of each source sentence. The decoder contains the states for the decoder LSTM layers and the offsets for the start of gold sentence. It additionally contains the attention for each time step (if the model uses attention).
We've uploaded English <-> German models trained on 4 million sentences from Workshop on Machine Translation 2015. Download link is below:
These models are 4-layer LSTMs with 1000 hidden units and essentially replicates the results from Effective Approaches to Attention-based Neural Machine Translation, Luong et al. EMNLP 2015.
Our implementation utilizes code from the following: