dksifoua/nmt

State of the art of Neural Machine Translation with PyTorch and TorchText.

Neural Machine Translation

State of the art of Neural Machine Translation with PyTorch and TorchText.

Quick start

I used the europarl-v7 parallel corpora to build models. The data is downloadable here!

• Install requirements and Download data

Run commands below:

$ pip install -r requirements.txt

$ python -m spacy download fr
$ python -m spacy download en
$ python -m spacy download fr_core_news_lg
$ python -m spacy download en_core_web_lg

$ mkdir -p ./checkpoints
$ mkdir -p ./data
$ mkdir -p ./images
$ mkdir -p ./logs

$ wget --no-check-certificate \
    http://www.statmt.org/europarl/v7/fr-en.tgz \
    -O ./data/fr-en.tgz

$ tar -xzvf ./data/fr-en.tgz -C ./data

$ rm -rf ./data/fr-en.tg

or simply run:

$ ./init.sh

• Build datasets

$ python -m scripts.build_datasets --help
usage: build_datasets.py [-h] [--src_lang SRC_LANG] [--dest_lang DEST_LANG]
                         [--n_samples N_SAMPLES] [--min_len MIN_LEN]
                         [--max_len MAX_LEN] [--min_freq MIN_FREQ]
                         [--save SAVE]

Build and save train, validation, and test datasets.

optional arguments:
  -h, --help            show this help message and exit
  --src_lang SRC_LANG   The source language. Default: fr.
  --dest_lang DEST_LANG
                        The destination language. Default: en.
  --n_samples N_SAMPLES
                        The number of samples. Default: 200000.
  --min_len MIN_LEN     The min length of an example. Default: 10.
  --max_len MAX_LEN     The max length of an example. Default: 25.
  --min_freq MIN_FREQ   The min freq of an words in vocabulary. Default: 5.
  --save SAVE           To whether or not save datasets and fields.

Modeling

Encoder-Decoder architecture

The encoder-decoder architecture is a neural network design pattern. As shown in the figure below, the architecture is partitioned into two parts, the encoder and the decoder. The encoder's role is to encode the inputs into state, which often contains several tensors. Then the state is passed into the decoder to generate the outputs. In machine translation, the encoder transforms a source sentence, e.g., Hello world!., into state vector, that captures its semantic information. The decoder then uses this state to generate the translated target sentence, e.g., Bonjour le monde !

Sequence-to-Sequence model

The sequence-to-sequence model is based on the encoder-decoder architecture to generate a sequence output for a sequence input, as demonstrated below. Both the encoder and the decoder commonly use recurrent neural networks (RNNs) to handle sequence inputs of variable length. The hidden state of the encoder is used directly to initialize the decoder hidden state to pass information from the encoder to the decoder. In this project, I tried several sequence-to-sequence models with LSTMs, Attention mechanisms, CNNs and Transformers.

Results

Training

$ python -m scripts.train --help
usage: train.py [-h] --model MODEL [--src_lang SRC_LANG]
                [--dest_lang DEST_LANG] [--batch_size BATCH_SIZE]
                [--init_lr INIT_LR] [--n_epochs N_EPOCHS]
                [--grad_clip GRAD_CLIP] [--tf_ratio TF_RATIO]

Train a model

optional arguments:
  -h, --help            show this help message and exit
  --model MODEL         The model name (SeqToSeqLSTM, SeqToSeqBiLSTM,
                        SeqToSeqLuongAttentionLSTM,
                        SeqToSeqBadhanauAttentionLSTM).
  --src_lang SRC_LANG   The source language. Default: fr.
  --dest_lang DEST_LANG
                        The destination language. Default: en.
  --batch_size BATCH_SIZE
                        The batch size. Default: 64.
  --init_lr INIT_LR     The learning rate. Default: 1e-05.
  --n_epochs N_EPOCHS   The number of epochs. Default: 15.
  --grad_clip GRAD_CLIP
                        The value of gradient clipping. Default: 1.0.
  --tf_ratio TF_RATIO   The teacher forcing ratio. Default: 1.0.

Models	learning rate	loss	val_loss	acc (%)	val_acc (%)	bleu-4 (%)	time/epoch
SeqToSeqLSTM	3.76E-04	2.753	3.125	9.942	9.382	15.012	02min 30s
SeqToSeqBiLSTM	3.76E-04	2.655	3.165	10.132	9.313	14.564	02min 40s
SeqToSeqLuongAttentionLSTM	1.87E-04	3.226	3.491	8.897	8.645	12.027	03min 17s
SeqToSeqBadhanauAttentionLSTM

Evaluation

$ python -m scripts.evaluate --help
usage: evaluate.py [-h] --model MODEL [--src_lang SRC_LANG]
                   [--dest_lang DEST_LANG]

Train a model

optional arguments:
  -h, --help            show this help message and exit
  --model MODEL         The model name (SeqToSeqLSTM, SeqToSeqBiLSTM,
                        SeqToSeqLuongAttentionLSTM,
                        SeqToSeqBadhanauAttentionLSTM).
  --src_lang SRC_LANG   The source language. Default: fr.
  --dest_lang DEST_LANG
                        The destination language. Default: en.

References