This example is for training for the WMT’14 English to German news translation task. It will use on the fly tokenization with sentencepiece and sacrebleu for evaluation.

Step 0: Download the data and prepare the subwords model

Preliminary steps are defined in the examples/scripts/ The following command will download the necessary datasets, and prepare a sentencepiece model:

chmod u+x

Note: you should have installed sentencepiece binaries before running this script.

Step 1. Build the vocabulary.

We need to setup the desired configuration with 1. the data 2. the tokenization options:

# wmt14_en_de.yaml
save_data: data/wmt/run/example
## Where the vocab(s) will be written
src_vocab: data/wmt/run/example.vocab.src
tgt_vocab: data/wmt/run/example.vocab.tgt

# Corpus opts:
        path_src: data/wmt/
        path_tgt: data/wmt/
        transforms: [sentencepiece, filtertoolong]
        weight: 23
        path_src: data/wmt/
        path_tgt: data/wmt/
        transforms: [sentencepiece, filtertoolong]
        weight: 19
        path_src: data/wmt/
        path_tgt: data/wmt/
        transforms: [sentencepiece, filtertoolong]
        weight: 3
        path_src: data/wmt/valid.en
        path_tgt: data/wmt/
        transforms: [sentencepiece]

### Transform related opts:
#### Subword
src_subword_model: data/wmt/wmtende.model
tgt_subword_model: data/wmt/wmtende.model
src_subword_nbest: 1
src_subword_alpha: 0.0
tgt_subword_nbest: 1
tgt_subword_alpha: 0.0
#### Filter
src_seq_length: 150
tgt_seq_length: 150

# silently ignore empty lines in the data
skip_empty_level: silent

Then we can execute the vocabulary building script. Let’s set -n_sample to -1 to compute the vocabulary over the whole corpora:

onmt_build_vocab -config wmt14_en_de.yaml -n_sample -1

Step 2: Train the model

We need to add the following parameters to the YAML configuration:


# General opts
save_model: data/wmt/run/model
keep_checkpoint: 50
save_checkpoint_steps: 5000
average_decay: 0.0005
seed: 1234
report_every: 100
train_steps: 100000
valid_steps: 5000

# Batching
queue_size: 10000
bucket_size: 32768
world_size: 2
gpu_ranks: [0, 1]
batch_type: "tokens"
batch_size: 4096
valid_batch_size: 16
batch_size_multiple: 1
max_generator_batches: 0
accum_count: [3]
accum_steps: [0]

# Optimization
model_dtype: "fp32"
optim: "adam"
learning_rate: 2
warmup_steps: 8000
decay_method: "noam"
adam_beta2: 0.998
max_grad_norm: 0
label_smoothing: 0.1
param_init: 0
param_init_glorot: true
normalization: "tokens"

# Model
encoder_type: transformer
decoder_type: transformer
enc_layers: 6
dec_layers: 6
heads: 8
rnn_size: 512
word_vec_size: 512
transformer_ff: 2048
dropout_steps: [0]
dropout: [0.1]
attention_dropout: [0.1]
share_decoder_embeddings: true
share_embeddings: true

Step 3: Translate and evaluate

We need to tokenize the testset with the same sentencepiece model as used in training:

spm_encode --model=data/wmt/wmtende.model \
    < data/wmt/test.en \
    > data/wmt/test.en.sp
spm_encode --model=data/wmt/wmtende.model \
    < data/wmt/ \
    > data/wmt/

We can translate the testset with the following command:

for checkpoint in data/wmt/run/model_step*.pt; do
    echo "# Translating with checkpoint $checkpoint"
    base=$(basename $checkpoint)
    onmt_translate \
        -gpu 0 \
        -batch_size 16384 -batch_type tokens \
        -beam_size 5 \
        -model $checkpoint \
        -src data/wmt/test.en.sp \
        -tgt data/wmt/ \
        -output data/wmt/${base%.*}.sp

Prior to evaluation, we need to detokenize the hypothesis:

for checkpoint in data/wmt/run/model_step*.pt; do
    base=$(basename $checkpoint)
    spm_decode \
        -model=data/wmt/wmtende.model \
        -input_format=piece \
        < data/wmt/${base%.*}.sp \
        > data/wmt/${base%.*}

Finally, we can compute detokenized BLEU with sacrebleu:

for checkpoint in data/wmt/run/model_step*.pt; do
    echo "$checkpoint"
    base=$(basename $checkpoint)
    sacrebleu data/wmt/ < data/wmt/${base%.*}